Paper deep dive
Performance at What Cost? A Sustainability-Aware Performance Index for Cell and Nucleus Instance Segmentation
Eiram Mahera Sheikh, Alaa Tharwat, Wolfram Schenck
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/11/2026, 4:25:16 AM
Summary
This paper introduces the Sustainability-Aware Performance Index (SAPI), a metric combining segmentation performance, energy consumption, and model size, to evaluate 19 pretrained and foundation models for cell and nucleus instance segmentation. Using the CellBinDB benchmark across six datasets, the study finds that larger models do not consistently justify their higher computational and energy costs with proportional performance gains, highlighting the need for sustainable model selection in biomedical image analysis.
Entities (11)
Relation Signals (10)
SAPI â combines â segmentation performance
confidence 95% · SAPI, a configurable metric that combines segmentation performance, energy consumption, and model size.
SAPI â combines â energy consumption
confidence 95% · SAPI, a configurable metric that combines segmentation performance, energy consumption, and model size.
SAPI â combines â Model Size
confidence 95% · SAPI, a configurable metric that combines segmentation performance, energy consumption, and model size.
CellBinDB â usedforbenchmarking â 19 pretrained and foundation models
confidence 95% · We benchmark 19 pretrained and foundation models across six CellBinDB datasets under zero-shot inference
PQ â evaluates â instance recognition and segmentation quality
confidence 90% · PQ jointly evaluates instance recognition and the segmentation quality of correctly matched objects.
SAM â isusedby â CellSAM
confidence 90% · CellSAM follows a different design: its CellFinder object detector first localizes cells and generates bounding-box prompts, after which the Segment Anything model (SAM) produces the corresponding masks.
AJI+ â measures â instance-level overlap
confidence 90% · AJI + measures instance-level overlap while accounting for unmatched reference and predicted objects.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Pretrained models for cell and nuclear instance segmentation differ substantially in architecture, pretraining data and objectives, parameter count, inference strategy, adaptation requirements, postprocessing pipeline, and computational demand. Large pretrained and foundation models are increasingly adopted because of their strong zero-shot capabilities, but their use also imposes greater energy consumption, memory requirements, computational demands, adaptation costs, and operational carbon emissions. Whether these additional demands are justified by meaningful gains in segmentation performance remains unclear. We address this question by introducing the Sustainability-Aware Performance Index (SAPI), a configurable metric that combines segmentation performance, energy consumption, and model size. We benchmark 19 pretrained and foundation models across six CellBinDB datasets under zero-shot inference and evaluate 16 fine-tunable models using few-shot adaptation with both frozen encoder and full-model fine-tuning. We estimate energy consumption for GPU, CPU, and RAM using software-based monitoring tools. Our results show that larger and more computationally demanding models do not consistently achieve proportionate improvements in segmentation quality. While few-shot adaptation benefits several models, the gains and resource costs vary considerably across architectures, datasets, and adaptation strategies, causing SAPI-based rankings to differ from rankings based on performance alone. This study provides a practical framework for comparing segmentation models more comprehensively and supports more computationally accessible and environmentally responsible model selection in biomedical image analysis.
Tags
Links
- Source: https://arxiv.org/abs/2610.10324v1
- Canonical: https://arxiv.org/abs/2610.10324v1
Trouble viewing inline? Open PDF directly â
Full Text
131,788 characters extracted from source content.
Expand or collapse full text
Performance at What Cost? A Sustainability-Aware Performance Index for Cell and Nucleus Instance Segmentation Eiram Mahera Sheikh, Alaa Tharwat, Wolfram Schenck Department of Engineering and Mathematics, Bielefeld University of Applied Sciences and Arts, Interaktion 1, 33619 Bielefeld eiram_mahera.sheikh, alaa.othman, wolfram.schenck@hsbi.de Abstract Pretrained models for cell and nuclear instance segmentation differ substantially in architecture, pre- training data and objectives, parameter count, inference strategy, adaptation requirements, post- processing pipeline, and computational demand. Large pretrained and foundation models are increas- ingly adopted because of their strong zero-shot capabilities, but their use also imposes greater energy consumption, memory requirements, computational demands, adaptation costs, and operational carbon emissions. Whether these additional demands are justified by meaningful gains in segmentation perfor- mance remains unclear. We address this question by introducing the Sustainability-Aware Performance Index (SAPI), a configurable metric that combines segmentation performance, energy consumption, and model size. We benchmark 19 pretrained and foundation models across six CellBinDB datasets under zero-shot inference and evaluate 16 fine-tunable models using few-shot adaptation with both frozen- encoder and full-model fine-tuning. We estimate energy consumption for GPU, CPU, and RAM using software-based monitoring tools. Our results show that larger and more computationally demanding models do not consistently achieve proportionate improvements in segmentation quality. While few-shot adaptation benefits several models, the gains and resource costs vary considerably across architectures, datasets, and adaptation strategies, causing SAPI-based rankings to differ from rankings based on per- formance alone. This study provides a practical framework for comparing segmentation models more comprehensively and supports more computationally accessible and environmentally responsible model selection in biomedical image analysis. Keywords: Green AI, Sustainable artificial intelligence, Cell segmentation, Nucleus segmentation, Vi- sion Foundation models, Few-shot learning, Energy-efficient deep learning, Biomedical image analysis 1 Introduction Cell and nucleus instance segmentation is a foundational step in quantitative microscopy because many down- stream analyses require individual objects to be identi- fied and delineated. Reliable object instance boundaries enable cell counting, morphometric measurement, phe- notypic profiling, spatial analysis of cell populations, and the assignment of molecular signals to individual cells. In digital pathology, nuclear segmentation supports the extraction of interpretable cellular features and the char- acterization of tissue architecture; in multiplex tissue imaging, whole-cell and nuclear masks convert spatially resolved marker measurements into cell-level data; and in experimental biology, segmentation underpins high- content assays in which large image collections must be transformed into reproducible quantitative readouts. These applications span basic cell biology, computational pathology, spatial omics, drug discovery, and biopharma- ceutical development, but they share the same practical requirement: segmentation must remain reliable across variation in tissue, staining chemistry, imaging modality, cell morphology, object density, and image quality. Er- rors introduced at this stage propagate into subsequent measurements, making segmentation quality consequen- tial for the validity of the biological or clinical analysis constructed from the resulting instances [1]. Cell and nucleus segmentation was initially ap- proached through classical image-processing pipelines that combined intensity thresholding, filtering, morpho- logical operations, contour detection, distance trans- forms, and watershed-based separation. These methods were often computationally efficient and interpretable, but their performance depended strongly on manually selected parameters and assumptions about image con- trast, object shape, and boundary visibility. Variations in staining, illumination, noise, cell morphology, and ob- ject density could therefore require extensive retuning and limit transfer across datasets. Deep convolutional neural networks changed this paradigm by learning hier- archical image representations directly from annotated data and combining contextual encoding with precise spatial localization, as demonstrated by the encoderâ decoder architecture and skip connections introduced in U-Net [2]. Once reliable pixel-level prediction be- came possible, subsequent work increasingly addressed the more difficult problem of converting foreground pre- dictions into distinct cellular instances. Convolutional architectures therefore incorporated object-centered ge- ometric representations [3], learned displacement fields [4], and dedicated output branches to distinguish ad- 1 arXiv:2610.10324v1 [cs.CV] 7 Oct 2026 joining or overlapping nuclei in crowded microscopy and histopathology images [5]. These developments moved instance segmentation beyond generic foreground clas- sification by allowing object shape, relative pixel po- sition, and instance boundaries to be learned as part of the prediction task. The next major architectural shift arose with vision transformers, which represented images as sequences of patches and used self-attention to learn interactions among image regions, particularly when pre-trained on large image collections [6]. Trans- former encoders were then incorporated into biomedi- cal image segmentation architectures by combining pre- trained encoders with task-specific decoders for cell and nucleus delineation [7]. Large-scale promptable segmen- tation introduced a foundation-model paradigm in which a high-capacity image encoder trained on an extensive collection of images and masks, could support mask gen- eration across varied image distributions and segmen- tation tasks [8]. This paradigm was then extended to microscopy by adapting foundationâmodel representa- tions and segmentation interfaces to biological images and microscopy-specific annotation workflows [9]. This progression has culminated in large vision transformerâ based foundation models pre-trained on extensive and diverse image collections. The practical appeal of pre-trained and foundation models lies in their ability to segment images from a new dataset without additional target-specific training. In this zero-shot setting, a model is applied using knowl- edge acquired during pre-training, thereby avoiding the time and annotation effort required to construct a new model for each imaging study. However, recent com- parative evaluations indicate that zero-shot performance remains strongly dependent on the characteristics of the target data [10]. When zero-shot predictions do not pro- vide sufficient accuracy, the pre-trained representation can instead serve as an initialization for adaptation to the target domain. Fine-tuning with a small support set of annotated target images allows the model to learn dataset-specific appearance and morphological charac- teristics without requiring the large annotation collec- tions typically needed to train a segmentation network from scratch. Few-shot fine-tuning studies have shown that pre-trained models can be adapted using only a limited number of annotated samples from the target dataset, although the resulting performance depends on which images are selected for annotation and how well they represent the target distribution [11]. The growing use of large pre-trained models has made the computational cost of artificial intelligence an in- creasingly important consideration alongside predictive performance. This concern motivated the emergence of Green AI, which argues that methodological progress should be evaluated not only by improvements in accu- racy, but also by the computational resources required to develop and use a model [12]. Early empirical evidence from natural language processing showed that training, hyperparameter optimization, and repeated experimen- tation with large neural networks could entail substantial energy consumption, financial cost, and associated car- bon emissions [13]. Subsequent analyses of large-scale neural network training further demonstrated that envi- ronmental impact is shaped not only by the amount of computation performed, but also by model architecture, hardware efficiency, data-centre infrastructure, and the carbon intensity of the electricity supply [14]. These find- ings led to calls for more systematic measurement and reporting of energy use and carbon emissions in machine- learning research, so that improvements in model quality can be interpreted in relation to the resources required to achieve them [15]. Accordingly, improvements in pre- dictive performance should be interpreted together with the computational resources required to obtain them. This requires explicit consideration of the costs asso- ciated with model development, target-domain adapta- tion, and repeated inference, particularly when large pre- trained models are intended for broad or sustained use. Despite substantial progress in cell and nucleus in- stance segmentation, model evaluation remains domi- nated by predictive performance. Existing benchmarks compare segmentation accuracy across imaging modali- ties, tissues, and model families, but they provide limited evidence about the computational resources required to obtain and sustain that performance. This omission is increasingly important because contemporary segmenta- tion methods differ markedly in model scale, inference requirements, and the amount of target-domain adap- tation they require. A model that achieves the highest segmentation score may also consume considerably more energy during inference, require costly fine-tuning, or de- pend on hardware that is unavailable in many research and deployment settings. Conversely, a computation- ally efficient model may provide comparable performance at substantially lower operational cost. However, these trade-offs cannot be determined from accuracy, parame- ter count, or runtime alone. To our knowledge, no con- trolled benchmark has jointly evaluated zero-shot per- formance, few-shot adaptation, inference and fine-tuning energy, and model complexity across a diverse collection of specialist, generalist, and foundation models for cell and nucleus instance segmentation. As a result, it re- mains unclear whether the performance gains offered by larger pretrained models are proportionate to the addi- tional resources required to adapt and deploy them. Ad- dressing this gap is necessary to support model selection that reflects not only segmentation quality, but also com- putational accessibility, operational efficiency, and the practical requirements of the intended application. Motivated by this gap, we ask whether the segmenta- tion performance gains offered by larger pretrained and foundation models justify their additional inference en- ergy, adaptation energy, and model complexity. To ad- dress this question, our study makes three main contri- butions. First, we conduct a controlled benchmark of 19 cell and nucleus instance-segmentation models across six CellBinDB datasets [10] covering DAPI, single-stranded DNA, multiplex immunofluorescence, and H&E imaging modalities, with 1,044 images distributed across 30 tis- sue types and containing over 100,000 annotated cell and nuclear instances. Specifically, we evaluate the mod- els under zero-shot condition to determine the perfor- mance provided by their existing pretrained representa- tions without target-domain optimization. For the 16 models with fine-tuning API available to us, we compare frozen-encoder and full-model fine-tuning using tissue- stratified few-shot support sets, thereby examining both 2 the benefits and the computational costs of adapting pre- trained models to unknown datasets. Second, we mea- sure energy consumption (GPU, CPU, and RAM) dur- ing inference and fine-tuning on a fixed hardware us- ing software based estimates, and quantify model com- plexity via total parameter count. Third, we introduce the Sustainability-Aware Performance Index (SAPI), a configurable metric that associates segmentation perfor- mance with energy demand and model scale. Together, these contributions provide a unified framework for de- termining when additional model capacity and computa- tional expenditure produce a proportionate improvement in segmentation performance. The broader objective is to make sustainability and computational accessibility integral components of methodological quality in cell and nuclear segmentation, rather than considerations introduced only after a model has been selected. This perspective does not imply that the smallest model is inherently preferable or that en- ergy efficiency should take precedence over segmenta- tion quality. Instead, additional computational demand should be justified by improvements that are meaningful for the intended application. This principle is relevant to high-throughput pathology, spatial molecular profil- ing, multiplex imaging, pharmaceutical screening, and academic research, where segmentation models may be applied repeatedly to large image collections under sub- stantially different hardware, memory, and energy con- straints. Resource-aware evidence can also strengthen reproducibility by revealing whether reported perfor- mance gains depend on computationally intensive adap- tation or deployment conditions that may not be acces- sible across institutions. By jointly evaluating segmenta- tion performance, adaptation requirements, operational energy consumption, and model scale, this study pro- vides a basis for determining whether improvements in accuracy are proportionate to the resources required to achieve and sustain them. More broadly, it frames model selection as a balance among technical performance, gen- eralizability, computational accessibility, and environ- mental responsibility. Such an approach can support the development and adoption of segmentation methods that are not only accurate, but also reproducible, practically feasible, and appropriate for the biological, clinical, or industrial setting in which they will be used. Finally, the remainder of this paper is organized as follows. Section 2 describes the CellBinDB datasets and the model zoo, together with the segmentation per- formance metrics, energy measurement protocols, the Sustainability-Aware Performance Index, and the exper- imental study design. Section 3 presents the results of zero-shot and few-shot evaluations, including segmenta- tion performance and energy consumption. Section 4 discusses the implications of our findings for sustainable model selection, and Section 5 concludes with a summary of the key findings. 2 Methods 2.1 Datasets We evaluate all models on the CellBinDB datasets [10], which is a multimodal benchmark for cell in- Table 1: Details of the six CellBinDB datasets. DatasetImages Cell instances Tissues Image size mIF606,0133 256Ă 256 10xGenomics_HE1007,0873 256Ă 256 10xGenomics_DAPI1007,7452 256Ă 256 DAPI203 16,6576 256Ă 256 ssDNA276 23,86720 256Ă 256 HE305 41,1116 512Ă 512 stance segmentation that contains expert instance an- notations across six imaging modalities. The col- lection combines in-house Stereo-seq acquisitions with publicly released 10xGenomics FFPE data, spanning four fluorescence modalities (multiplex immunofluores- cence (mIF), single-stranded DNA (ssDNA), DAPI, and 10xGenomics_DAPI) and two haematoxylinâeosin modalities (HE and 10xGenomics_HE). In total, the datasets have 1,044 images containing 102,480 annotated cell instances, distributed across 30 distinct tissue types over the six modalities. Table 1 summarizes the com- position of each dataset and Figure 1 shows one sample image with its reference mask from each dataset. We chose CellBinDB for three reasons. First, it spans several microscopy domains within a single resource that was consistently annotated, so the same evaluation pro- tocol applies unchanged to every model and every do- main. Second, none of the evaluated checkpoints was trained on CellBinDB, which makes it an independent test set for both zero-shot evaluation and target-specific adaptation. Third, it varies along the axes that drive the difficulty of the task: staining, tissue type, image source, and image size. The fluorescence and H&E images differ in their intensity distributions and visual structure, and cell appearance changes further with the staining method and the tissue type. Many images contain densely packed or touching cells, which are hard to separate into indi- vidual instances. Others show large variation in cell size, shape, and staining intensity. Together, these differences amount to real domain shifts between the six datasets and tests whether a model holds its performance across heterogeneous microscopy data. We inverted the mIF images following the same pro- tocol as in the original CellBinDB paper. For the H&E images we used colour-deconvolution [16] to extract the hematoxylin channel. No other preprocessing was ap- plied by us to any dataset. 2.2 Model zoo The model zoo was designed to represent the principal methodological choices currently encountered in cell and nuclear instance segmentation. The benchmark includes 19 pretrained models from 11 model families, span- ning task-specific convolutional networks, transformer- based segmentation models, and adaptations of general- purpose vision foundation models. These models dif- fer in their encoder architecture and scale, instance- reconstruction strategy, separation of touching objects, pretraining domain, and suitability for different mi- croscopy modalities. Their evaluation across the six 3 mIF 10xGenomics DAPIDAPIssDNA 10xGenomics HEHE Figure 1: Sample images and their corresponding instance segmentation masks from the six CellBinDB datasets. Each column shows one dataset. The top row shows the microscopy image and the bottom row shows the reference instance masks. datasets was therefore determined by the modality and domain of the data on which the corresponding models were originally trained. Models developed specifically for fluorescence microscopy were evaluated on the fluo- rescence datasets, models developed for histopathology were evaluated on the H&E datasets, and models with pretraining that supports both modalities were evalu- ated across all six CellBinDB datasets. This separation avoids evaluating models outside their intended input domain, where differences in performance could reflect unsupported modality shifts rather than the segmenta- tion capability of the model itself. At the same time, it preserves the practical model-selection setting in which users choose models according to the imaging modality and training domain for which they were developed. As shown in Figure 2, their total parameter counts range from 1.4 million to approximately 700 million, covering more than two orders of magnitude in model size. The patterned bars additionally indicate whether each model was evaluated on the fluorescence datasets only, the H&E datasets only, or all the six datasets. A central distinction among the evaluated models is how they convert pixel-level predictions into separate object instances. StarDist [3] imposes an explicit ge- ometric prior by representing each object as a star- convex polygon described by radial distances from a candidate centre. Cellpose [4] on the other hand pre- dicts spatial flows that guide pixels towards their cor- responding object centres. Cellpose-SAM [17] follows this same decoding and post-processing pipeline; how- ever, it uses a vision transformer encoder. MEDIAR [18] also uses a similar flow-based formulation, while com- bining it with a hierarchical transformer encoder and a multiscale decoder developed for heterogeneous mi- croscopy images. HoVer-Net [5] predicts nuclear fore- ground together with horizontal and vertical displace- ment maps, which are used to separate touching nuclei in histopathology images. Mesmer [19] and DeepCell [20] use multiscale convolutional segmentation networks and distance-based post-processing. InstanSeg [21] instead learns object-centred pixel representations from which individual instances are recovered. CellSAM [22] follows a different design: its CellFinder object detector first localizes cells and generates bounding-box prompts, af- ter which the Segment Anything model (SAM) [8] pro- duces the corresponding masks. These approaches ad- dress the same instance separation problem through dif- ferent inductive biases. CellViT-256 and CellViT-SAM- H [7] share the same general histopathology segmen- tation framework but use markedly different encoders: CellViT-256 employs a comparatively compact DINO- pretrained vision transformer, whereas CellViT-SAM-H uses the substantially larger ViT-H encoder initialized from SAM. Micro-SAM [9] and PathoSAM [23] adapt SAM directly to biomedical image domains through fine- tuning. The three Micro-SAM configurations use light- microscopy-adapted ViT-T, ViT-B, and ViT-L encoders with an automatic instance-segmentation decoder. The three PathoSAM configurations use ViT-B, ViT-L, and ViT-H encoders adapted specifically for histopathology. Within each family, varying the encoder while retaining a common application domain and segmentation formu- lation creates a controlled model-scale comparison. A comparison of all these models therefore helps to deter- mine whether computationally more elaborate encoders provide greater practical value than carefully designed object representations and post-processing procedures. Comparing these models is important because a larger foundation-model encoder may increase representational capacity, but also increases parameter storage, infer- ence energy, memory requirements, and fine-tuning cost. Overall, the model zoo provide structured variation along four dimensions: instance-reconstruction strategy, en- coder architecture and scale, modality specialization, and adaptation regime. The inclusion of related variants within Cellpose, CellViT, Micro-SAM, and PathoSAM permits relatively controlled within-family comparisons, whereas the broader set supports comparisons across fun- damentally different formulations. 4 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 2 8 32 128 512 Number of parameters (millions) 1.4 1.4 3.8 6.66.6 21.2 26.0 45.0 46.8 103.3 104.8 119.4 121.3 289.6 304.6 321.9 323.4 652.1 699.7 Fluorescence-only H&E-only All datasets Figure 2: Model size and dataset compatibility of the 19 evaluated models. Bar height represents the total number of parameters (N Ξ ) on a logarithmic scale, with exact counts reported in millions above each bar. Hatch patterns indicate whether a model was used with the fluorescence datasets, the H&E datasets, or all six CellBinDB datasets. 2.3 Segmentation performance evalua- tion We evaluate segmentation performance using the Aggre- gated Jaccard Index plus (AJI + ) [5], Panoptic Quality (PQ) [24], and Normalized Surface Dice (NSD) [25] met- rics, which capture complementary aspects of segmenta- tion quality. AJI + measures instance-level overlap while accounting for unmatched reference and predicted ob- jects. PQ jointly evaluates instance recognition and the segmentation quality of correctly matched objects. NSD measures agreement between the predicted and reference boundaries within a specified spatial tolerance. All three metrics range from zero to one, with higher values indi- cating better segmentation quality. Let G = G 1 ,...,G K denote the set of reference in- stances in an image, and let R = R 1 ,...,R L denote the corresponding set of predicted instances, where K and L are the total numbers of reference and predicted instances, respectively. AJI + uses a one-to-one assign- ment between reference and predicted instances. In our implementation, this assignment was obtained using the Hungarian algorithm without imposing an intersection- over-union threshold. Let M denote the resulting set of matched pairs, and letU G and U R denote the sets of un- matched reference and predicted instances, respectively. AJI + is defined as: AJI + = X (k,l)âM |G k â© R l | X (k,l)âM |G k âȘ R l | + X kâU G |G k | + X lâU R |R l | (1) The one-to-one assignment ensures that each reference or predicted instance participates in at most one match. Unmatched instances contribute only to the denomi- nator. Consequently, AJI + penalizes missed instances, false-positive detections, merges, and splits while mea- suring the overall agreement between the reference and predicted instance segmentations. PQ is calculated using the standard instance-matching criterion. A reference instance and a predicted instance are considered a true-positive match when their inter- section over union is strictly greater than 0.5. Let TP denote the set of matched instance pairs, and let FP and FN denote the sets of unmatched predicted and reference instances, respectively. Let IoU denote the intersection- over-union between the two masks. PQ is defined as: PQ = X (G k ,R l )âTP IoU(G k ,R l ) |TP| + 1 2 |FP| + 1 2 |FN| (2) The numerator measures the segmentation quality of the matched instances, whereas the denominator accounts for both matched instances and recognition errors aris- ing from false-positive and false-negative detections. PQ therefore reflects both the accuracy with which individ- ual objects are identified and the spatial quality of their predicted masks. For the NSD, the reference and predicted instance la- bel maps are converted into binary foreground masks. NSD therefore evaluates the boundary of the complete foreground region without considering the identities of individual instances. Let S G and S R denote the sur- faces of the reference and predicted foreground masks, respectively. Let w(x) denote the length associated with surface element x, and let d(x,S) denote the shortest distance from x to surface S. For a spatial tolerance Ï, NSD is defined as: NSD Ï = P xâS G w(x) I[d(x,S R )â€ Ï ] + P yâS R w(y) I[d(y,S G )â€ Ï ] P xâS G w(x) + P yâS R w(y) (3) 5 Here, I[·] denotes the indicator function. A surface el- ement contributes to the numerator when its nearest point on the opposing surface lies within the specified tolerance. In our experiments, NSD is calculated using a tolerance of Ï = 2 pixels and a pixel spacing of [1, 1] for all the datasets. AJI + , PQ, and NSD are first calculated independently for each test image. Let A i , Q i , and D i denote the corre- sponding AJI + , PQ, and NSD values for image i. When both the reference and predicted masks contained in- stances but no valid instance matches were identified, we set AJI + and PQ to zero, and calculate NSD from the corresponding binary foreground surfaces. For a test set containing N images, the dataset-level metrics are calculated as: A = 1 N N X i=1 A i , Q = 1 N N X i=1 Q i , D = 1 N N X i=1 D i (4) This aggregation assigns equal weight to every test im- age, preventing images with a larger number of cells or a greater foreground area from disproportionately influ- encing the dataset-level results. Because no individual metric provides a complete char- acterization of segmentation quality [26, 27], we com- bined the three dataset-level metrics using their har- monic mean into a single segmentation performance score, denoted by P. The harmonic mean is used be- cause it requires balanced performance across instance overlap, instance recognition, and boundary agreement. A low value for any one component substantially reduces the combined score and cannot be fully compensated for by high values in the remaining components. The per- formance score is defined as: P = ïŁ± ïŁŽ ïŁČ ïŁŽ ïŁł 0,min A,Q,D = 0, 3 A â1 +Q â1 +D â1 , otherwise. (5) 2.4 Energy measurements We measure energy consumption separately for each in- ference and fine-tuning phase using hardware-level en- ergy counters exposed through the operating system and GPU management interface. Direct wall-socket measure- ment with an external power meter was not used because the experiments were conducted on a shared data-center compute cluster rather than on a dedicated workstation with an accessible, isolated power connection. In such an environment, an external meter installed at the rack, power-distribution unit, or server power input would measure the aggregate electrical consumption of the cor- responding infrastructure and would not directly isolate the energy attributable to the selected GPU and proces- sor domains. Obtaining component-level measurements with external instrumentation would additionally require physical access to the server power-delivery path and coordination with the data-center infrastructure, which was outside the scope of available experimental setup. This is why, we used software accessible hardware coun- ters to provide a practical and reproducible means of obtaining component-level measurements without modi- fying the cluster infrastructure. GPU energy measurement was obtained using the NVIDIA Management Library (NVML), accessed through the nvidia-ml-py Python bindings [28, 29]. NVML is the NVIDIA-supported programmatic inter- face underlying nvidia-smi and provides access to GPU power measurements for supported data-center GPUs, including the Tesla V100 used in this study [29]. The use of NVML is also established in empirical studies of en- ergy consumption on NVIDIA GPUs [30, 31]. Processor- package and DRAM energy were obtained from Intel Running Average Power Limit (RAPL) energy coun- ters exposed through the Linux powercap interface [32â 34]. RAPL has been extensively evaluated against exter- nal power measurements and has been reported to ex- hibit strong correlation with plug-level power consump- tion while introducing negligible measurement overhead, making it suitable for energy monitoring on servers and computing clusters [33]. These interfaces measure energy at the hardware-domain level and thus do not include power consumed by the power supply, cooling system, storage devices, networking equipment, or other infras- tructure outside the monitored GPU, CPU-package, and DRAM domains. The compute node was allocated exclusively to the ex- periments, and each experiment used one selected GPU without competing GPU processes. The complete energy increment recorded for the selected GPU, CPU packages, and DRAM domains during each measurement interval was attributed to the benchmark workload. This ex- perimental isolation reduces interference from concurrent computational workloads and makes the component-level measurements comparable across models. The same hardware, measurement interfaces, measurement proce- dure, and measurement scope were used for all evaluated models, which is particularly important for the relative comparison of energy consumption. The measurement interval encompassed the complete inference or fine-tuning phase executed by the experi- ment. For inference, this included network instantia- tion, checkpoint loading, transfer to the GPU, CUDA ini- tialization, model-specific normalization, augmentations, channel adaptation, resizing, rescaling, tiling, forward propagation, transfer of outputs to the CPU, and model- specific instance reconstruction and post-processing. For fine-tuning, the measured interval additionally included loss calculation, backpropagation, parameter updates, and all training epochs. Thus, model-specific compu- tational operations such as Cellpose flow integration, StarDist non-maximum suppression, and SAM-based mask generation were included in the measured energy. Dataset-level transformations performed before model invocation, logging, and metric calculations were ex- cluded because they were not part of the model execution pipeline being evaluated. GPU operations were synchro- nized immediately before the initial and final counter readings to ensure that asynchronously queued GPU op- erations were completed before the corresponding energy values were recorded. Because NVML and RAPL report hardware-domain energy rather than wall-socket electricity, their measure- ments are not assumed to be exact substitutes for ex- ternal power-meter measurements. Previous validation studies have shown that the agreement between on-chip 6 energy sensors and external meters depends on the hard- ware architecture and workload, particularly for GPU measurements [31]. The present study therefore empha- sizes consistency of the measurement procedure and con- trolled comparison across models rather than interpret- ing the reported values as exact facility-level electricity consumption. 2.4.1 GPU energy GPU energy was measured using the cumulative device- energy counter provided by NVML. After synchronizing the selected GPU, the counter was read at the start of each measured inference or fine-tuning phase. The GPU was synchronized again after the phase to ensure that all queued operations had completed before the counter was read a second time. Let G start and G end denote the counter readings at the start and end boundaries, respectively. Both values represent the cumulative energy consumed by the GPU since the NVIDIA driver was loaded and are reported by NVML in millijoules (mJ). They do not represent instantaneous GPU power. The GPU energy consumed during the measured phase is calculated as: E GPU = G end â G start (6) where E GPU is expressed in millijoules (mJ). Although the counter is cumulative, subtracting the start reading from the end reading isolates the energy consumed be- tween the two measurement boundaries. This device-level measurement represents the energy recorded for the complete selected GPU rather than for an individual process. Because the selected GPU was used exclusively by our experiment, the complete counter increment was assigned to the evaluated workload. 2.4.2 CPU and memory energy CPU and host-memory energy were obtained using the cumulative RAPL energy counters exposed through the Linux powercap interface. Counter values were recorded at the beginning and end of each measured phase. Processor-package energy is calculated as: E CPU = X sâS âR pkg,s 10 3 (7) where S is the set of detected physical CPU packages and âR pkg,s is the change in the package energy counter for the package s, expressed in microjoules. Dividing by 10 3 converts microjoules to millijoules. The package do- main includes the processor cores and other components within the CPU package. Host-memory energy was calculated separately by summing the increments of all detected DRAM energy counters: E RAM = X dâD âR DRAM,d 10 3 (8) where D is the set of detected DRAM energy domains and âR DRAM,d is the counter increment for DRAM do- main d, expressed in microjoules. Thus, E RAM is ex- pressed in millijoules. Counter wraparound was han- dled using the maximum counter range reported by the powercap interface. 2.4.3 Total IT energy The total energy consumed by either a training or an inference phase is calculated as: E IT = E GPU + E CPU + E RAM (9) It excludes datacenter infrastructure overhead and uninstrumented components such as storage, network- ing, motherboard circuitry, and power-supply conversion losses. 2.4.4 Inference energy Inference energy is the total energy consumed by the inference phase amortized over the number of images in the test set: E i = E IT N test (10) where E IT is the total IT energy measured over the com- plete inference phase and N test is the number of test images processed during that phase. Therefore, E i is expressed in millijoules (mJ) per test image. 2.4.5 Fine-tuning energy Fine-tuning energy is the total energy consumed during the fine-tuning phase amortized over the number of im- ages in the support set: E ft = E IT N support (11) where E IT is the total IT energy measured over the com- plete fine-tuning phase, including all training epochs, and N support is the number of unique images in the sup- port set. Therefore, E ft is expressed in millijoules (mJ) per support image and represents the total energy of the complete fine-tuning procedure. 2.5 Sustainability-Aware Performance Index (SAPI) Several composite metrics have been proposed to move model evaluation beyond predictive performance alone. NetScore [35] was introduced to quantify the balance be- tween accuracy, architectural complexity, and computa- tional complexity. For a network N, NetScore is defined as: âŠ(N ) = 20 log 10 a(N ) 2 p p(N )· m(N ) ! (12) where a(N ) is predictive accuracy, p(N ) is the parameter count, and m(N ) is the number of multiply-accumulate operations (MAC) required for inference. NetScore represented an important step towards resource-aware model comparison, but its computational term is an an- alytical operation count only. MAC count does not cap- ture hardware efficiency, memory access, data movement, software implementation, or other system-level effects that can cause models with similar operation counts to consume different amounts of energy. 7 The authors in the paper [36] subsequently proposed the Sustainable-Accuracy Metric (SAM) to combine pre- dictive accuracy with electricity consumption. SAM is defined as: SAM = ÎČ a α log 10 (E) (13) where a is accuracy and E is electricity consumption in kilowatt-hours. This formulation replaces an analyt- ical computational proxy with electricity consumption. Here, ÎČ is only a global scaling constant; it changes the numerical range of SAM but cannot change the weight of accuracy or energy. Also, SAM contains no independent representation of model scale, meaning that two mod- els with similar accuracy and measured energy receive similar scores even when their storage and memory re- quirements differ substantially. To address these limitations, we propose a novel met- ric, Sustainability-Aware Performance Index (SAPI). It combines predictive performance, energy consumption during inference and fine-tuning, and model parameter count through a weighted multiplicative formulation: SAPI = P α [log 10 (E)] ÎČ Â· [log 10 (N Ξ )] Îł (14) where P is the segmentation performance as defined in Equation 5; E is the energy consumed by the model, and N Ξ is the total number of parameters in the model. The base-10 logarithm is applied to E and N Ξ be- cause both quantities vary over several orders of mag- nitude across model architectures, whereas the perfor- mance score (P) is bounded between zero and one. Using the unscaled energy and parameter values would allow their substantially larger numerical ranges to dominate the overall SAPI score and would impose a penalty that increases directly with absolute scale. The logarithmic transformation compresses these ranges while preserving their ordering, so that a model with greater energy con- sumption or more parameters always receives a larger penalty. It also introduces diminishing marginal penal- ties: an equal multiplicative increase in energy or param- eter count produces an equal increment on the logarith- mic scale, irrespective of the initial magnitude. Base 10 is used because it provides a direct interpretation in terms of orders of magnitude; for example, a tenfold increase in either quantity increases its logarithm by one. Under these fixed conventions, the logarithmic terms provide interpretable and comparable penalties for operational energy demand and model scale without allowing either component to overwhelm the performance term. The exponents α, ÎČ, and Îł are positive weights satis- fying the constraint: α + ÎČ + Îł = 1(15) The exponents α, ÎČ and Îł control the relative influ- ence of performance, energy consumption, and model complexity, respectively. Requiring α,ÎČ,Îł > 0 preserves the intended monotonic behaviour of the index: higher performance must increase SAPI, whereas higher energy consumption or a larger parameter count must decrease it. A zero exponent would remove the corresponding di- mension from the index, while a negative exponent would reverse its meaning by rewarding lower performance or greater resource use. The constraint (α+ÎČ+Îł = 1) is im- posed to normalize the exponents into a common weight- ing scale. Without this constraint, multiplying all three exponents by the same positive constant would change the numerical magnitude and dispersion of SAPI without changing the relative weighting among its components or the resulting model ordering, because the transformed score would simply be a positive power of the original score. Enforcing a unit sum therefore removes this arbi- trary degree of freedom and makes each weighting con- figuration uniquely defined and directly comparable with alternative configurations. It also permits relative prior- ities, such as (3:2:1), to be represented unambiguously as (α=1/2), (ÎČ=1/3), and (Îł=1/6). For a modelâs zero-shot inference, the SAPI score is defined as: SAPI ZS = P α ZS [log 10 (E i )] ÎČ Â· [log 10 (N Ξ )] Îł (16) where P ZS is the zero-shot performance of the model and E i is the inference energy per test image as defined in Equation 10. For a fine-tuned model, the energy term combines the inference energy of the adapted model per test image with the energy of the complete fine-tuning procedure per support image: E FT+I = λE i + (1â λ)E ft (17) where λ â [0, 1]. E ft is the fine-tuning energy per sup- port image as defined in Equation 11 and E i is the post fine-tuning inference energy per test image as defined in Equation 10. The parameter λ controls the relative emphasis as- signed to inference energy, whereas (1â λ) controls the emphasis assigned to fine-tuning energy. This weighting is motivated by the different temporal roles of the two energy costs. Fine-tuning is typically performed once, or only occasionally, for a given model and target dataset, whereas inference is repeated for every input processed after deployment. Consequently, the contribution of the initial adaptation cost becomes progressively amortized as the number of inferences increases, while cumulative inference energy continues to grow with deployment vol- ume [37â39]. For fine-tuning followed by inference, SAPI is defined as: SAPI FT+I = P α FT [log 10 (E FT+I )] ÎČ Â· [log 10 (N Ξ )] Îł (18) where P FT is the performance of the fine-tuned model. 2.6 Experimental Study Design We evaluated the models under three experimental con- ditions: zero-shot inference, few-shot full-model fine- tuning, and few-shot fine-tuning with a frozen encoder. Because CellBinDB does not provide predefined training, validation, and test partitions, we defined the support and test sets for the few-shot experiments as described below. The same evaluation protocol was used for all applicable model-dataset combinations. 8 2.6.1 Zero-shot Evaluation All 19 model configurations were evaluated in the zero- shot setting. Each model was applied to all the images in all the datasets supported by its respective pretrained checkpoint. No image was used to update model pa- rameters, select inference thresholds, or modify model- specific post-processing settings. Predictions were gen- erated using the native inference pipeline of each imple- mentation. No manually supplied point, box, or mask prompts were used. Models that support prompt-based inference were instead evaluated using their automatic instance-segmentation procedures. Model-specific auto- matic settings, such as automatic diameter estimation in Cellpose, were retained where applicable. This experi- mental setting allowed us to assess how the models per- form on the CellBinDB datasets out of the box, without any dataset-specific adaptation. 2.6.2 Few-shot Adaptation Pretrained segmentation models can be adapted to a tar- get microscopy domain using a small number of labeled images [9, 40, 41]. However, the effectiveness of few-shot adaptation can depend on the images selected for fine- tuning [11]. This is particularly relevant for microscopy instance segmentation, where producing accurate object- level annotations is laborious and time-consuming [42]. For few-shot fine-tuning, we used the following proce- dure for selecting the support set and test set. For each dataset, we randomly selected one image, using a pre- defined seed, from a tissue group within the dataset as a support image. The total number of support images therefore equals to the number of tissues present in the dataset, as reported in Table 1. All the remaining images in the dataset formed the test set. For a given dataset and seed, the same support and test sets was used for every compatible model. We did not create a separate validation set because splitting the already small support set would further re- duce the number of labeled images available for adapta- tion. An exploratory analysis of training-loss behavior was used to select a common training duration of 100 epochs. All fine-tuning experiments were therefore run for 100 epochs, and the model obtained after the final epoch was used for evaluation. Model-specific default training hyperparameters were retained with a batch size of 8 used for all models. Models can be fine-tuned either by updating all pa- rameters (full-model regime) or by keeping the encoder weights frozen while updating only the remaining train- able components (frozen-encoder regime). In the full- model regime, all model parameters were trainable from the first epoch. In the frozen-encoder regime, the en- coder or backbone parameters were held fixed, while the model-specific decoder, prediction heads, and other des- ignated adaptation components remained trainable. The relative effectiveness of these strategies depends on the architecture, the size of the support set, and the de- gree of domain shift between the pretraining data and the target dataset. Freezing the encoder can preserve transferable pretrained representations and reduce the number of trainable parameters, but it may also restrict the modelâs ability to adapt its feature representation to the target domain. Previous studies have therefore reported differing outcomes for frozen-encoder and full- model adaptation [43, 44]. We evaluate both possibili- ties under comparable conditions. We train and evaluate each fine-tunable model in both regimes using exactly the same support set and test test. Fine-tuning was performed for 16 of the 19 model con- figurations. DeepCell NuclearSegmentation, Mesmer, and CellSAM were evaluated only in the zero-shot setting because their fine-tuning interfaces were not available to us. Information on the exact model checkpoint names and software versions are provided in the Supplementary material. 2.6.3 Repeated Experiments Each experiment was repeated ten times using seeds 0 to 9. For each seed, a tissue-stratified support set was sampled independently. Repeating the experiments across these support sets reduced dependence on a single support-set selection. For a given dataset and seed, the support set and test was kept the same across all the models and both fine-tuning regimes. The seeds were propagated to Python, NumPy, Py- Torch, and TensorFlow where applicable. Deterministic cuDNN execution was enabled and cuDNN benchmark- ing was disabled. These settings reduced uncontrolled variation from sampling, data ordering, augmentation, model initialization where applicable, and stochastic framework operations, although complete bitwise deter- minism cannot be guaranteed for every third-party im- plementation. 2.6.4 Computational Setup All experiments were conducted on the same compute node equipped with two NVIDIA Tesla V100 PCIe GPUs with 32 GB memory each, two Intel Xeon Gold 6226R processors at 2.90 GHz with 16 physical cores per pro- cessor, and 501 GiB of system memory. Only GPU 0 was used for the experiments, and no other GPU processes were running on the node. To reduce the effect of the GPUâs thermal state on the energy measurements, the GPU was allowed to cool to 45 ⊠C before each measurement. The same cooldown procedure was applied before each experiment. Energy consumption was then measured during the execution of the corresponding inference or fine-tuning workload using the procedure described in Section 2.4. 3 Results All primary sustainability-aware comparisons were con- ducted using the predefined SAPI weights α=1/2, ÎČ=1/3, and Îł=1/6, corresponding to a performanceâ energyâsize weighting ratio of (3:2:1). This weighting gives segmentation performance the greatest influence because predictive accuracy remains the primary require- ment for model utility, while assigning substantial im- portance to energy consumption and a smaller, but non- negligible, contribution to model size. The combined energy term was calculated using λ=0.8, so that fine- tuning energy consumption contributed 20% and post 9 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model mIF 10xGenomics_DAPI DAPI ssDNA 10xGenomics_HE HE Dataset 0.5480.1450.5360.5400.1790.4120.3680.5340.5230.5840.5160.427 0.6960.7520.6680.6980.6290.6310.7090.4530.7430.6410.6000.709 0.5620.6790.5810.5130.5560.5360.6560.5660.6540.6020.7120.647 0.6180.6810.5240.4640.5280.5050.5960.4050.6110.5380.6420.569 0.6160.6500.5810.6640.4900.6760.6940.6930.4620.6610.5740.6340.7020.7060.696 0.5610.3420.3580.4500.1170.4970.5900.6580.4980.6220.2010.5950.6410.6540.645 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Performance (P) Figure 3: Zero-shot segmentation performance across compatible modelâdataset combinations. Each cell reports the mean performance score P, calculated as the harmonic mean of AJI + , PQ, and NSD. Grey cells denote modelâdataset combinations that were not evaluated because the model is not intended to be used with that dataset modality. fine-tuning inference energy consumption contributed 80%. This choice represents a deployment-oriented set- ting in which inference is expected to be performed re- peatedly after adaptation, while still accounting for the computational cost incurred during fine-tuning. Unless otherwise stated, all SAPI rankings reported below use these fixed values. 3.1 Zero-shot inference We performed zero-shot inference to compare the mod- els without any adaptation to the CellBinDB datasets. The results showed clear differences in segmenta- tion performance, inference energy consumption, and sustainability-aware utility across models and datasets. Performance also depended strongly on the dataset, and the models with the highest segmentation scores did not always achieve the highest SAPI ZS values. We therefore present the results in four parts: predictive performance, inference energy consumption, the joint relationship be- tween performance, energy use, and model size, and the resulting SAPI ZS rankings. In the supplementary mate- rial we provide qualitative comparisons of the predicted masks produced by every compatible model for one sam- ple image from each dataset, together with the corre- sponding input image and ground-truth mask. 3.1.1 Performance Figure 3 shows the zero-shot performance score P for ev- ery compatible model-dataset combination. Across these combinations, P ranged from 0.117 for Mesmer on H&E to 0.752 for InstanSeg on 10xGenomics_DAPI, demon- strating substantial variation in zero-shot transferability across models and imaging modalities. The strongest model differed among the four fluores- cence datasets. InstanSeg achieved the highest perfor- mance on 10xGenomics_DAPI (P=0.752) and ssDNA (P=0.681), whereas CellposeSAM performed best on DAPI (P=0.712). CellSAM attained the highest score on mIF (P=0.584). The corresponding AJI + , PQ, and NSD metrics are reported in the Supplementary material. The highest H&E performance was obtained predominantly by histopathology-specific models. PathoSAM ViT-H achieved the highest score on both 10xGenomics_HE (P=0.706) and HE (P=0.654) datasets. 3.1.2 Inference energy consumption We measured the zero-shot inference energy used by the GPU, CPU, and RAM for each model. Figure 4 shows the mean energy consumed per test image, averaged across experimental seeds and the datasets compatible with each model. The total energy consumption varied widely across the model zoo. InstanSeg consumed the least energy, with a mean of 8.36 J per image. The two StarDist configurations also had relatively low energy requirements, consuming around 8 J per image. MEDIAR followed at 28.27 J per image. In contrast, CellSAM consumed the most energy at 437.65 J per image. CellposeSAM and HoVer- Net had similarly high energy requirements, consuming 410.91 and 410.72 J per image, respectively. CellSAM therefore consumed approximately 52 times more energy per image than InstanSeg. The contribution of each hardware component also dif- fered among models. CPU energy formed the largest component for several of the lower-energy models, in- cluding InstanSeg and both StarDist configurations. GPU energy dominated the total consumption of most of the higher-energy models, including CellposeSAM, HoVer-Net, and the PathoSAM variants. CellSAM was a notable exception: despite having the highest total energy consumption, most of its measured energy came from the CPU. RAM made the smallest contribution for every model. These results show that zero-shot inference energy differed not only in total magnitude but also in how the workload was distributed across the measured hardware components. 10 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0 100 200 300 400 Energy (Joules/image) Energy source RAM CPU GPU Figure 4: Mean zero-shot inference energy per test image, separated into GPU, CPU, and RAM contributions. Energy values were averaged across experimental seeds and the datasets compatible with each model. The total height of each stacked bar represents the total measured IT energy per image. 3.1.3 Performance vs Energy vs Model size We compared mean zero-shot performance with infer- ence energy and model size in Figure 5. Across the 19 model configurations, Spearman correlation analysis showed almost no monotonic association between per- formance and inference energy (Ï = â0.020) and only a weak positive association between performance and pa- rameter count (Ï = 0.197). In contrast, parameter count showed a strong positive association with inference en- ergy (Ï = 0.747). Thus, larger models generally required more energy for inference, but greater model size or en- ergy consumption was not consistently associated with higher zero-shot segmentation performance. Several relatively small and energy-efficient models achieved competitive performance. StarDist (Fluo), for example, reached a mean performance of 0.606 while us- ing only 18.47 J per image with approximately 1.4 mil- lion parameters. MEDIAR also combined strong perfor- mance with low energy use. It achieved a mean perfor- mance of 0.636 across all six datasets while consuming 28.27 J per image. In comparison, CellposeSAM con- sumed 410.91 J per image but achieved a lower mean performance of 0.616 over the same datasets. CellSAM and HoVer-Net also had high energy requirements with- out reaching the highest performance levels. These re- sults show that high inference energy consumption did not necessarily translate into better zero-shot segmenta- tion. Among the models evaluated on the two H&E datasets, the PathoSAM variants and the two Cel- lViT configurations achieved the highest mean perfor- mance. However, this pattern did not extend to all pathology-specific models, as HoVer-Net and StarDist (H&E) achieved lower mean performance values. The PathoSAM family provides a clear example of the trade-off between performance, energy consumption, and model size. PathoSAM ViT-B achieved a mean perfor- mance of 0.675 with 104.8 million parameters and con- sumed 164.15 J per image. PathoSAM ViT-H increased the mean performance only slightly to 0.680, while in- creasing the model size to 652.1 million parameters and the energy consumption to 342.71 J per image. Thus, the ViT-H variant used approximately twice as much en- ergy and more than six times as many parameters as the ViT-B variant for only a tiny increase in segmentation performance. 3.1.4 SAPI ZS We calculated SAPI ZS for each compatible model- dataset combination to assess segmentation performance together with inference energy consumption and model size. Figure 6a shows that the SAPI ZS score for each compatible model-dataset combination. InstanSeg achieved the highest SAPI ZS on four datasets: 10xGenomics_DAPI (0.4019), DAPI (0.3869), ssDNA (0.3890), and 10xGenomics_H&E (0.3729). 11 0100200300400 Energy per Image (Joules) 0.45 0.50 0.55 0.60 0.65 Performance (P) Cellpose cyto2 Cellpose cyto3 CellposeSAM CellSAM CellViT-256 CellViT-SAM-H DeepCell HoverNet InstanSeg MEDIAR Mesmer MicroSAM ViT-B MicroSAM ViT-L MicroSAM ViT-T PathoSAM ViT-B PathoSAM ViT-H PathoSAM ViT-L StarDist (Fluo) StarDist (H&E) Model Size (Parameters) 1M 10M 100M 500M Figure 5: Relationship between zero-shot segmentation performance, inference energy consumption, and model size. Each point represents one model. Performance and energy were averaged across ten runs and then across the datasets compatible with that model. The horizontal axis shows total inference energy per image on a logarithmic scale, and marker area represents the total number of model parameters. StarDist (Fluo) ranked first on mIF with a score of 0.3336, while StarDist (H&E) ranked first on the HE dataset with a score of 0.3404. These results show that SAPI ZS rankings differ from the rankings based on per- formance alone. Figure 6b compares the mean SAPI ZS across the four fluorescence datasets. StarDist (Fluo) achieved the high- est mean score at 0.3544, followed by MEDIAR at 0.3419 and InstanSeg at 0.3380. Although InstanSeg ranked first on three of the four fluorescence datasets, its sub- stantially lower score on mIF reduced its overall mean. MEDIAR achieved the highest mean segmentation per- formance across the fluorescence datasets, but StarDist (Fluo) ranked higher under SAPI ZS because it combined competitive performance with much lower energy con- sumption and a smaller model size. By contrast, Cell- poseSAM achieved strong segmentation performance but received a lower mean SAPI ZS of 0.3111 because of its greater energy and parameter requirements. A similar pattern appeared across the two H&E datasets (Figure 6c). StarDist (H&E) achieved the high- est mean SAPI ZS at 0.3495, followed by MEDIAR at 0.3426, CellViT-256 at 0.3392, and PathoSAM ViT-B at 0.3349. The larger PathoSAM variants achieved some of the highest mean segmentation scores on the H&E datasets, but their higher energy consumption and pa- rameter counts reduced their SAPI ZS values. For ex- ample, PathoSAM ViT-H achieved the highest mean H&E performance but obtained a lower mean SAPI ZS of 0.3242. Overall, SAPI ZS did not simply reproduce the per- formance ranking. It favoured models that maintained strong segmentation performance while using less energy and fewer parameters. The variation across datasets also shows that the most favourable model depended on the imaging modality and dataset characteristics. 3.2 Few-shot Adaptation For each dataset, we randomly selected one image from every tissue group to form the support set and used all the remaining images as the test set. We fine-tuned the model on the support set using two strategies i.e. full-model fine-tuning and frozen-encoder fine-tuning. In full-model fine-tuning, we updated all model parame- ters. In frozen-encoder fine-tuning, we kept the en- coder weights fixed and updated the remaining train- able parameters. We evaluated the zero-shot check- point and both fine-tuned variants on the same test set. We repeated the complete support-set selection, fine- tuning, and evaluation procedure with ten different ran- dom seeds to capture variation caused by support set and test set composition. 3.2.1 Performance Few-shot adaptation improved segmentation perfor- mance for most compatible model-dataset combina- tions (Figures 7â12). The largest gains generally oc- curred when the corresponding zero-shot model per- formed poorly. This pattern was particularly clear on HE, DAPI, and ssDNA, where adaptation substan- tially improved Cellpose, InstanSeg, Hover-Net, and Mi- croSAM models. By contrast, the gains were usually smaller on the two 10xGenomics datasets, where several 12 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model mIF 10xGenomics_DAPI DAPI ssDNA 10xGenomics_HE HE Dataset 0.330.170.320.320.180.280.250.300.310.300.280.26 0.380.400.360.370.330.340.350.280.370.320.310.34 0.340.390.340.320.310.320.340.320.350.310.330.33 0.360.390.320.300.300.310.320.270.340.290.320.31 0.360.370.330.360.300.340.360.340.280.350.300.320.330.330.34 0.340.270.260.290.140.280.310.330.290.340.180.300.320.320.31 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 SAPI ZS (a) Zero-shot Sustainability-Aware Performance Index (SAPI ZS ) across compatible model-dataset combinations. Each cell reports the mean score across ten runs. Grey cells indicate combinations that were not evaluated because the model is not intended to be used with the corresponding image modality. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer MicroSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L Model 0.28 0.30 0.32 0.34 0.36 Sustainable AI Performance Index ( SAPI ZS ) 0.3544 0.3380 0.3333 0.3256 0.2805 0.3126 0.3155 0.2902 0.3419 0.3027 0.3111 0.3087 Fluorescence-only All datasets (b) SAPI ZS scores average across ten runs and the four fluorescence datasets: mIF, DAPI, ssDNA and 10xGenomics_DAPI. Hatching pattern distinguishes fluorescence-only models from models evaluated on all the datasets. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 Mesmer HoverNet CellViT-256 PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.22 0.24 0.26 0.28 0.30 0.32 0.34 0.36 Sustainable AI Performance Index ( SAPI ZS ) 0.3495 0.3208 0.2963 0.3236 0.2236 0.3077 0.3392 0.3349 0.2855 0.3426 0.2372 0.3069 0.3273 0.3242 0.3287 H&E-only All datasets (c) SAPI ZS scores average across ten runs and the two H&E datasets: HE and 10xGenomics_HE. Hatching pattern distinguishes H&E-only models from models evaluated on all the datasets. Figure 6: Zero-shot sustainability-aware performance across models and datasets. 13 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.2 0.3 0.4 0.5 0.6 Performance (P) Zero-shotFull-modelFrozen-encoder Figure 7: Segmentation performance P on the mIF dataset under zero-shot inference, full-model fine-tuning and frozen-encoder fine-tuning. The plot reports the mean value of P across ten runs for each compatible model with error bars indicating standard deviation. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.60 0.65 0.70 0.75 Performance (P) Zero-shotFull-modelFrozen-encoder Figure 8: Segmentation performance P on the 10xGenomics_DAPI dataset under zero-shot inference, full-model fine-tuning and frozen-encoder fine-tuning. The plot reports the mean value of P across ten runs for each compatible model with error bars indicating standard deviation. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.50 0.55 0.60 0.65 0.70 0.75 Performance (P) Zero-shotFull-modelFrozen-encoder Figure 9: Segmentation performance P on the DAPI dataset under zero-shot inference, full-model fine-tuning and frozen-encoder fine-tuning. The plot reports the mean value of P across ten runs for each compatible model with error bars indicating standard deviation. 14 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.45 0.50 0.55 0.60 0.65 0.70 Performance (P) Zero-shotFull-modelFrozen-encoder Figure 10: Segmentation performance P on the ssDNA dataset under zero-shot inference, full-model fine-tuning and frozen-encoder fine-tuning. The plot reports the mean value of P across ten runs for each compatible model with error bars indicating standard deviation. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.575 0.600 0.625 0.650 0.675 0.700 0.725 0.750 Performance (P) Zero-shotFull-modelFrozen-encoder Figure 11: Segmentation performance P on the 10xGenomics_HE dataset under zero-shot inference, full-model fine- tuning and frozen-encoder fine-tuning. The plot reports the mean value of P across ten runs for each compatible model with error bars indicating standard deviation. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 Performance (P) Zero-shotFull-modelFrozen-encoder Figure 12: Segmentation performance P on the HE dataset under zero-shot inference, full-model fine-tuning and frozen-encoder fine-tuning. The plot reports the mean value of P across ten runs for each compatible model with error bars indicating standard deviation. 15 models already achieved relatively strong zero-shot per- formance. Full-model fine-tuning generally produced stronger performance than frozen-encoder fine-tuning. Updating the encoder allowed the model to adjust its learned rep- resentation to the target images, which was especially useful for models that transferred poorly in the zero-shot setting. However, frozen-encoder fine-tuning sometimes remained close to full-model fine-tuning and occasionally performed slightly better. For example, freezing the en- coder improved the results of CellViT-256 on both the H&E datasets. InstanSeg on mIF shows a more pro- nounced exception, as the frozen-encoder strategy out- performed full-model fine-tuning considerably. The response to adaptation also differed across datasets. Performance on mIF remained lower and more variable than on the other datasets, although StarDist (Fluo) retained strong performance and sev- eral general-purpose models improved after fine-tuning. DAPI and ssDNA showed more consistent gains, partic- ularly for Cellpose cyto2 and Cellpose cyto3. On the H&E datasets, adaptation substantially improved mod- els with weak zero-shot transfer, including InstanSeg and the Cellpose cyto models. CellposeSAM and ME- DIAR achieved strong adapted performance on several datasets, while the preferred model still depended on the image modality and dataset. Adaptation did not improve every model. Cellpose cyto3 showed little benefit on mIF, while InstanSeg on 10xGenomics_DAPI and CellViT-256 on 10xGe- nomics_HE performed slightly worse than their zero- shot baselines. While the performance of most mod- els is stable across the ten runs, CellposeSAM with frozen-encoder finetuning shows greater variance in per- formance on most datasets. Also, InstanSeg and ME- DIAR have larger performance variations for the mIF dataset. These exceptions show that few-shot adaptation depends on both the target dataset and the pretrained model. They also show that updating all the parameters does not always provide the most stable or better result. 3.2.2 Energy Consumption Fine-tuning energy E ft varied considerably across models (Figure 13). The figure reports the energy consumed per support image, averaged across the datasets compatible with each model and across the repeated runs. Cell- pose cyto2 and cyto3 required the least energy under both adaptation strategies. InstanSeg, MEDIAR, and StarDist also remained among the lower-energy models. In contrast, the larger MicroSAM, PathoSAM, and Cel- lViT variants occupied the high-energy end of the com- parison. PathoSAM ViT-H consumed the most energy during full-model fine-tuning. Freezing the encoder during fine-tuning reduced the training energy E ft for all models. However, the amount of the reduction depended strongly on the model. The largest reductions appeared for CellposeSAM, CellViT- SAM-H, Hover-Net, and PathoSAM ViT-H. Freezing the encoder also produced clear savings for the larger Mi- croSAM and PathoSAM variants. By contrast, the dif- ference between the two fine-tuning strategies was small for StarDist, Cellpose, InstanSeg, MEDIAR, CellViT- 256, and MicroSAM ViT-T. Across all the models, GPU and CPU activity ac- counted for the largest shares of fine-tuning energy (E ft ), while RAM contributed relatively little. Larger models generally consumed more fine-tuning energy, but model size alone did not explain the observed differences. Mod- els with similar parameter counts sometimes showed dis- tinct energy profiles, and encoder freezing produced dif- ferent levels of savings across architectures. Fine-tuning energy therefore reflected both model scale and the com- putational structure of the adaptation process. 3.2.3 SAPI FT+I The SAPI FT+I results show clear differences across mod- els and datasets (Figure 14). Smaller models that combined strong adapted performance with low fine- tuning and inference energy generally received the high- est scores. As a result, the sustainability-aware ranking did not always follow the performance ranking. Several large models achieved competitive segmentation perfor- mance after adaptation, but their higher energy con- sumption and parameter counts reduced their SAPI FT+I scores. Across the fluorescence datasets, InstanSeg achieved the highest average SAPI FT+I . Cellpose cyto2, Cellpose cyto3, and MEDIAR followed closely. StarDist (Fluo) also remained competitive. The heatmap shows that these rankings were not consistent across all fluorescence datasets. InstanSeg performed strongly on DAPI, s- DNA, and 10xGenomics_DAPI, but obtained a much lower score on mIF. The mIF dataset also produced lower scores for several MicroSAM variants, showing that the sustainability-aware benefit of adaptation remained dataset-dependent. A similar pattern appeared on the two H&E datasets. InstanSeg obtained the highest average score, followed by MEDIAR. Cellpose cyto2, Cellpose cyto3, and StarDist (H&E) also achieved strong results. In contrast, the larger PathoSAM and CellViT variants generally re- ceived lower scores, despite their competitive adapted performance because their additional energy and model- complexity costs were not fully offset by their perfor- mance gains. Fine-tuning also changed the relative SAPI ranking of several models. In the zero-shot setting, StarDist (Fluo) and MEDIAR ranked among the strongest models be- cause they combined competitive segmentation perfor- mance with low inference energy and small model size. After few-shot adaptation, InstanSeg moved to the high- est average SAPI FT+I score across both fluorescence and H&E datasets, while Cellpose cyto2 and cyto3 also im- proved their relative positions. This shift reflects the effect of adaptation on both segmentation performance and the total energy cost considered by SAPI FT+I . Mod- els that gained substantially in segmentation perfor- mance could therefore improve their sustainability-aware ranking, whereas models with high adaptation or infer- ence costs remained disadvantaged despite competitive adapted performance. The results show that the most favourable model under zero-shot inference is not nec- essarily the most favourable model after target-specific adaptation. 16 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000 8,000 Training energy per image (Joules) Energy source / Mode RAM CPU GPU Full Model Frozen Encoder Figure 13: Mean fine-tuning energy per support image under full-model and frozen-encoder adaptation. The total bar height represents the combined GPU, CPU, and RAM energy consumption averaged across compatible datasets. Overall, the results favor models that achieve a suit- able balance between segmentation performance, adap- tation cost, inference cost, and model complexity. No model achieved the highest score on every dataset, and some models showed markedly different results across image modalities. Nevertheless, InstanSeg remains the most sustainable model on an average across all the datasets, followed by MEDIAR and then the two Cell- pose cyto variants. The choice of model should therefore reflect both the target dataset and the sustainability re- quirements of the intended deployment scenario. 4 Discussion This study examined whether the segmentation gains from large pre-trained and foundation models were pro- portionate to their computational requirements. We evaluated 19 model configurations across six CellBinDB datasets and considered segmentation performance, in- ference energy, fine-tuning energy, and model complex- ity. Three findings define the main narrative of the study. First, increasing model size did not lead to a consistent or proportional improvement in segmentation performance. Second, large foundation models often achieved a marginally higher performance, but they re- quired order of magnitude more energy and trainable parameters. Third, several compact models remained competitive and provided a better balance between seg- mentation quality and computational cost. These results show that the model with the highest segmentation score is not necessarily the most suitable model for resource constrained deployment scenarios. The main issue is therefore proportionality. A larger computational burden is reasonable when it produces a clear and practically important improvement. How- ever, a marginal performance gain may not justify a large increase in inference energy, adaptation cost, or model complexity. The decision depends on the intended application. A marginal gain may be important when it prevents a consequential segmentation error, but it may have limited value when competing models already meet the required performance level. These findings sup- port the Green AI principle that predictive quality and computational efficiency should be evaluated together [12, 15]. Our study extends this principle to cell and nuclear instance segmentation through a controlled com- parison of predictive performance, energy use, and model complexity. 4.1 Model scale did not determine seg- mentation performance The results do not support a simple relationship between parameter count and segmentation quality. Several large pre-trained and foundation models performed strongly, but their advantage varied across datasets and adapta- tion settings. Smaller architectures also remained com- petitive across multiple datasets. Model size alone there- fore provided limited information about the performance that a model would achieve on a specific cell or nuclear segmentation task. This result reflects the demands of instance segmen- tation in biomedical images. A model must sepa- rate densely packed and touching objects, preserve fine boundaries, and remain robust to changes in staining, tissue structure, image contrast, object density, and imaging modality. Generic visual representations do not automatically satisfy these requirements. Performance also depends on the correspondence between the pre- training domain and the target images, the model ar- chitecture, the instance-separation method, the infer- ence procedure, and the model-specific pre-processing and post-processing steps. Large-scale pre-training can still provide useful rep- 17 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model mIF 10xGenomics_DAPI DAPI ssDNA 10xGenomics_HE HE Dataset 0.340.260.330.330.270.290.310.300.28 0.380.410.390.390.340.350.390.330.35 0.350.410.380.380.340.350.390.330.34 0.360.400.370.370.330.330.380.330.31 0.370.390.380.380.340.370.340.380.330.340.330.35 0.350.370.350.350.330.350.340.370.280.330.320.33 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 SAPI FT + I (a) SAPI FT+I for each compatible model-dataset combination. Grey cells indicate combinations that were not evaluated because the model is not intended to be used with the corresponding image modality. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.32 0.33 0.34 0.35 0.36 0.37 Sustainability-Aware Performance Index 0.3550 0.3704 0.3679 0.3672 0.3222 0.3284 0.3672 0.3231 0.3191 Fluorescence-only All datasets (b) Mean SAPI FT+I across the four fluorescence datasets: mIF, DAPI, ssDNA, and 10xGenomics_DAPI. Hatching distinguishes fluorescence-specific models from models evaluated on all datasets. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.30 0.32 0.34 0.36 0.38 Sustainability-Aware Performance Index 0.3637 0.3820 0.3644 0.3678 0.3312 0.3557 0.3391 0.3714 0.3058 0.3315 0.3253 0.3425 H&E-only All datasets (c) Mean SAPI FT+I across the HE and 10xGenomics_HE datasets. Hatching distinguishes H&E-specific models from models evaluated on all datasets. Figure 14: Sustainability-aware performance after few-shot adaptation. The results combine segmentation perfor- mance with fine-tuning and post-fine-tuning inference energy and model size. Scores are averaged across ten indepen- dent runs. 18 resentations. Foundation models may transfer across datasets and can support tasks that are not measured in this benchmark, including prompted segmentation and reuse across several applications [8, 45]. Their broader capabilities may justify their size when a single shared model replaces several specialized systems or supports a wide range of workloads. However, these benefits should be demonstrated for the intended deployment setting. For a fixed and repetitive segmentation task, a special- ized model may provide a better balance when the ad- ditional capabilities of a general-purpose model remain unused. The results also show why parameter count and mea- sured energy should not be treated as equivalent mea- sures. Parameter count describes model scale and influ- ences hardware requirements, but it does not fully deter- mine computational cost. Energy use also depends on activation sizes, operator types, memory transfers, nu- merical precision, image tiling, prompt generation, post- processing, framework overhead, and hardware utiliza- tion. A model with fewer parameters may therefore con- sume more energy than its size suggests. Conversely, a larger architecture may use the hardware more efficiently. Direct energy measurements provide stronger evidence for deployment decisions than parameter count alone. Energy and complexity should therefore be reported as related but distinct properties. Large-scale pre-training often improve cell and nuclear segmentation. However, the additional performance was usually accompanied by greater inference energy and a much larger parameter footprint. The relevant question is therefore not only whether a foundation model per- forms better, but whether the size of the improvement justifies the additional recurring costs. This distinction is especially important for inference. Fine-tuning normally occurs a limited number of times, whereas inference may be repeated over thousands or millions of images. A difference that appears small at the level of one image can become substantial in a high-volume pipeline. Repeated inference affects energy consumption, processing time, accelerator availability, infrastructure requirements, cooling demand, and the number of analyses that can run concurrently. These ef- fects are relevant to digital pathology, spatial biology, flu- orescence microscopy, and high-content screening, where segmentation is often applied to large image collections. The practical importance of the performance gain must therefore guide model selection. A more computa- tionally demanding model may be justified when it cor- rects errors that affect downstream measurements, bi- ological conclusions, or clinical interpretation. In con- trast, the lower-cost model is usually more preferable when two models produce results that are sufficiently similar for the intended task. This does not mean that energy should take priority over segmentation quality. Candidate models should first meet a task-specific per- formance requirement. Models that satisfy this require- ment can then be compared using inference energy, run- time, memory demand, parameter count, maintainabil- ity, and hardware accessibility. This approach avoids two problematic choices. It pre- vents the selection of an inadequate model only because it consumes less energy. It also prevents a marginal per- formance gain from automatically outweighing a much larger computational burden. The appropriate weight will differ between applications and should reflect the consequences of segmentation errors in the downstream workflow. 4.2 Few-shot adaptation should be justi- fied by its incremental benefit The comparison between zero-shot inference, frozen- encoder adaptation, and full-model fine-tuning shows that adaptation should be treated as a deployment in- vestment. Fine-tuning can improve the alignment be- tween a pre-trained model and a target dataset. How- ever, the gain should be evaluated against the required annotation effort, training energy, memory demand, and implementation complexity. Full-model fine-tuning is difficult to justify when it provides only a tiny improve- ment over zero-shot inference or frozen-encoder adapta- tion. Frozen-encoder adaptation can provide a useful com- promise. It updates only part of the model and can reduce the computational and memory requirements of training. It may work well when the pre-trained en- coder already provides suitable representations for the target images. However, freezing the encoder can re- strict adaptation when the target modality differs sub- stantially from the pre-training domain. The relative value of frozen and full adaptation therefore depends on both the model and the dataset. Our results do not es- tablish one adaptation strategy as universally superior. The controlled few-shot protocol is important for this interpretation. Each model received the same tissue- stratified support set and the same training budget. This design allowed us to compare how efficiently different models used a limited support set. It did not attempt to find the maximum possible performance of every archi- tecture. Model-specific tuning of the learning rate, aug- mentation, training duration, or support-set size could improve some results. However, such optimization would require additional computation and would reduce com- parability between models. The benchmark instead ad- dresses a practical question: under the same limited an- notation and adaptation budget, which models convert these resources into useful performance gains most effec- tively? The value of adaptation also depends on how often the resulting model will be used. An expensive fine- tuning procedure may be reasonable when it produces a meaningful improvement and the adapted model is then applied to a large number of images. The same proce- dure may be less attractive when models must be re- trained frequently for different sites, staining protocols, tissues, or imaging devices. Adaptation and inference costs should therefore be interpreted in relation to the expected deployment volume. The SAPI FT+I formulation makes this distinction ex- plicit by combining fine-tuning and post-fine-tuning in- ference energy. The parameter λ controls their relative contribution. A larger value gives more importance to recurring inference, while a smaller value gives greater importance to adaptation. This weighting represents an assumed deployment pattern rather than a complete life- 19 cycle model. It should therefore be selected and reported according to the expected use of the adapted model. 4.3 SAPI makes the performance- resource trade-off explicit No single measurement was sufficient to identify the most appropriate model. A ranking based only on segmenta- tion performance favors the most accurate models with- out considering the resources they require. A ranking based only on energy or parameter count favors compact models even when their segmentation quality is insuf- ficient. The joint analysis of performance, energy, and model size showed that models occupied different posi- tions across these dimensions. The Sustainability-Aware Performance Index (SAPI) addresses this problem by combining the three dimen- sions in a transparent manner. Under the default weights, (α=1/2, ÎČ=1/3, and Îł=1/6), performance, energy, and complexity have a relative importance of (3:2:1). Performance receives the greatest weight be- cause segmentation quality remains the primary require- ment. Energy receives the second-largest weight because inference can create a recurring operational cost. Param- eter count receives the smallest weight but still captures relevant constraints related to storage, memory, model transfer, and access to suitable hardware. This weighting does not reward low energy use at the expense of inadequate segmentation. Instead, it assesses whether the achieved performance is proportionate to the required energy and model complexity. A compact model can rank favorably when it achieves competitive performance at substantially lower cost. A larger model can also rank highly when its performance advantage is sufficiently large to compensate for its greater resource consumption. SAPI does not define a universal and permanent or- dering of models. Its ranking depends on the priorities expressed through the weights. A high-throughput labo- ratory may assign more importance to inference energy. An application with severe consequences for segmenta- tion errors may place greater emphasis on performance. A resource-constrained setting may give additional im- portance to model complexity. A change in ranking un- der different weights is therefore not a failure of the SAPI index. Instead, it shows that the preferred model de- pends on the intended use case. 4.4 Energy consumption and implica- tions for operational carbon emis- sions Energy consumption has direct implications for the op- erational carbon emissions associated with computation. Electricity-related operational emissions depend on the energy consumed and the carbon intensity of the electric- ity supplying the computation. Under the same electric- ity and data-center conditions, a proportional increase in energy consumption therefore produces a proportional increase in operational emissions. This makes energy effi- ciency an important consideration when selecting models for deployment. The implications become more substantial as compu- tational workloads increase. A small difference in energy per image can accumulate when a model processes thou- sands or millions of images or is deployed continuously across multiple sites. This is particularly relevant for high-throughput applications such as digital pathology, biopharmaceutical screening, and large-scale microscopy analysis. When models provide only marginal perfor- mance gains but require substantially more computa- tion, the additional energy demand can translate into a corresponding increase in electricity related operational emissions. Fine-tuning and repeated model adaptation add further computational demand and should therefore also be considered when evaluating the overall resource requirements of a model. Carbon-aware practices can reduce the emissions as- sociated with a given computational workload by us- ing electricity with lower carbon intensity or shifting computation to periods when cleaner electricity is avail- able. These measures can complement improvements in model and workflow efficiency, which directly reduce the amount of computation required. The energy results re- ported in this study should, however, not be interpreted as a complete environmental assessment. They quantify computational demand within the defined experimen- tal setting and do not capture embodied emissions from hardware and infrastructure or the energy and emissions associated with pre-training externally developed mod- els. Our results, therefore, provide an important indi- cator of the potential operational carbon implications of model choice. 4.5 Implications for sustainable biomed- ical image segmentation Our findings have direct implications for the design and deployment of biomedical image-segmentation systems. Cell and nuclear segmentation often forms an early stage of a larger analytical pipeline. Its outputs may be used for feature extraction, cell counting, spatial analysis, population statistics, phenotypic profiling, treatment- response modeling, or quality control. The value of a performance improvement should therefore be judged by its effect on these downstream analyses rather than by the segmentation score alone. This consideration is particularly important in high- content screening, drug discovery, toxicology, spatial bi- ology, and large microscopy studies. These applications may process large numbers of images across many ex- perimental conditions. A small increase in per-image en- ergy or runtime can accumulate over the full workload. A more expensive model may also reduce throughput or require additional accelerators. The resulting infras- tructure burden may limit the number of samples, com- pounds, tissues, or experimental conditions that can be analyzed within a fixed budget. Our results indicate that researchers should not se- lect the largest available model by default. They should first define the minimum segmentation quality required by the downstream task. Among models that meet this requirement, they should prefer the model that provides the most favorable balance of inference energy, adapta- tion energy, runtime, memory use, parameter count, and 20 maintenance effort. This approach makes computational cost part of the model-selection process. Model developers should also report efficiency mea- surements alongside segmentation results. Parameter count alone is not sufficient. Studies should report measured inference energy, runtime, memory use, hard- ware, numerical precision, image size, tiling strategy, and relevant pre-processing and post-processing bound- aries. Fine-tuning studies should additionally report the training budget and the energy required for adaptation. These details are necessary for making informed deci- sions about using the model. Adaptation strategies should follow the same princi- ple. Full-model fine-tuning should not be treated as the automatic choice. Frozen-encoder adaptation may be preferable when it achieves a similar improvement at lower energy cost. Zero-shot inference may remain the most suitable option when adaptation provides marginal benefit. The selected strategy should depend on the in- cremental performance gain, the frequency of retraining, and the expected number of future inferences. Our findings are also relevant to shared research infras- tructures and institutions with limited access to high- memory accelerators. A compact model that runs on widely available hardware may support broader repro- duction and adoption, even when its performance is marginally below that of the largest model. Compu- tational accessibility affects which laboratories can use a model, the size of the studies they can conduct, and their ability to reproduce published results. Efficiency is therefore connected not only to environmental sustain- ability but also to scientific accessibility. A foundation model may still provide the best system- level solution when it supports several modalities or tasks and removes the need to maintain separate specialized models. In that case, the relevant comparison should include the full anticipated workload. The cost of one shared model should be compared with the combined cost and maintenance burden of the systems it replaces. Our benchmark does not directly measure this multi-task value, but it shows that broad capability should not be assumed to compensate automatically for high recurring cost. For clinical and translational applications, computa- tional efficiency can influence processing time, local hard- ware requirements, service capacity, and maintainability. However, this study did not evaluate clinical decision- making, patient outcomes, regulatory requirements, or prospective workflow performance. Our results concern computational and operational suitability. Any clinical application would still require external validation, de- tailed failure analysis, robustness testing, and prospec- tive evaluation. The broader implication is that sustainable model selection does not mean selecting the model with the lowest energy consumption. It means using computa- tional resources in proportion to the value they provide. The most suitable model is one that meets the required performance standard without imposing an unnecessary computational burden and adaptation cost. 4.6 Limitations Several limitations affect the interpretation of our re- sults. First, we performed the measurements on one fixed GPU and CPU configuration. This choice controlled hardware variation and supported consistent compar- isons. However, absolute energy consumption may differ across accelerators, numerical precisions, software envi- ronments, and deployment platforms. Relative efficiency may also change when an architecture is better opti- mized for a different hardware system. The reported energy values are therefore empirical measurements for the evaluated configuration rather than fixed properties of the model architectures. Second, the benchmark covers operational energy within defined inference and fine-tuning boundaries. It does not include the energy consumed during the orig- inal pre-training of the published checkpoints. Reli- able checkpoint-specific pre-training information was not available for all models. Third, the fixed adaptation protocol was designed to enable controlled comparison across models rather than maximize the performance of each model individually. The use of limited support sets, a fixed training du- ration, and no validation-based early stopping ensures that all models are evaluated under the same adaptation conditions, but may favor models that adapt efficiently within these constraints. Model-specific optimization, such as using additional support images, longer train- ing, different learning-rate schedules, or model-specific augmentation, could improve the performance of indi- vidual models. However, such optimization would intro- duce additional sources of variation, increase computa- tional cost, and make direct comparisons across models less controlled. Fourth, the six CellBinDB subsets cover several imag- ing modalities and tissue strata, but they do not repre- sent the full diversity of microscopy and pathology data. Differences in specimen preparation, staining, imaging devices, acquisition settings, artifacts, and disease com- position may affect model performance. External evalu- ation on independently collected datasets is needed be- fore extending the reported rankings to other settings. Undocumented overlap between model pre-training data and the source data underlying a benchmark may also be difficult to exclude for externally developed checkpoints. Fifth, for the zero-shot inference experiments, we eval- uated model-native automatic inference without dataset- specific threshold optimization or manual prompting. This protocol supports scalable and consistent compar- ison, but it may not show the maximum performance of interactive or prompt-dependent models. The conclu- sions therefore apply to automatic segmentation under the evaluated settings. They should not be directly ex- tended to interactive workflows in which expert prompts or manual corrections form part of the inference process. Sixth, parameter count represents only one aspect of model complexity. It does not directly describe activa- tion memory, operator efficiency, dependency manage- ment, loading time, software stability, or maintenance effort. The measured energy and memory values provide complementary information, but a complete deployment analysis would also require throughput testing, reliabil- 21 ity assessment, hardware cost, and integration with the intended analytical system. Finally, SAPI contains explicit choices. The weights assigned to performance, energy, and parameter count, and the value of λ affect the resulting ranking. The in- dex is configurable so that these choices can reflect dif- ferent deployment priorities. However, every application should state and justify its selected weights. SAPI rank- ings should always be reported together with its compo- nent weights. 5 Conclusion This study presents a sustainability-aware benchmark of 19 pretrained cell and nuclear instance segmenta- tion models across six CellBinDB datasets. We evalu- ated zero-shot inference and few-shot adaptation using segmentation performance, operational energy consump- tion, and model size. Our results show that these fac- tors vary substantially across models and datasets, and that higher performance does not always justify greater computational cost. Model selection should therefore consider image modality, adaptation needs, deployment scale, and resource use rather than segmentation accu- racy alone. We introduced SAPI, a novel measure, to provide a transparent way to examine the trade-offs between per- formance, energy use, and model complexity. The ex- plicit weightage assigned to each component of SAPI al- lows researchers and practitioners to adjust the weighting according to their priorities, providing a practical frame- work for selecting cell segmentation models that are both effective and efficient. CRediT authorship contribution statement Eiram Mahera Sheikh: Conceptualization, Methodol- ogy, Software, Formal analysis, Investigation, Visualiza- tion, Writing - Original Draft. Alaa Tharwat: Writing - Review & Editing. Wolfram Schenck: Supervision, Funding acquisition, Writing - Review & Editing Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this pa- per. Acknowledgments This research was conducted within the framework of the project âSAIL: SustAInable Lifecycle of Intelligent So- cioTechnical Systemsâ (grant no. NW21-059B). SAIL is receiving funding from the programme âNetzwerke 2021â, an initiative of the Ministry of Culture and Science of the State of North Rhine-Westphalia, Germany. The sole re- sponsibility for the content of this publication lies with the authors. We thank Christoph Ostrau for reviewing the manuscript and providing valuable comments and suggestions that contributed to its improvement. Data availability The code developed for this study is publicly available at https://github.com/eiram-mahera/sapi. The Cell- BinDB datasets analyzed in this study are publicly avail- able through Zenodo at https://doi.org/10.5281/ zenodo.15370205. Supplementary data Supplementary material associated with this article is provided in Appendix A. References [1] Jonathan Mitchel, Teng Gao, Viktor Petukhov, Eli Cole, and Peter V Kharchenko. Impact and cor- rection of segmentation errors in spatial transcrip- tomics. Nature Genetics, pages 1â11, 2026. doi: 10.1038/s41588-025-02497-4. [2] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234â241. Springer, 2015. doi: 10.1007/978-3-319-24574-4_28. [3] Uwe Schmidt, Martin Weigert, Coleman Broad- dus, and Gene Myers. Cell detection with star- convex polygons. In International conference on medical image computing and computer-assisted in- tervention, pages 265â273. Springer, 2018. doi: 10.1007/978-3-030-00934-2_30. [4] Carsen Stringer, Tim Wang, Michalis Michae- los, and Marius Pachitariu. Cellpose: a gen- eralist algorithm for cellular segmentation. Na- ture methods, 18(1):100â106, 2021. doi: 10.1038/ s41592-020-01018-x. [5] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi- tissue histology images. Medical image analysis, 58: 101563, 2019. doi: 10.1016/j.media.2019.101563. [6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. doi: 10.48550/arXiv.2010. 11929. [7] Fabian Hörst, Moritz Rempe, Lukas Heine, Con- stantin Seibold, Julius Keyl, Giulia Baldini, Selma Ugurel, Jens Siveke, Barbara GrĂŒnwald, Jan Eg- ger, et al. Cellvit: Vision transformers for precise 22 cell segmentation and classification. Medical im- age analysis, 94:103143, 2024. doi: 10.1016/j.media. 2024.103143. [8] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan- Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on com- puter vision, pages 4015â4026, 2023. doi: 10.48550/ arXiv.2304.02643. [9] Anwai Archit, Luca Freckmann, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Carolin Teuber, Melanie Spitzner, Con- stanza Tapia Contreras, et al. Segment anything for microscopy. Nature methods, 22(3):579â591, 2025. doi: 10.1038/s41592-024-02580-4. [10] Can Shi, Jinghong Fan, Zhonghan Deng, Huanlin Liu, Qiang Kang, Yumei Li, Jing Guo, Jingwen Wang, Jinjiang Gong, Sha Liao, Ao Chen, Ying Zhang, and Mei Li. CellBinDB: a large-scale multi- modal annotated dataset for cell segmentation with benchmarking of universal models. GigaScience, 14: 1â15, 2025. doi: 10.1093/gigascience/giaf069. [11] Eiram Mahera Sheikh, Alaa Tharwat, Constanze Schwan, and Wolfram Schenck. Cgmd: Centrality- guided maximum diversity for annotation-efficient fine-tuning of pretrained cell segmentation models. In 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), pages 1â5. IEEE, 2026. doi: 10.1109/ISBI61048.2026.11515602. [12] Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. Green ai. Commun. ACM, 63(12): 54â63, November 2020. ISSN 0001-0782. doi: 10. 1145/3381831. URL https://doi.org/10.1145/ 3381831. [13] Emma Strubell, Ananya Ganesh, and Andrew Mc- Callum. Energy and policy considerations for deep learning in NLP. In Anna Korhonen, David Traum, and LluĂs MĂ rquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computa- tional Linguistics, pages 3645â3650, Florence, Italy, July 2019. Association for Computational Linguis- tics. doi: 10.18653/v1/P19-1355. [14] David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. Carbon emissions and large neural network training. arXiv preprint arXiv:2104.10350, 2021. doi: 10.48550/ arXiv.2104.10350. [15] Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau. To- wards the systematic reporting of the energy and carbon footprints of machine learning. Jour- nal of Machine Learning Research, 21(248):1â 43, 2020. URL http://jmlr.org/papers/v21/ 20-312.html. [16] Arnout C Ruifrok, Dennis A Johnston, et al. Quan- tification of histochemical staining by color decon- volution. Analytical and quantitative cytology and histology, 23(4):291â299, 2001. [17] Marius Pachitariu, Michael Rariden, and Carsen Stringer. Cellpose-sam: superhuman generalization for cellular segmentation. BioRxiv, pages 2025â04, 2025. doi: 10.1101/2025.04.28.651001. [18] Gihun Lee, SangMook Kim, Joonkee Kim, and Se- Young. Yun. Mediar: Harmony of data-centric and model-centric for multi-modality microscopy. In Jun Ma, Ronald Xie, Anubha Gupta, JosĂ© Guil- herme de Almeida, Gary D. Bader, and Bo Wang, editors, Proceedings of The Cell Segmentation Chal- lenge in Multi-modality High-Resolution Microscopy Images, volume 212 of Proceedings of Machine Learning Research, pages 1â16. PMLR, 28 Novâ 09 Dec 2023. URL https://proceedings.mlr. press/v212/lee23a.html. [19] Noah F Greenwald, Geneva Miller, Erick Moen, Alex Kong, Adam Kagel, Thomas Dougherty, Christine Camacho Fullaway, Brianna J McIn- tosh, Ke Xuan Leow, Morgan Sarah Schwartz, et al. Whole-cell segmentation of tissue im- ages with human-level performance using large- scale data annotation and deep learning. Nature biotechnology, 40(4):555â565, 2022. doi: 10.1038/ s41587-021-01094-0. [20] David A Van Valen, Takamasa Kudo, Keara M Lane, Derek N Macklin, Nicolas T Quach, Mialy M DeFelice, Inbal Maayan, Yu Tanouchi, Euan A Ash- ley, and Markus W Covert. Deep learning automates the quantitative analysis of individual cells in live- cell imaging experiments. PLoS computational bi- ology, 12(11):e1005177, 2016. doi: 10.1371/journal. pcbi.1005177. [21] Thibaut Goldsborough, Ben Philps, Alan OâCallaghan, Fiona Inglis, Leo Leplat, Andrew Filby, Hakan Bilen, and Peter Bankhead. Instanseg: an embedding-based instance segmentation algo- rithm optimized for accurate, efficient and portable cell segmentation. arXiv preprint arXiv:2408.15954, 2024. doi: 10.48550/arXiv.2408.15954. [22] Markus Marks, Uriah Israel, Rohit Dilip, Qilin Li, Changhua Yu, Emily Laubscher, Ahamed Iqbal, Elora Pradhan, Ada Ates, Martin Abt, et al. Cell- sam: a foundation model for cell segmentation. Nature Methods, pages 1â9, 2025. doi: 10.1038/ s41592-025-02879-w. [23] Titus Griebel, Anwai Archit, and Constantin Pape. Segment anything for histopathology. arXiv preprint arXiv:2502.00408, 2025. doi: 10.48550/ arXiv.2502.00408. [24] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr DollĂĄr. Panoptic seg- mentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogni- tion (CVPR), pages 9404â9413, 2019. doi: 10.1109/ CVPR.2019.00963. 23 [25] Stanislav Nikolov, Sam Blackwell, Alexei Zverovitch, Ruheena Mendes, Michelle Livne, Jeffrey De Fauw, Yojan Patel, Clemens Meyer, Harry Askham, Bernardino Romera-Paredes, Christopher Kelly, Alan Karthikesalingam, Carlton Chu, Dawn Carnell, Cheng Boon, Derek DâSouza, Syed Ali Moinuddin, Bethany Garie, Yasmin McQuinlan, Sarah Ireland, Kiarna Hampton, Krystle Fuller, Hugh Montgomery, Geraint Rees, Mustafa Suleyman, Trevor Back, CĂan Owen Hughes, Joseph R Ledsam, and Olaf Ronneberger. Clinically applicable segmentation of head and neck anatomy for radiotherapy: deep learning algorithm development and validation study. Journal of Medical Internet Research, 23(7):e26151, 2021. doi: 10.2196/26151. [26] Annika Reinke, Minu D Tizabi, Michael Baum- gartner, Matthias Eisenmann, Doreen Heckmann- Nötzel, A Emre Kavur, Tim RĂ€dsch, Carole H Su- dre, Laura Acion, Michela Antonelli, et al. Un- derstanding metric-related pitfalls in image analy- sis validation. Nature methods, 21(2):182â194, 2024. doi: 10.1038/s41592-023-02150-0. [27] Lena Maier-Hein, Annika Reinke, Patrick Go- dau, Minu D Tizabi, Florian Buettner, Evangelia Christodoulou, Ben Glocker, Fabian Isensee, Jens Kleesiek, Michal Kozubek, et al. Metrics reloaded: recommendations for image analysis validation. Na- ture methods, 21(2):195â212, 2024. doi: 10.1038/ s41592-023-02151-z. [28] NVIDIA Corporation. nvidia-ml-py: Python bind- ings for the nvidia management library, 2026. URL https://pypi.org/project/nvidia-ml-py/. [software]. [29] NVIDIA Corporation. NVML API Reference Guide. NVIDIA Corporation, 2026. URL https: //docs.nvidia.com/deploy/nvml-api/. [30] Zhenheng Tang, Yuxin Wang, Qiang Wang, and Xi- aowen Chu. The impact of gpu dvfs on the energy and performance of deep learning: An empirical study. In Proceedings of the Tenth ACM Interna- tional Conference on Future Energy Systems, pages 315â325, 2019. doi: 10.1145/3307772.3328315. [31] Muhammad Fahad, Arsalan Shahid, Ravi Reddy Manumachu, and Alexey Lastovetsky. A compar- ative study of methods for measurement of energy of computing. Energies, 12(11):2204, 2019. doi: 10.3390/en12112204. [32] Marcus HĂ€hnel, Björn Döbel, Marcus Völp, and Hermann HĂ€rtig. Measuring energy consumption for short code paths using rapl. ACM SIGMET- RICS Performance Evaluation Review, 40(3):13â17, 2012. doi: 10.1145/2425248.2425252. [33] Kashif Nizam Khan, Mikael Hirki, Tapio Niemi, Jukka K Nurminen, and Zhonghong Ou. Rapl in ac- tion: Experiences in using rapl for power measure- ments. ACM Transactions on Modeling and Per- formance Evaluation of Computing Systems (TOM- PECS), 3(2):1â26, 2018. doi: 10.1145/3177754. [34] Yehia Arafa, Ammar ElWazir, Abdelrahman ElKa- nishy, Youssef Aly, Ayatelrahman Elsayed, Abdel- Hameed Badawy, Gopinath Chennupati, Stephan Eidenbenz, and Nandakishore Santhi. Verified instruction-level energy consumption measurement for nvidia gpus. In Proceedings of the 17th ACM International Conference on Computing Frontiers, pages 60â70, 2020. doi: 10.1145/3387902.3392613. [35] Alexander Wong. Netscore: towards universal met- rics for large-scale performance analysis of deep neural networks for practical on-device edge us- age. In International Conference on Image Analysis and Recognition, pages 15â26. Springer, 2019. doi: 10.1007/978-3-030-27272-2_2. [36] Shreyank N Gowda, Xinyue Hao, Gen Li, Shashank Narayana Gowda, Xiaobo Jin, and Laura Sevilla-Lara. Watt for what: Rethinking deep learn- ingâs energy-performance relationship. In European Conference on Computer Vision, pages 388â405. Springer, 2024. doi: 10.1007/978-3-031-92089-9_ 24. [37] Andrew A Chien, Liuzixuan Lin, Hai Nguyen, Var- sha Rao, Tristan Sharma, and Rajini Wijayawar- dana. Reducing the carbon impact of generative ai inference (today and in 2035). In Proceedings of the 2nd workshop on sustainable computer systems, pages 1â7, 2023. doi: 10.1145/3604930.3605705. [38] Amine Lbath and Ibtissam Labriji. Energy effi- ciency in ai for 5g and beyond: A deeprx case study. In 2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), pages 1151â1156. IEEE, 2024. doi: 10. 1109/EuCNC/6GSummit60053.2024.10597065. [39] Shih-Kai Chou, Jernej Hribar, Vid HanĆŸel, Mihael MohorÄiÄ, and Carolina Fortuna. The energy cost of artificial intelligence lifecycle in communication net- works. IEEE Journal on Selected Areas in Commu- nications, 2025. doi: 10.1109/JSAC.2025.3642835. [40] Marius Pachitariu and Carsen Stringer. Cell- pose 2.0: how to train your own model. Nature methods, 19(12):1634â1641, 2022. doi: 10.1038/ s41592-022-01663-4. [41] Youssef Dawoud, Arij Bouazizi, Katharina Ernst, Gustavo Carneiro, and Vasileios Belagiannis. Know- ing what to label for few shot microscopy image cell segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3568â3577, 2023. doi: arXiv:2211. 10244v1. [42] Fabian Englbrecht, Iris E Ruider, and Andreas R Bausch. Automatic image annotation for fluorescent cell nuclei segmentation. PloS one, 16(4):e0250093, 2021. doi: 10.1371/journal.pone.0250093. [43] Weiyi Xie, Nathalie Willems, Shubham Patil, Yang Li, and Mayank Kumar. Sam fewshot finetuning for anatomical segmentation in medical images. In Proceedings of the IEEE/CVF winter conference 24 on applications of computer vision, pages 3253â 3261, 2024. URL https://openaccess.thecvf. com/content/WACV2024/html/Xie_SAM_Fewshot_ Finetuning_for_Anatomical_Segmentation_in_ Medical_Images_WACV_2024_paper.html. [44] Hanxue Gu, Haoyu Dong, Jichen Yang, and Ma- ciej A Mazurowski. How to build the best med- ical image segmentation algorithm using founda- tion models: a comprehensive empirical study with segment anything model. arXiv preprint arXiv:2404.09957, 2024. doi: arXiv:2404.09957v3. [45] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature communications, 15(1):654, 2024. 25 Table S1: Models, pretrained checkpoints, and software versions used in the benchmark. Model nameSoftware nameSoftware versionCheckpoint name StarDist (Fluo)stardist0.9.22D_versatile_fluo StarDist (HE)stardist0.9.22D_versatile_he InstanSeginstanseg-torch0.1.2single_channel_nuclei Cellpose cyto2cellpose3.1.1cyto2torch_0 Cellpose cyto3cellpose3.1.1cyto3 Cellpose-SAMcellpose4.0.8cpsam DeepCellDeepCell0.12.10NuclearSegmentation-8.tar.gz MesmerDeepCell0.12.10MultiplexSegmentation-9.tar.gz CellSAMDeepCell0.12.10cellsam-models_v1.2.tar.gz MicroSAM ViT-T micro-sam1.7.7vit_t_lm MicroSAM ViT-B micro-sam1.7.7vit_b_lm MicroSAM ViT-L micro-sam1.7.7vit_l_lm PathoSAM ViT-B micro-sam1.7.7vit_b_histopathology PathoSAM ViT-L micro-sam1.7.7vit_l_histopathology PathoSAM ViT-H micro-sam1.7.7vit_h_histopathology HoverNet github.com/vqdang/hover_net. git main branchhovernet_original_kumar_notype_ pytorch.tar MEDIAR github.com/Lee-Gihun/MEDIARmain branchfrom_phase2.pth CellViT-256 github.com/TIO-IKIM/CellViTmain branchCellViT-256-x20.pth CellViT-SAM-H github.com/TIO-IKIM/CellViTmain branchCellViT-SAM-H-x20.pth Appendix A Supplementary Material This supplementary material provides additional information on model and software, together with detailed quanti- tative and qualitative results. The experimental protocol, metric definitions, energy measurement procedure, carbon- emission calculation, and formulation of the Sustainable AI Performance Index are described in the main manuscript. Unless stated otherwise, quantitative results report the mean across ten runs. Qualitative zero-shot examples use the experiment associated with random seed 0. A.1 Model and Software Details Table S1 lists every evaluated model in the benchmark, the exact pre-trained checkpoint used to initialize it, and the software version used to run the experiments. For software installed directly from a source repository, the table reports the corresponding Git commit instead of a package version. The exact inference and training configurations and the code developed for this study are publicly available at https://github.com/eiram-mahera/sapi. A.2 Zero-Shot Evaluation A.2.1 Performance Metrics We evaluated each pre-trained model without dataset-specific adaptation. Each model was assessed only on datasets compatible with its intended image modality. The main manuscript reports the composite performance score (P), whereas Figures S1, S2 and S3 provide the corresponding AJI+, Panoptic Quality (PQ), and Normalized Surface Dice (NSD) results. A.2.2 Qualitative Predictions To provide a qualitative comparison of the segmentation outputs, we selected one test image from each dataset using a fixed random seed of 0. For each selected image, the corresponding supplementary figure (Figures S4, S5, S6, S7, S8 and S9) presents the original image, the ground-truth instance mask, an overlay of the ground-truth instance mask on the original image, and the predictions generated by all applicable models, overlaid on the original image. A.3 Few-Shot Adaptation We grouped the images in a dataset by tissue and sampled one image from every tissue group to form the support set for fine-tuning. The remaining images formed the test set. We first evaluated each compatible model on this test set before adaptation and then evaluated it again after fine-tuning on the support set. Fine-tuning was performed for 100 epochs under two regimes: full-model fine-tuning, in which all the model parameters were updated, and frozen-encoder fine-tuning, in which the encoder weights were not updated during training. We repeated the complete procedure ten times using random seeds 0â9. 26 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model mIF 10xGenomics_DAPI DAPI ssDNA 10xGenomics_HE HE Dataset 0.5600.1090.6180.6240.2080.4620.3900.5330.5340.5920.6400.456 0.6840.7350.6840.7020.6250.6810.6930.5370.7260.6750.6600.700 0.5340.6490.5710.5200.5220.5520.6280.5500.6260.6170.6910.618 0.5990.6560.5320.5030.5150.5460.5860.4470.6010.5620.6370.557 0.5850.6240.5830.6520.5330.6510.6670.6690.4770.6430.6000.6300.6800.6820.672 0.5280.3490.3480.4300.1710.4410.5470.6190.4880.5850.3480.5600.6030.6150.610 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 AJI+ Figure S1: Zero-shot AJI+ scores across compatible modelâdataset combinations. Grey cells denote modelâdataset combinations that were not evaluated because the model is not intended to be used with that image modality. StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model mIF 10xGenomics_DAPI DAPI ssDNA 10xGenomics_HE HE Dataset 0.4620.1360.5660.5630.1690.2920.3610.4290.4840.5350.6060.423 0.6130.6660.6060.6280.5460.5900.6290.3900.6630.5690.5610.633 0.4660.5860.5000.4310.4790.4660.5660.4610.5710.5240.6300.558 0.5170.5790.4240.3730.4380.4230.4930.3030.5130.4400.5480.472 0.5740.5790.5320.6100.4430.6110.6200.6310.3920.6020.5300.5790.6400.6440.624 0.4890.2480.2880.3750.0740.4340.5190.5740.3940.5410.1250.5180.5550.5680.554 0.0 0.1 0.2 0.3 0.4 0.5 0.6 Panoptic Quality (PQ) Figure S2: Zero-shot Panoptic Quality (PQ) scores across compatible modelâdataset combinations. Grey cells denote modelâdataset combinations that were not evaluated because the model is not intended to be used with that image modality. 27 StarDist (Fluo) StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T Mesmer HoverNet CellViT-256 MicroSAM ViT-B PathoSAM ViT-B DeepCell MEDIAR CellSAM CellposeSAM MicroSAM ViT-L PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model mIF 10xGenomics_DAPI DAPI ssDNA 10xGenomics_HE HE Dataset 0.6540.2390.4520.4590.1650.5930.3540.7080.5550.6330.3850.406 0.8210.8860.7270.7800.7470.6280.8360.4560.8670.6930.5890.818 0.7580.8560.7090.6210.7150.6120.8250.7610.8040.6870.8470.813 0.8000.8650.6720.5590.6870.5740.7720.5340.7720.6570.7830.737 0.7050.7790.6360.7440.5020.7930.8260.8010.5410.7570.5960.7040.8080.8120.820 0.7140.5340.4950.5950.1630.6810.7540.8320.6940.7930.2480.7530.8200.8320.827 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Normalised Surface Dice (NSD) Figure S3: Zero-shot Normalized Surface Dice (NSD) scores across compatible modelâdataset combinations. Grey cells denote modelâdataset combinations that were not evaluated because the model is not intended to be used with that image modality. A.3.1 Aggregated Jaccard Index Plus (AJI + ) Figures S10, S11, S12, S13, S14, and S15 visualize the AJI + results for each dataset. A.3.2 Panoptic Quality (PQ) Figures S16, S17, S18, S19, S20, and S21 visualize the PQ metrics for each dataset. A.3.3 Normalized Surface Dice (NSD) Figures S22, S23, S24, S25, S26, and S27 visualize the NSD metrics for each dataset. 28 Original Ground Truth Ground Truth Overlay Cellpose cyto2Cellpose cyto3 CellSAMDeepCell InstanSeg MEDIARMesmerMicroSAM ViT-B MicroSAM ViT-LMicroSAM ViT-TStarDist (Fluo) Figure S4: Qualitative zero-shot comparison for one image sampled from the mIF dataset using random seed 0. The figure shows the input image, the ground-truth instance mask, the ground-truth overlay, and predictions from all models compatible with the image modality. Each prediction is overlaid on the input image. 29 Original Ground Truth Ground Truth Overlay Cellpose cyto2Cellpose cyto3 CellSAMDeepCell InstanSeg MEDIARMesmerMicroSAM ViT-B MicroSAM ViT-LMicroSAM ViT-TStarDist (Fluo) Figure S5: Qualitative zero-shot comparison for one image sampled from the DAPI dataset using random seed 0. The figure shows the input image, the ground-truth instance mask, the ground-truth overlay, and predictions from all models compatible with the image modality. Each prediction is overlaid on the input image. 30 Original Ground Truth Ground Truth Overlay Cellpose cyto2Cellpose cyto3 CellSAMDeepCell InstanSeg MEDIARMesmerMicroSAM ViT-B MicroSAM ViT-LMicroSAM ViT-TStarDist (Fluo) Figure S6: Qualitative zero-shot comparison for one image sampled from the ssDNA dataset using random seed 0. The figure shows the input image, the ground-truth instance mask, the ground-truth overlay, and predictions from all models compatible with the image modality. Each prediction is overlaid on the input image. 31 Original Ground Truth Ground Truth Overlay Cellpose cyto2Cellpose cyto3 CellSAMDeepCell InstanSeg MEDIARMesmerMicroSAM ViT-B MicroSAM ViT-LMicroSAM ViT-TStarDist (Fluo) Figure S7: Qualitative zero-shot comparison for one image sampled from the 10xGenomics_DAPI dataset using random seed 0. The figure shows the input image, the ground-truth instance mask, the ground-truth overlay, and predictions from all models compatible with the image modality. Each prediction is overlaid on the input image. 32 Original Ground Truth Ground Truth Overlay Cellpose cyto2Cellpose cyto3 CellSAMCellViT-256 CellViT-SAM-HDeepCellHoverNet InstanSeg MEDIARMesmerPathoSAM ViT-BPathoSAM ViT-H PathoSAM ViT-LStarDist (H&E) Figure S8: Qualitative zero-shot comparison for one image sampled from the 10xGenomics_HE dataset using random seed 0. The figure shows the input image, the ground-truth instance mask, the ground-truth overlay, and predictions from all models compatible with the image modality. Each prediction is overlaid on the input image. 33 Original Ground Truth Ground Truth Overlay Cellpose cyto2Cellpose cyto3 CellSAMCellViT-256 CellViT-SAM-HDeepCellHoverNet InstanSeg MEDIARMesmerPathoSAM ViT-BPathoSAM ViT-H PathoSAM ViT-LStarDist (H&E) Figure S9: Qualitative zero-shot comparison for one image sampled from the HE dataset using random seed 0. The figure shows the input image, the ground-truth instance mask, the ground-truth overlay, and predictions from all models compatible with the image modality. Each prediction is overlaid on the input image. 34 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.1 0.2 0.3 0.4 0.5 0.6 0.7 AJI+ Zero-shotFull-modelFrozen-encoder Figure S10: AJI + results for few-shot adaptation on the mIF dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.600 0.625 0.650 0.675 0.700 0.725 0.750 0.775 AJI+ Zero-shotFull-modelFrozen-encoder Figure S11: AJI + results for few-shot adaptation on the 10xGenomics_DAPI dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.50 0.55 0.60 0.65 0.70 AJI+ Zero-shotFull-modelFrozen-encoder Figure S12: AJI + results for few-shot adaptation on the DAPI dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. 35 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.50 0.55 0.60 0.65 0.70 AJI+ Zero-shotFull-modelFrozen-encoder Figure S13: AJI + results for few-shot adaptation on the ssDNA dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.575 0.600 0.625 0.650 0.675 0.700 0.725 AJI+ Zero-shotFull-modelFrozen-encoder Figure S14: AJI + results for few-shot adaptation on the 10xGenomics_HE dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 AJI+ Zero-shotFull-modelFrozen-encoder Figure S15: AJI + results for few-shot adaptation on the HE dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. 36 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.1 0.2 0.3 0.4 0.5 0.6 Panoptic Quality (PQ) Zero-shotFull-modelFrozen-encoder Figure S16: PQ results for few-shot adaptation on the mIF dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.525 0.550 0.575 0.600 0.625 0.650 0.675 0.700 Panoptic Quality (PQ) Zero-shotFull-modelFrozen-encoder Figure S17: PQ results for few-shot adaptation on the 10xGenomics_DAPI dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.45 0.50 0.55 0.60 0.65 Panoptic Quality (PQ) Zero-shotFull-modelFrozen-encoder Figure S18: PQ results for few-shot adaptation on the DAPI dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. 37 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.35 0.40 0.45 0.50 0.55 0.60 0.65 Panoptic Quality (PQ) Zero-shotFull-modelFrozen-encoder Figure S19: PQ results for few-shot adaptation on the ssDNA dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.525 0.550 0.575 0.600 0.625 0.650 0.675 0.700 Panoptic Quality (PQ) Zero-shotFull-modelFrozen-encoder Figure S20: PQ results for few-shot adaptation on the 10xGenomics_HE dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 Panoptic Quality (PQ) Zero-shotFull-modelFrozen-encoder Figure S21: PQ results for few-shot adaptation on the HE dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. 38 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.2 0.3 0.4 0.5 0.6 Normalised Surface Dice (NSD) Zero-shotFull-modelFrozen-encoder Figure S22: NSD results for few-shot adaptation on the mIF dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.60 0.65 0.70 0.75 0.80 0.85 0.90 Normalised Surface Dice (NSD) Zero-shotFull-modelFrozen-encoder Figure S23: NSD results for few-shot adaptation on the 10xGenomics_DAPI dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.60 0.65 0.70 0.75 0.80 0.85 0.90 Normalised Surface Dice (NSD) Zero-shotFull-modelFrozen-encoder Figure S24: NSD results for few-shot adaptation on the DAPI dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. 39 StarDist (Fluo) InstanSeg Cellpose cyto2Cellpose cyto3 MicroSAM ViT-T MicroSAM ViT-B MEDIAR CellposeSAM MicroSAM ViT-L Model 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 Normalised Surface Dice (NSD) Zero-shotFull-modelFrozen-encoder Figure S25: NSD results for few-shot adaptation on the ssDNA dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.65 0.70 0.75 0.80 0.85 Normalised Surface Dice (NSD) Zero-shotFull-modelFrozen-encoder Figure S26: NSD results for few-shot adaptation on the 10xGenomics_HE dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. StarDist (H&E) InstanSeg Cellpose cyto2Cellpose cyto3 HoverNet CellViT-256 PathoSAM ViT-B MEDIAR CellposeSAM PathoSAM ViT-L PathoSAM ViT-H CellViT-SAM-H Model 0.5 0.6 0.7 0.8 0.9 Normalised Surface Dice (NSD) Zero-shotFull-modelFrozen-encoder Figure S27: NSD results for few-shot adaptation on the HE dataset. The figure compares compatible, fine-tunable models under zero-shot inference, frozen-encoder and full-model fine-tuning. Each value reports the mean across ten runs. 40