Paper deep dive
CLoPA: Continual Low Parameter Adaptation of Interactive Segmentation for Medical Image Annotation
Parhom Esmaeili, Chayanin Tangwiriyasakul, Eli Gibson, Sebastien Ourselin, M. Jorge Cardoso
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/13/2026, 12:19:24 AM
Summary
CLoPA is a continual learning strategy for interactive medical image segmentation that fine-tunes a small subset of parameters (instance normalization and shallow convolution kernels) in foundation models like nnInteractive. It operates within existing annotation workflows, requiring no new parameters or inference changes, and significantly improves performance to expert-level across diverse Medical Segmentation Decathlon tasks.
Entities (5)
Relation Signals (3)
CLoPA → adapts → nnInteractive
confidence 100% · We propose CLoPA, a continual adaptation strategy that tunes a small fraction of nnInteractive's parameters
CLoPA → evaluatedon → Medical Segmentation Decathlon
confidence 95% · Across eight Medical Segmentation Decathlon tasks spanning diverse anatomical targets
nnInteractive → benchmarkedagainst → nnU-Net
confidence 90% · Expert-level thresholds are defined as the task-specific mean Dice of nnU-Net
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Interactive segmentation enables clinicians to guide annotation, but existing zero-shot models like nnInteractive fail to consistently reach expert-level performance across diverse medical imaging tasks. Because annotation campaigns produce a growing stream of task-specific labelled data, online adaptation of the segmentation model is a natural complement to zero-shot inference. We propose CLoPA, a continual adaptation strategy that tunes a small fraction of nnInteractive's parameters on the annotation cache, triggered by lightweight episode scheduling. CLoPA requires no new parameters or changes to the inference pipeline, and operates entirely within the existing annotation workflow. Across eight Medical Segmentation Decathlon tasks spanning diverse anatomical targets and imaging characteristics, CLoPA rapidly elevates performance to expert-level, even for tasks where nnInteractive previously failed, with the majority of gains realised after a single training episode. We show that the benefits of tuning different parameter groups depends on task characteristics and data regimes. Also, that for targets with complex geometries (e.g., hepatic vessels), instance normalisation and low-level feature tuning saturates, suggesting a need for deeper feature-representation alignment in the most challenging scenarios.
Tags
Links
- Source: https://arxiv.org/abs/2603.06426v1
- Canonical: https://arxiv.org/abs/2603.06426v1
Trouble viewing inline? Open PDF directly →
Full Text
27,226 characters extracted from source content.
Expand or collapse full text
CLoPA: Continual Low-Parameter Adaptation of Interactive Segmentation for Medical Image Annotation Parhom Esmaeili 1 , Chayanin Tangwiriyasakul 1 , Eli Gibson 2 , Sebastien Ourselin 1 , and M. Jorge Cardoso 1 1 School of Biomedical Engineering and Imaging Sciences, KCL, UK 2 Siemens Healthineers, Princeton NJ, USA parhom.esmaeili@kcl.ac.uk Abstract.Interactive segmentation enables clinicians to guide annota- tion, but existing zero-shot models like nnInteractive fail to consistently reach expert-level performance across diverse medical imaging tasks. Be- cause annotation campaigns produce a growing stream of task-specific labelled data, online adaptation of the segmentation model is a natu- ral complement to zero-shot inference. We propose CLoPA, a contin- ual adaptation strategy that tunes a small fraction of nnInteractive’s parameters on the annotation cache, triggered by lightweight episode scheduling. CLoPA requires no new parameters or changes to the in- ference pipeline, and operates entirely within the existing annotation workflow. Across eight Medical Segmentation Decathlon tasks spanning diverse anatomical targets and imaging characteristics, CLoPA rapidly elevates performance to expert-level, even for tasks where nnInteractive previously failed, with the majority of gains realised after a single train- ing episode. We show that the benefits of tuning different parameter groups depends on task characteristics and data regimes. Also, that for targets with complex geometries (e.g., hepatic vessels), instance normali- sation and low-level feature tuning saturates, suggesting a need for deeper feature-representation alignment in the most challenging scenarios. Keywords:Interactive segmentation, continual learning, medical image analysis, dataset annotation, parameter-efficient fine-tuning 1 Introduction Large-scale annotated datasets in medical imaging are bottlenecked by data- sharing restrictions [10] and the cost of manual segmentation [11]. Domain shifts across medical centres further limit the applicability of static pre-trained seg- mentation models like nnU-Net [6] for annotation. Interactive segmentation ad- dresses this by letting clinicians guide the process through prompts such as clicks, scribbles, or bounding boxes [7]. Among recent models, nnInteractive [7] achieves strong zero-shot generalisation across diverse anatomies. Yet even nnIn- teractive cannot consistently reach expert-level performance with low-effort click arXiv:2603.06426v1 [cs.CV] 6 Mar 2026 2P. Esmaeili et al. Fig. 1.Overview of CLoPA. As annotated samples are produced, the annotation cache grows and periodically triggers training episodes that fine-tune only a small subset of nnInteractive’s parameters, progressively improving segmentation quality. 3D-Image ViewGold-Standard CLoPA nnInteractive prompts across all tasks [5], making zero-shot models impractical for large-scale annotation campaigns where both speed and reliability are essential. Because annotation is inherently repetitive, the annotated samples produced during a campaign constitute a growing task-specific dataset that can be ex- ploited for online model adaptation. This motivates a continual learning ap- proach: rather than relying solely on the zero-shot model, we progressively fine- tune it on the annotation stream to close the gap to expert-level performance. Crucially, adaptation must remain lightweight—both to avoid overfitting on the small, growing cache and to preserve the strong initialisation from pre-training. In this work, we make the following contributions: (1) we propose CLoPA, a continual adaptation strategy that fine-tunes only a small subset parameters of a base foundation model (e.g., nnInteractive) during the annotation workflow, requiring no new parameters and no changes to inference; (2) we demonstrate across eight MSD tasks that this lightweight adaptation rapidly achieves expert- level performance, including on tasks where the base model previously failed; and (3) we extend the interactive segmentation evaluation protocol of Esmaeili et al. [5] with trajectory metrics that capture adaptation dynamics over time. 1.1 Related Work Interactive segmentation algorithms incorporate image data and user guided prompts (e.g., clicks, scribbles, bounding boxes or lassos [7]) to generate output segmentations on prompted targets. Current state-of-the-art methods employ deep learning to design these algorithms; and during training networks are typ- ically trained with a synthetic user-in-the-loop to iteratively prompt from the error-region between a reference annotation and a prediction [3, 4, 15, 12, 17, 7]. Zero-shot models, intended for general out-of-the-box use, generate training ex- amples by pooling huge quantities of different segmentation datasets [15, 7, 3, 4] to sample image-annotation pairs. Training data diversity is introduced by pseudo-labelling strategies: but classical methods like SLIC [1] produce coarse Title Suppressed Due to Excessive Length3 parcellations, while SAM-based projection [9, 12] used by nnInteractive yields mostly blobby instances due to the domain gap from natural images [7]. Existing methods address prompt ambiguity with multiple candidate masks [9, 12], composite training annotations [7], or more restrictive prompts such as bounding boxes and lassos [4, 7]. However, restrictive prompts are laborious for complex geometries, and no current method reliably achieves expert level per- formance using only low-effort click interactions [5]. Additionally, fixed patch sizes and zooming heuristics [4, 7] used for volumetric inference are suboptimal for targets with low volume fractions (sparse, branching structures), where CNN based methods particularly struggle [7]. In summary, zero-shot models lack the inductive biases to consistently reach expert-level performance, making them unsuited for large-scale annotation. Since annotation is inherently repetitive, adaptation to the task context is a natural complement. In-context learning approaches [16] plateau below specialist perfor- mance and do not scale with dataset size. Continual parameter tuning has been explored in 2D [18], but to our knowledge has not been extended to volumetric interactive segmentation. 2 Methodology We consider a practical data-annotation scenario: a single, fixed binary seg- mentation task where annotated samples are cached as they are produced. In this work, we adapt a base interactive-segmentation foundation model, namely nnInteractive [7], on a stream of incoming annotations using full memory re- play, without modifying the inference pipeline. CLoPA has two components: (1)whento trigger adaptation (training episodes), and (2)howto adapt. A training episode is triggered once the annotation cache contains at least 25% of the dataset (k D = 0.25) and at least 5 unassigned samples (fromk M = 0.2, i.e., ≥1/k M samples to allow for a validation split). 2.1 Training Configurations Tunable parameters:We freeze all pretrained weights bar two configurations. 1)unfreezing only instance normalisation [14] affine parameters (scale and bias); we denote this configurationCLoPA-I.Nin all tables.2)tuning all instance normalisation parameters and the convolution kernels in first stage of the U-Net’s [13] encoder and the last stage of the decoder (segmentation layers); we denote this configurationCLoPA-C.N. Instance normalisation modulates per-channel feature statistics independently for each sample, making it a natural fit for task- specific style and contrast adaptation without altering the learned spatial filters. Because these parameters constitute a tiny fraction of the total model (<0.01%), the risk of catastrophic forgetting or overfitting on the small annotation cache is minimal, while still providing capacity to recalibrate feature distributions for the target anatomy. Convolution kernels determine the learned feature represen- tations; making it a natural fit for more challenging target anatomies. Tuning at shallow depths allows for aligning low-level features while avoiding overfitting. 4P. Esmaeili et al. Data synthesis:Following nnU-Net [6], we sample192 3 patches uniformly across foreground and background classes from the annotation cache, apply- ing nnInteractive’s preprocessing and augmentation pipeline. For each gradient step, we simulate click-based interaction: one foreground and one background point sampled from false-negative regions per interaction step. We currently employ fine-tuning with only click-prompting as this is the least laborious form of user-interaction. For training we use an unweighted Dice Cross-Entropy loss [8] averaged across interaction time-steps, same as Wong et al. [17]. For a maximum number of interaction time-steps,Nour loss is: L= 1 N N ∑ i=1 L Dice (m i , y) +L CE (m i , y)(1) Wherem i indicates the softmaxed prediction for the batch at interaction-step i. Each loss term is averaged across batch samples, if a batch sample reaches a termination condition (Dice Score of 1) then it is not accounted for on subsequent iteration time-steps in the loss calculation. For each gradient update we employ N= 5interaction steps per gradient update and a fixed initial batch size of 2. For each training episode, an initial learning rate of1e −3 is used for all parameters being fine-tuned with an ADAM optimiser. We train with a fixed quantity of10 epochs, and for each epoch, we perform a fixed quantity of50gradient updates. 3 Experiments Tasks:We extend the four axes of task complexity examined by Esmaeili et al. [5] to the following (conducting evaluations in the native image spaces):(i)Vari- ation in image voxel count (i.e. volume size),(i)image spacing/anisotropy,(i) target geometry (spherical vs irregular vs jagged targets occupying low-volume fractions),(iv)target size(v)target detectability and(vi)training dataset size. In line with Esmaeili et al. [5] we chose binary semantic segmentation tasks from the Medical Segmentation Decathlon (MSD) [2]. For multi-sequence datasets, the sequence that best visualises the target was used. Prompting:As with Esmaeili et al. [5], all simulations use one point per itera- tion per class randomly sampled from false-negative foreground and background regions. Simulations consist of interactive initialisation and 100 editing steps. Model Evaluation:We extend the evaluation metrics of Esmaeili et al. [5]. On a per-sample basis, we compute Dice and Normalised Surface Dice (NSD) with MSD [2] tolerances at each interaction step, as well as interaction-count- normalised AUCs for both metrics (nAUC). Following Esmaeili et al. [5], we estimate the number of interactions to expert-level performance (NoI) on a per- sample Dice basis, normalised by the maximum interaction count to obtain a percentage (nNoI). Expert-level thresholds are defined as the task-specific mean Dice of nnU-Net [6] trained on the full training set with optimal configuration selection. All expected metric values are reported as dataset-wide means. The percentage of samples that did not reach the performance target (NoF) is also Title Suppressed Due to Excessive Length5 Table 1.Task characteristics. All tasks are binary semantic segmentation from the MSD [2]. Dataset sizes are post-split. TaskSequence Key characteristicsN HippocampusT1 Small volume, fine detail130 Brain tumour core T2w Irregular, ambiguous boundaries240 PancreasCT Large, blobby140 LiverCT Very large, blobby, easy detection65 ProstateT2w Spherical, highly anisotropic16 Lung lesionCT Small target in large volume31 Hepatic vesselsCT Sparse, branching, low volume fraction 151 Colon cancerCT Hard to detect63 reported. For adaptation studies, we extend the evaluation by tracking the trajec- tory of each metric’s expected performanceE[m j (t)]as a function of dataset size t, evaluating after each training episode. We report AUCs over these trajectories (trajectory AUCs). Significance rankings follow the MSD protocol [2] with pair- wise algorithm comparisons: Wilcoxon signed-rank tests for all metrics, except episodic NoF where a McNemar test is used since comparisons are between bi- nary outcomes. For trajectory measures, pairing is performed along data-sample index rather than test samples. Significance is reported at theα= 0.05level. Data Splitting and Evaluation Runs:MSD training datasets were split 50- 50 into a train and holdout set. For each training run the adaptive algorithms iterate through the training set; with the sequence of the training data stream permuted across 3 runs, to obtain 3 runs of episodic checkpoints. Inference is also performed on 3 runs to simulate prompting stochasticity, with the corresponding training-run model checkpoints. For static models we use the same checkpoint across inference runs. For episodic performance measures we first average metrics across runs before reporting statistics. For trajectory measures, we obtain the trajectories of expected performances and then average across training runs, as training is not generally triggered synchronously across runs. 4 Results and Discussion Table 2 compares nnInteractive, CLoPA-I.N and CLoPa-C.N after the final train- ing episode (approximately nnU-Net-equivalent quantity of training data). Tasks where the base model converges:For tasks where nnInteractive al- ready achieves consistent convergence to nnU-Net-level performance (NoF≲5%: liver, prostate, pancreas, lung lesion, colon cancer)—typically blobby targets or those where detectability is the main challenge—CLoPA-I.N and CLoPA C.N maintains convergence rates (except pancreas, though not at a statistically sig- nificant rate) while improving all other metrics. Initialisation Dice and NSD are substantially higher, indicating that task-alignment accelerates annotation effi- ciency, especially for large targets like liver where adaptation provides sufficient context to trigger nnInteractive’s auto-zoom mechanism with minimal prompt- 6P. Esmaeili et al. Table 2.Final-episode performance. CLoPA-I.N: instance normalisation (I.N) only, CLoPA-C.N: Instance-normalisation and shallow depth convolution kernel tuning. All metrics report means except NoF (percentage of failed samples). Bold indicates first- ranked method per metric and task according to significance rankings. TaskAlgorithm DiceNSDNoI Init.Iter. 100nAUCInit.Iter. 100nAUCnNoINoF Brain Tumour nnInteractive0.5050.7530.7420.5830.9220.90429.624.4 CLoPA-I.N0.6820.8150.8170.8120.9680.96314.812.4 CLoPA-C.N0.6940.8110.8140.8270.9680.96415.413.2 Liver nnInteractive0.3730.9700.9630.3840.9920.9842218.2 CLoPA-I.N0.9120.9740.9730.9100.9940.99211.76.06 CLoPA-C.N0.8480.9710.9690.8480.9910.98815.910.6 Hippocampus nnInteractive0.5850.8050.7880.6310.8900.8709997.7 CLoPA-I.N0.8750.9030.9020.9570.9840.98345.640 CLoPA-C.N0.8760.9110.9090.9570.9880.9873831.5 Prostate nnInteractive0.7740.9220.9210.8150.9850.9834.60 CLoPA-I.N0.8820.9350.9320.9410.9920.9901.80 CLoPA-C.N0.8870.9360.9320.9450.9940.99120 Lung Lesion nnInteractive0.6980.8490.8520.7430.9400.93620 CLoPA-I.N0.7640.8610.8560.8190.9500.9421.50 CLoPA-C.N0.7620.8580.8550.8200.9460.9411.60 Pancreas nnInteractive0.4540.8950.8810.5380.9830.9719.82.13 CLoPA-I.N0.6680.8970.8880.7600.9850.9798.32.84 CLoPA-C.N0.6870.8950.8870.7810.9840.9788.52.84 Hepatic Vessels nnInteractive0.1130.1650.2090.1220.2320.28984.782.9 CLoPA-I.N0.5110.6980.6870.6820.8710.85619.611.8 CLoPA-C.N0.5210.6920.6810.6930.8640.8492314.5 Colon Cancer nnInteractive0.4750.7280.7250.5200.7880.7902.40 CLoPA-I.N0.6260.8190.8130.7210.9130.9111.20 CLoPA-C.N0.6150.7810.7790.7160.8740.87420 ing. Final-iteration Dice and NSD, as well as nAUC metrics, also improve— particularly for tasks where zero-shot performance was not already saturated, such as colon cancer. Thus, task-alignment raises the performance ceiling and im- proves editing stability even for easier tasks. For these tasks CLoPA-I.N generally outperforms CLoPA-C.N. This is likely because the existing low-level features are sufficient, and with relatively fewer samples (all these tasks are in medium dataset size at most) instance normalisation tuning is more robust. Tasks where the base model struggles:For more challenging tasks (NoF≳ 20%: brain tumour core, hippocampus, hepatic vessels), adaptation yields large performance gains across all metrics: substantially better initialisation, faster an- notation (nNoI), more robust editing stability (nAUC), and higher performance ceilings. These tasks share characteristics that expose limitations in nnInterac- tive’s zero-shot capabilities: ambiguous segmentation boundaries (brain tumour core), fine detail requirements at image sizes deviating from the model’s192 3 input patch (brain tumour core, hippocampus), and sparse branching structures with low volume fractions (hepatic vessels). For brain tumour core and hip- pocampus, the base model plateaus early but remains stable (nAUC slightly below final-iteration performance), indicating a performance ceiling when seg- Title Suppressed Due to Excessive Length7 Fig. 2.Comparison across methods for the trajectory of mean Dice Score after the editing terminates, as a function of data samples received. Left to right, trajectory on tasks: hepatic vessel, brain tumour core, hippocampus. Vertical lines indicate the number of samples (NoS) to reach nnU-Net performance (horizontal red line). 1 11111 %"$' 1 !#%% # %! "%$$$ T("*&(-*(!) %%%*(*!,("*&(- &("*&(- &("*&(- () &#)+$(&$'#)& %%*(&($% && && () &#)+$(&$'#)& %%*(&($% && && 1 111 1 %"$' 1 # &!&# !# ("*&(-*(!) %%%*(*!,("*&(- &("*&(- &("*&(- () &#)+$(&$'#)& %%*(&($% %%%*(*!,& && && () &#)+$(&$'#)& %%*(&($% %%%*(*!,& && && 11 111 %"$' 1 !""!"&$ ("*&(-*(!) %%%*(*!,("*&(- &("*&(- &("*&(- () &#)+$(&$'#)& %%*(&($% && () &#)+$(&$'#)& %%*(&($% && menting ambiguous targets with sparse point prompts. Instance normalisation adaptation partially mitigates this but improvements plateau after the first train- ing episode (Fig. 2), likely because the underlying feature representations are not adjusted for ambiguous targets or images smaller than the patch size. Fur- ther shallow-depth convolutional kernel tuning noticeably boosts performance for hippocampus, but not for brain tumour core. This is due to an underly- ing challenge with hippocampus being the small voxel-count, allowing low-level feature tuning to elevate the base model to nnU-Net performance (Fig. 2). For brain tumour core, ambiguous tissue boundaries likely require deep representa- tion tuning (hence mixed-results with instance normalisation only tuning). Hepatic vessels present a different challenge: sparse branching structures with low volume fractions, likely underrepresented in nnInteractive’s training data. The base model exhibits volatile peak-and-dip editing behaviour (nAUC much higher than final-iteration metrics). Post adaptation, initialisation performance immediately spikes and the editing performance does not exhibit this behaviour. After the first training episode (Fig. 2), instance normalisation tuning allows expected post-editing performance to exceed nnU-Net performance within ap- proximately 20% of the editing budget. However, with NoF at11.8%, perfor- mance still saturates—as illustrated by the plateauing after the initial episode in Fig. 2. This saturation likely reflects the inability of instance normalisation alone to learn task-specific feature representations. Contributing factors may include the CNN architecture’s limitations for targets with long-range depen- dencies (requiring adjusting convolution kernels). However, CLoPA-C.N does not provide improvements, indicating that tuning at shallow depth is limited or that the short interaction window during training (N=5edits), may be limiting exposure to gradient updates near segmentation completion. Trajectory analysis:Table 3 confirms the same trends across the full dataset- size trajectory, though improvements are less pronounced than in the final- episode snapshot because the trajectory AUC averages over all episodes, in- cluding before adaptation takes effect. As shown in Fig. 2, at least one adap- 8P. Esmaeili et al. Table 3.Trajectory AUC performance across all tasks. CLoPA-I.N: instance nor- malisation (I.N) only, CLoPA-C.N: I.N and shallow depth convolution kernel tuning. Metrics summarise AUCs over the trajectory of expected performance as a function of dataset size, except NoF (percentage of failed samples). Bold indicates first-ranked method per metric and task according to significance rankings. TaskAlgorithm DiceNSDNoI Init.Iter. 100nAUCInit.Iter. 100nAUCnNoINoF Brain Tumour nnInteractive0.5050.7530.7420.5830.9220.90426.820.4 CLoPA-I.N0.6230.8060.8020.7400.9600.95015.811.7 CLoPA-C.N0.6320.8020.7990.7510.9580.95016.813.1 Liver nnInteractive0.3730.9700.9630.3840.9920.98417.513.6 CLoPA-I.N0.7340.9720.9680.7370.9920.98815.810.2 CLoPA-C.N0.7170.9710.9670.7190.9930.9871610 Hippocampus nnInteractive0.5850.8050.7880.6310.8900.8709894.9 CLoPA-I.N0.8030.8800.8750.8740.9620.95655.448.6 CLoPA-C.N0.8030.8830.8780.8740.9630.95752.745.8 Prostate nnInteractive0.7740.9220.9210.8150.9850.9834.20 CLoPA-I.N0.8500.9320.9290.8990.9910.9883.70.694 CLoPA-C.N0.8510.9320.9300.9010.9910.9882.80 Lung Lesion nnInteractive0.6980.8490.8520.7430.9400.9361.80 CLoPA-I.N0.7480.8600.8570.7970.9470.9391.70 CLoPA-C.N0.7440.8570.8540.7950.9390.93120.243 Pancreas nnInteractive0.4540.8950.8810.5380.9830.9719.32.13 CLoPA-I.N0.5930.8990.8870.6830.9860.9778.12.01 CLoPA-C.N0.6000.8910.8800.6880.9820.97310.33.8 Hepatic Vessels nnInteractive0.1130.1650.2090.1220.2320.2898279.8 CLoPA-I.N0.4070.5550.5560.5420.7160.71539.431.7 CLoPA-C.N0.4050.5620.5630.5420.7090.71037.530.3 Colon Cancer nnInteractive0.4750.7280.7250.5200.7880.7902.40 CLoPA-I.N0.5760.7830.7790.6600.8670.8671.70 CLoPA-C.N0.5720.7840.7780.6580.8720.86920.256 tive method is always capable of reaching nnU-Net performance on unattain- able tasks for the base model (hepatic vessel and hippocampus). Moreover, we see that instance normalisation tuning rapidly elevates expected performance to near expert-level after the first training episode, but subsequently plateaus across all tasks. This is practically significant: it means that clinicians benefit from improved predictions early in the annotation campaign, reducing cumula- tive user effort over the full dataset. Since the initial episode yields the majority of the benefit, a two-phase strategy—triggering instance normalisation adapta- tion early, then transitioning to deeper feature-representation tuning as more data becomes available—would likely yield further gains. This is supported by the adaptation of convolution kernels producing higher peak performances, but being less stable in small data regimes. Such a curriculum could also incorporate dynamic prompting strategies, for instance extending the interaction window, to maximise the quality of the training signal provided to the model. In sum- mary, low-parameter adaptation elevates performance ceilings, stabilises editing behaviour, boosts annotation efficiency and enables specialist performance to be reached on all tasks using low-effort clicking, with only a fraction of the data. Title Suppressed Due to Excessive Length9 References 1. Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., Süsstrunk, S.: SLIC Superpixels Compared to State-of-the-Art Superpixel Methods. IEEE Transac- tions on Pattern Analysis and Machine Intelligence34(11), 2274–2282 (11 2012). https://doi.org/10.1109/TPAMI.2012.120 2. Antonelli, M., Reinke, A., Bakas, S., et al.: The Medical Segmentation Decathlon. Nature Communications13(1), 4128 (7 2022). https://doi.org/10.1038/s41467- 022-30695-9 3. Cheng, J., Ye, J., Deng, Z., Chen, J., Li, T., Wang, H., Su, Y., Huang, Z., Chen, J., Jiang, L., Sun, H., He, J., Zhang, S., Zhu, M., Qiao, Y.: SAM-Med2D (8 2023) 4. Du, Y., Bai, F., Huang, T., Zhao, B.: SegVol: Universal and Interactive Volumetric Medical Image Segmentation. In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (eds.) Advances in Neural Information Processing Systems. p. 110746–110783. Curran Associates Inc, Vancouver (2024) 5. Esmaeili, P., Borges, P., Fernandez, V., Gibson, E., Ourselin, S., Cardoso, M.J.: A Methodology for Clinically Driven Interactive Segmentation Evaluation. In: Guo, X., Jin, Y., Lamdouar, H., Men, Q., Ouyang, C., Sahu, M., Vedula, S.S. (eds.) International Workshop on Human-AI Collaboration. p. 13–22. Springer Nature Switzerland, Cham (2026). https://doi.org/10.1007/978-3-032-08970-0\_2 6. Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmenta- tion. Nature Methods18(2), 203–211 (2 2021). https://doi.org/10.1038/s41592- 020-01008-z 7. Isensee, F., Rokuss, M., Krämer, L., Dinkelacker, S., Ravindran, A., Stritzke, F., Hamm, B., Wald, T., Langenberg, M., Ulrich, C., Deissler, J., Floca, R., Maier- Hein, K.: nnInteractive: Redefining 3D Promptable Segmentation. arXiv preprint (3 2025) 8. Jadon, S.: A survey of loss functions for semantic segmentation. In: 2020 IEEE Conference on Computational Intelligence in Bioinformatics and Com- putational Biology (CIBCB). p. 1–7. IEEE, Via del Mar (10 2020). https://doi.org/10.1109/CIBCB48159.2020.9277638 9. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment Anything. arXiv preprint (4 2023) 10. de Kok, J.W.T.M., de la Hoz, M.A.A., de Jong, Y., et al.: A guide to sharing open healthcare data under the General Data Protection Regulation. Scientific Data 10(1), 404 (6 2023). https://doi.org/10.1038/s41597-023-02256-2 11. Lenchik, L., Heacock, L., Weaver, A.A., et al.: Automated Segmentation of Tissues Using CT and MRI: A Systematic Review. Academic Radiology26(12), 1695–1706 (12 2019). https://doi.org/10.1016/j.acra.2019.07.006 12. Ravi, N., Gabeur, V., Hu, Y.T., et al.: SAM 2: Segment Anything in Images and Videos. arXiv preprint (8 2024) 13. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomed- ical Image Segmentation. arXiv preprint (5 2015) 14. Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance Normalization: The Missing Ingredient for Fast Stylization. arXiv preprint (11 2017) 15. Wang, H., Guo, S., Ye, J., Deng, Z., Cheng, J., Li, T., Chen, J., Su, Y., Huang, Z., Shen, Y., Fu, B., Zhang, S., He, J., Qiao, Y.: SAM-Med3D. arXiv preprint (10 2023) 10P. Esmaeili et al. 16. Wong, H.E., Ortiz, J.J.G., Guttag, J., Dalca, A.V.: MultiverSeg: Scalable Inter- active Segmentation of Biomedical Imaging Datasets with In-Context Guidance. arXiv preprint (12 2024) 17. Wong, H.E., Rakic, M., Guttag, J., Dalca Adrian V.: ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image. In: European Con- ference on Computer Vision (ECCV). Milan (2024) 18. Xu, W., Liang, Z., Anthony, H., Ibrahim, Y., Cohen, F., Yang, G., Kamnitsas, K.: You Point, I Learn: Online Adaptation of Interactive Segmentation Models for Handling Distribution Shifts in Medical Imaging. In: The Fourteenth International Conference on Learning Representations (12 2025)