Paper deep dive
CarveMix-RC: Addressing Rare-Class Imbalance Through Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation
Md Shibly Sadique, Md Fayaz Bin Hossen, Michael L. Evans, Walia Farzana, Asfaqur Rahman, Ahmed Temtam, Khan M. Iftekharuddin
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Accurate segmentation of post-treatment brain metastases is essential for treatment planning, longitudinal disease monitoring, and quantitative assessment of therapeutic response. The BraTS-MET 2026 Task 1 challenge introduces a clinically relevant segmentation problem involving four anatomically distinct tumor subregions: non-enhancing tumor core (NETC), surrounding non-enhancing FLAIR hyperintensity (SNFH), enhancing tumor (ET), and the resection cavity (RC). Among these, RC segmentation is particularly challenging because of its low prevalence, heterogeneous postoperative appearance, and lesion-wise evaluation protocol, leading conventional segmentation networks to prioritize dominant tumor classes during optimization. The proposed nnU-Net-based framework explicitly addresses RC segmentation through four complementary components: (i) RC-weighted Dice and Cross-Entropy optimization to alleviate class imbalance, (ii) anatomically consistent cavity augmentation to increase the diversity of postoperative cavity appearances, (iii) a residual encoder architecture for enhanced multi-scale feature learning, and (iv) lesion-aware morphological post-processing to suppress false-positive cavity predictions while preserving anatomically plausible structures. The framework is evaluated on the BraTS-MET 2026 Task 1 online validation benchmark. Among the evaluated configurations, the ensemble model (Residual Encoder nnU-Net + nnU-Net + RC-aware CarveMix) achieves the best performance, with lesion-wise Dice scores of 0.732, 0.752, 0.708, and 0.575 and corresponding NSD scores of 0.794, 0.798, 0.727, and 0.474 for ET, TC, WT, and RC, respectively. These experimental results show that integrating RC-aware optimization, anatomically consistent augmentation, and lesion-aware post-processing provides an effective strategy for improving rare resection cavity segmentation in post-treatment brain metastases.
Tags
Links
- Source: https://arxiv.org/abs/2609.35195v1
- Canonical: https://arxiv.org/abs/2609.35195v1
Trouble viewing inline? Open PDF directly β
Full Text
36,629 characters extracted from source content.
Expand or collapse full text
This work has been submitted to the 29 th International Conference on Medical Image Computing and Computer Assisted Intervention for possible publication. Copyright may be transferred without notice, after which this ver- sion may no longer be accessible. CarveMix-RC: Addressing Rare-Class Imbalance Through Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation M. S. Sadique 1 [0000β0002β6734β6802], MD Fayaz Bin Hossen 1 [0009-0008-5109-2466], Michael L. Evans 1 [0009-0002-6880-2950], W. Farzana 1 [0000β0003β1995β2426], Asfaqur Rahman 1 [0009-0007-9624-2766] A. Temtam 1 [0000β0002β1983β4422], and K. M. Iftekharuddin 1,2 [0000β0001β8316β4163] 1 Vision Lab, Department of Electrical and Computer Engineering, Old Dominion University, Norfolk, VA 23529, USA 2 Interdisciplinary Schools, Old Dominion University, Norfolk, VA 23529, USA https://sites.wp.odu.edu/VisionLab/ msadi002, mhoss006, mevan028, wfarz001, arahm003, atemt001, kiftekha@odu.edu Abstract. Accurate segmentation of post-treatment brain metastases is essential for treat- ment planning, longitudinal disease monitoring, and quantitative assessment of therapeutic re- sponse. The BraTS-MET 2026 Task 1 challenge introduces a clinically relevant segmentation problem involving four anatomically distinct tumor subregions: non-enhancing tumor core (NETC), surrounding non-enhancing FLAIR hyperintensity (SNFH), enhancing tumor (ET), and the resection cavity (RC). Among these, RC segmentation is particularly challenging because of its low prevalence, heterogeneous postoperative appearance, and lesion-wise evaluation protocol, leading conventional segmentation networks to prioritize dominant tumor classes during optimi- zation. The proposed nnU-Net-based framework explicitly addresses RC segmentation through four complementary components: (i) RC-weighted Dice and Cross-Entropy optimization to alleviate class imbalance, (i) anatomically consistent cavity augmentation to increase the diversity of postoperative cavity appearances, (i) a residual encoder architecture for enhanced multi-scale feature learning, and (iv) lesion-aware morphological post-processing to suppress false-positive cavity predictions while preserving anatomically plausible structures. The proposed framework was evaluated on the BraTS-MET 2026 Task 1 online validation benchmark using multi-parametric MRI. Among the evaluated configurations, the ensemble model (Residual Encoder nnU-Net + nnU-Net + RC-aware CarveMix) achieved the best perfor- mance, with lesion-wise Dice scores of 0.732, 0.752, 0.708, and 0.575 and corresponding NSD scores of 0.794, 0.798, 0.727, and 0.474 for ET, TC, WT, and RC, respectively. These experi- mental results show that integrating RC-aware optimization, anatomically consistent augmenta- tion, and lesion-aware post-processing provides an effective strategy for improving rare resection cavity segmentation in post-treatment brain metastases. Code and implementation details are publicly available at https://github.com/shiblyg/carvemix-rc. Keywords: Brain Metastases, Post-treatment MRI, Resection Cavity Segmentation, nnU-Net, Class Imbalance, Lesion-wise Segmentation. 2 Sadique et al. 1 Introduction Brain metastases (BMs) are the most common intracranial malignancies in adults, affecting approximately 20-40% of patients with systemic cancer and representing a major cause of neurological morbidity and mortality [1-3]. Advances in surgery, stere- otactic radiosurgery (SRS), immunotherapy, and targeted therapies have improved pa- tient survival, increasing the need for longitudinal MRI to monitor treatment response and disease progression. Accurate segmentation of post-treatment tumor subregions is therefore essential for treatment planning, response assessment, recurrence monitoring, and quantitative imaging biomarker development. Post-treatment brain metastasis segmentation remains considerably more challeng- ing than segmentation of untreated tumors because surgical intervention fundamentally alters normal anatomy. Postoperative cavities, blood products, edema, tissue defor- mation, and treatment-related imaging changes introduce substantial anatomical and appearance variability that can closely resemble recurrent disease or radiation-induced effects, making reliable delineation difficult even for experienced neuroradiologists [4]. These challenges motivate the BraTS-MET 2026 Challenge, which provides a stand- ardized benchmark for post-treatment brain metastasis segmentation using multi-para- metric MRI. BraTS-MET Task 1 requires simultaneous segmentation of four clinically relevant tissue classes: the non-enhancing tumor core (NETC), surrounding non-enhancing FLAIR hyperintensity (SNFH), enhancing tumor (ET), and resection cavity (RC). Among these, RC is particularly challenging because it occupies a relatively small im- age volume while exhibiting substantial variability in morphology and appearance due to differences in surgical technique, healing stage, hemorrhage, cerebrospinal fluid ac- cumulation, and surrounding tissue response. As a result, RC is severely underrepre- sented during optimization, often leading to reduced sensitivity and inconsistent lesion detection in conventional segmentation models [5-9]. Recent advances in medical image segmentation have been largely driven by nnU- Net, which automatically configures preprocessing, network architecture, and training strategies for robust performance across diverse biomedical imaging tasks. However, its standard optimization and sampling strategy does not explicitly address severe fore- ground class imbalance. Consequently, optimization is dominated by abundant tumor regions, encouraging the network to focus on ET, TC, and WT while underrepresenting the rare RC class. This limitation becomes particularly evident in lesion-wise evalua- tion, where even a small missing resection cavity can substantially reduce overall per- formance despite accurate segmentation of larger tumor regions [10,11]. Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation 3 Addressing postoperative brain metastasis segmentation requires more than increas- ing network capacity. The rarity of resection cavities introduces a fundamental learning challenge: models are inherently biased toward anatomically dominant tumor regions, resulting in poor representation of uncommon postoperative structures. Although data augmentation, class-balanced optimization, and post-processing have each been ex- plored to mitigate class imbalance, they are typically investigated in isolation and do not explicitly account for the anatomical characteristics of postoperative cavities. Con- sequently, accurate segmentation of resection cavities remains a persistent limitation in existing methods. To address these challenges, we propose a unified lesion-aware segmentation frame- work that improves the representation and delineation of rare postoperative anatomical structures while preserving robust performance across all tumor subregions. Rather than relying on a single strategy, the framework integrates complementary mechanisms for class-aware optimization, anatomically consistent cavity augmentation, enhanced feature representation, and lesion-aware false-positive suppression within the nnU-Net paradigm. Together, these components improve learning of the underrepresented resec- tion cavities while maintaining anatomical consistency throughout the segmentation process. The proposed framework was evaluated on the BraTS-MET 2026 Task 1 benchmark using multi-parametric MRI. Experimental results demonstrate consistent improve- ments in resection cavity segmentation while maintaining competitive performance for enhancing tumor, tumor core, and whole tumor segmentation, indicating that integrat- ing lesion-aware augmentation with class-aware optimization provides an effective so- lution for postoperative brain metastasis segmentation. Our main contributions are summarized as follows: 1. Identification of resection cavity segmentation as the main challenge in post- treatment brain metastasis analysis and present a unified framework to address this. 2. Introducing an RC-weighted strategy with anatomically consistent cavity aug- mentation to enhance learning of underrepresented cavity regions while main- taining performance on dominant tumor sub-regions. 3. Post-processing incorporating residual encoder-based feature learning along- side lesion-aware morphological post-processing to enhance segmentation ro- bustness and minimize false-positive cavity predictions. 4 Sadique et al. 2 Related Work 2.1 Brain Metastasis Segmentation Brain metastases (BMs) represent the most prevalent intracranial malignancy in adults and exhibit considerable variability in size, morphology, anatomical location, and treatment response. Accurate delineation of metastatic lesions and postoperative tissue compartments is essential for stereotactic radiosurgery planning, longitudinal disease monitoring, and quantitative response assessment. While considerable progress has been achieved in automated brain tumor segmentation using deep learning, most existing studies [5,7] have focused on untreated gliomas or preoperative metastatic le- sions, where tumor boundaries are comparatively well defined. In contrast, post-treat- ment brain metastases introduce substantially greater anatomical complexity due to sur- gical resection, postoperative cavity formation, hemorrhage, tissue deformation, edema, and heterogeneous enhancement patterns, making accurate segmentation con- siderably more challenging [5]. The Brain Tumor Segmentation (BraTS) challenges [5] have significantly advanced the development of automated segmentation methods by providing standardized datasets, evaluation protocols, and benchmark leaderboards. While previous BraTS editions pri- marily emphasized glioma segmentation, the recently introduced BraTS-MET bench- mark extends these efforts to pre and post-treatment brain metastases, requiring simul- taneous delineation of the non-enhancing tumor core (NETC), the surrounding non- enhancing FLAIR hyperintensity (SNFH), the enhancing tumor (ET), and the resection cavity (RC). This benchmark highlights the unique challenges associated with postop- erative anatomy and lesion-wise evaluation [5,7]. 2.2 Deep Learning for Medical Image Segmentation Convolutional neural networks have become the dominant paradigm for medical im- age segmentation, with encoder-decoder architectures such as U-Net establishing the foundation for numerous subsequent developments. Variants incorporating residual learning, attention mechanisms, dense connectivity, transformer-based feature extrac- tion, and multi-scale context aggregation have consistently improved segmentation per- formance across diverse imaging modalities [12-14]. Despite these architectural ad- vances, segmentation accuracy remains strongly influenced by optimization strategy, training data diversity, and class distribution, particularly in highly imbalanced medical imaging problems [15-18]. Among contemporary segmentation frameworks, nnU-Net has emerged as the de facto baseline because of its automatic configuration of preprocessing, architecture se- lection, patch size, normalization, data augmentation, and training schedules. Rather than proposing a new network architecture, nnU-Net demonstrates that carefully opti- mized training protocols often outperform increasingly complex architectural modifi- cations. Consequently, nnU-Net has become the reference framework for numerous in- ternational segmentation challenges, including multiple BraTS competitions [10,11]. Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation 5 However, the default nnU-Net optimization strategy assumes a relatively balanced set of foreground classes. Under severe class imbalance, optimization gradients become dominated by frequent anatomical structures, leading to reduced sensitivity for minority classes. This limitation is particularly evident for postoperative resection cavities, which occupy only a small fraction of the imaging volume yet contribute substantially to lesion-wise evaluation metrics. 2.3 Learning Under Severe Class Imbalance Class imbalance remains a major challenge in medical image segmentation because lesion classes often differ substantially in size and frequency. Existing approaches ad- dress this problem through weighted loss functions, adaptive sampling, hard example mining, and curriculum learning [12,14,19]. While these strategies improve optimiza- tion, they cannot compensate for the limited anatomical diversity of rare postoperative structures. Consequently, improving optimization alone is often insufficient for reliable segmentation of underrepresented classes such as resection cavities. 2.4 Data Augmentation for Rare Anatomical Structures Data augmentation plays a central role in improving the robustness and generaliza- tion of deep segmentation networks. Conventional augmentation techniques-including random rotations, scaling, elastic deformations, intensity perturbations, and spatial cropping-primarily increase appearance diversity without fundamentally altering ana- tomical composition. Recently, region-level augmentation strategies such as CutMix, MixUp, ClassMix, and CarveMix have demonstrated improved learning by exchanging semantically meaningful image regions between training samples [20-23]. While these approaches increase foreground diversity, they are not specifically designed for post- operative anatomy. Resection cavities exhibit irregular geometry, heterogeneous signal characteristics, and complex interactions with surrounding edema and residual tumor tissue. Therefore, augmentation strategies for post-operative segmentation should pre- serve anatomical plausibility while increasing representation of rare cavity appear- ances. Therefore, we leverage region-based cavity augmentation within a post-opera- tive brain metastasis segmentation framework to improve representation of the un- derrepresented RC class during training. 2.5 Lesion-Aware Segmentation and Postprocessing Medical image segmentation is now often assessed using lesion-wise metrics that focus on accurately detecting individual pathological structures, rather than just voxel- wise overlap. In these evaluation methods, false-positive predictions and missed small lesions can significantly reduce overall performance, even when global Dice scores ap- pear satisfactory. Consequently, morphology-aware postprocessing remains an im- portant component of modern segmentation pipelines. Connected-component analysis, size-based filtering, topology-preserving pos-pro- cessing, and morphological operations have been widely employed to suppress spurious 6 Sadique et al. predictions while preserving anatomically plausible structures. Several BraTS-winning approaches [10,11,13] have demonstrated that simple, yet carefully designed postpro- cessing strategies can produce measurable improvements, particularly for small or in- frequent lesion classes. post-processing. Motivated by these observations, the proposed framework integrates lesion-aware post-processing as its final phase to enhance the an- atomical consistency of resection cavity predictions while preserving segmentation ac- curacy in the remaining tumor regions. 2.6 CarveMix Augmentation Postoperative resection cavities are underrepresented in the training data and exhibit substantial variability in size, shape, and anatomical location. To increase the diversity of cavity appearances, we incorporated CarveMix [20], an anatomy-aware augmenta- tion strategy. As illustrated in Figure 1, CarveMix generates additional training samples by transplanting resection cavity regions between anatomically compatible subjects while preserving the corresponding segmentation labels. Unlike conventional intensity or geometric augmentations, CarveMix directly augments the morphology and spatial distribution of postoperative cavities, increasing the variability of rare RC examples presented during training. The augmented samples are used together with the original training data during nnU-Net optimization. 3 Methodology Fig. 1 illustrates the overall framework of the proposed method. Starting from multi- parametric MRI volumes, we first increase the representation of the underrepresented resection cavity (RC) class using an anatomically consistent cavity synthesis strategy adapted from CarveMix. The augmented dataset is subsequently used to train a Fig. 1. Overview of the proposed RC-aware segmentation framework Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation 7 Residual Encoder nnU-Net with RC-aware optimization. During inference, lesion- aware morphological post-processing is applied to suppress anatomically implausible false-positive cavity predictions while preserving tumor structures. 3.1 RC-aware Optimization The BraTS-MET dataset exhibits severe class imbalance, with the resection cavity accounting for only a small proportion of the foreground voxels. Consequently, stand- ard optimization tends to prioritize larger anatomical structures, such as the enhancing tumor (ET), the non-enhancing tumor core (NETC), and the surrounding non-enhanc- ing FLAIR hyperintensity (SNFH), resulting in inferior RC segmentation. Given the four MRI modalities ν=ν1,ν1ν,ν2ν€,ν2ν, The segmentation network predicts voxel-wise posterior probabilities ν=ν ν (ν). The network is optimized using a weighted hybrid objective β=νβ ν·ννν +(1βν)β ννΆνΈ , where the weighted cross-entropy is defined as β ννΆνΈ =ββν€ ν νΆ ν=1 ν ν logβ‘(ν ν ), with an increased class weight assigned to the RC category, ν€ ν = ν€ ν νΆ ,ν=ν νΆ, 1,otherwise. This weighting increases the contribution of rare cavity voxels to optimization while maintaining stable learning in the remaining tumor compartments. In addition, nnU- Net foreground patch oversampling is increased to expose the network to RC- containing regions more frequently during training. In all RC-aware experiments, Dice and cross-entropy were equally weighted (ν€ ν·ννν = ν€ νΆνΈ = 1), the cross-entropy class weights were [1, 1, 1, 1, 3] (ν€ ν νΆ = 3), and the foreground oversampling was increased from 0.33 to 0.66. 3.2 RC-aware Anatomically Consistent Cavity Synthesis Although weighted optimization enhances gradient allocation, the limited number of RC-positive subjects constrains the diversity of cavity appearances encountered during training. To alleviate this limitation, we leverage CarveMix for postoperative cavity synthesis. For the primary augmentation setting, 300 synthetic RC-positive cases were generated; an extended setting used 450 synthetic cases. Donor cases with fewer than 50 RC voxels were excluded, and the signed-distance threshold parameter was sampled as lambda ~β‘ν(β3,5), where negative and positive values contract and expand the carved RC region, respectively. For an RC-positive source image, the cavity mask is converted into a signed distance representation ν·(νΉ ν ), and a random threshold 8 Sadique et al. νβΌ 1 2 ν(ν ν ,0)+ 1 2 ν(0,ν ν’ ) is sampled to generate a variable cavity region ν(ν£)= ν,β‘ ( ν« ( νΉ ν )( ν ) β€ν ) β‘ β‘ν,β‘νννννννννβ‘ Unlike the original CarveMix formulation, CarveMix-RC does not rely on source- target spatial correspondence. The carved RC is transferred at its native dimensions using its minimum bounding box to a uniformly sampled location for which the com- plete box lies within the target volume; infeasible placements are rejected, with no resizing or interpolation. Only the RC mask and corresponding four-modality intensi- ties are transferred, and RC takes precedence over existing labels within the insertion mask; no atlas-based placement or tumor-overlap exclusion is imposed. Before inser- tion, source intensities are harmonized independently for each modality by matching the local mean and standard deviation of the target region. The resulting synthetic NIfTI cases subsequently undergo standard nnU-Net v2 preprocessing, with the complete pro- cedure defined in Algorithm 1. Algorithm 1: RC-aware CarveMix Augmentation Input: Training dataset ν=(νΏ ν ,ν ν ) ν=ν ν΅ , RC label ν νΉνͺ , desired number ν»of synthetic cases. Output: Synthetic dataset ν ν =(νΏ Μ ,ν Μ ). 1 for ν‘=1,2,...,νdo 2 Randomly select an RC-positive source (νβ,νβ) containing at least 50 RC voxels and an RC-negative target (νβ,νβ). 3 Define the binary RC mask ν ν (ν£)β‘=β‘ν[ν ν (ν£)β‘=β‘ν ν νΆ ]. 4 Compute the signed distance transform D(Rβ), negative inside the RC and positive outside. 5 Sample the cavity-size parameter νβ‘βΌβ‘ν°(β3,5). 6 Construct the variable carved region ν ν (ν£)β‘=β‘ν[ν·(ν ν )(ν£)β‘<β‘ν]. 7 Compute the minimum axis-aligned bounding box ν΅β enclosing νβ. 8 Crop the multimodal source patch ν ν = νβ[ν΅β] and binary mask ν ν = νβ[ν΅β]. 9 if Bβ exceeds the target dimensions along any axis then 10 reject the pair and resample; no resizing or interpolation is performed. 11 end if 12 Sample a valid insertion origin q such that the complete patch lies within νβ, and let ν΅β(ν) denote the corresponding target box. 13 For each modality m, harmonize source voxels inside ν ν to the local target statistics: νβ² ν,ν β‘=β‘ ν ν,ν β‘ββ‘ν ν ,ν ν ν ,ν β‘+β‘ν (ν ν‘,ν β‘+β‘ν)β‘+β‘ν ν‘,ν ,β‘νβ‘=β‘10β»βΆ. The source and target statistics are evaluated only over voxels selected by ν ν . 14 Initialize ν β‘ββ‘νβ and, within ν΅β(ν), update ν ν [ν΅ ν‘ (ν)]β‘=β‘νβ² ν,ν β‘ββ‘ν ν β‘+β‘ν ν‘,ν [ν΅ ν‘ (ν)]β‘ββ‘(1β‘ββ‘ν ν ). 15 Initialize α»Έβ‘ββ‘νβ and assign α»Έ[ν΅ ν‘ (ν)](ν£)β‘=β‘ν ν νΆ β‘νννβ‘ννβ‘ν£β‘ν ν’νββ‘ν‘βνν‘β‘ν ν (ν£)β‘=β‘1, while all remaining target labels are unchanged. 16 Add (ν ,α»Έ) to νβ. 17 end for 18 return νβ. Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation 9 3.3 Residual Encoder Segmentation and Lesion-aware Post-processing The augmented dataset is used to train a Residual Encoder nnU-Net while preserving the automated configuration and training pipeline of nnU-Net. During inference, slid- ing-window prediction with Gaussian weighting generates voxel-wise probability maps, which are subsequently converted to discrete segmentation labels. Lesion-aware post-processing applies 26-connected component analysis: RC components <50 voxels and ET components <15 voxels are removed. Binary hole filling is then applied to the RC segmentation to reduce internal discontinuities. Together, these operations suppress small isolated false-positive predictions while preserving the boundaries of larger pre- dicted lesions. 4 Experimental Setup 4.1 Dataset Experiments were conducted on the BraTS-MET 2026 Task 1 dataset, which com- prises 1,296 pre- and post-treatment brain metastasis cases across four co-registered MRI modalities (T1, T1c, T2w, and T2f). Expert annotations include four semantic labels: non-enhancing tumor core (NETC), surrounding non-enhancing FLAIR hyper- intensity (SNFH), enhancing tumor (ET), and resection cavity (RC). 4.2 Implementation Details The proposed framework was implemented using nnU-Net v2 with a Residual En- coder backbone in PyTorch. Images were preprocessed using the default nnU-Net pipeline, including foreground cropping, intensity normalization, and automatic target spacing estimation. Training employed weighted Dice and Cross-Entropy loss with in- creased emphasis on the RC class, RC-aware patch oversampling, and anatomically consistent cavity augmentation. Unless otherwise specified, all remaining hyperparam- eters followed the default nnU-Net configuration. Training and inference were per- formed on NVIDIA H100 GPUs. Inference was performed using nnU-Net's sliding-window prediction with Gaussian weighting and test-time augmentation. Final predictions were refined using lesion- aware post-processing (RC-Weighted nnU-Net), including connected-component fil- tering, cavity hole filling, and morphology-based removal of anatomically implausible RC predictions. Model ensembling (RC-Weighted ResEncL Predictions) was per- formed by averaging the softmax probabilities, followed by lesion-aware post-pro- cessing. 10 Sadique et al. Fig. 2. Qualitative Comparison of Five Model Configurations on the BraTS-MET 2026 Online Validation Set 5 Results 5.1 Experimental Evaluation The proposed framework was evaluated on the BraTS-MET 2026 Task 1 validation dataset using the official challenge evaluation server. Performance was evaluated using the official BraTS-MET 2026 evaluation framework. Segmentation accuracy was as- sessed using lesion-wise Dice Similarity Coefficient (DSC) and Normalized Surface Dice (NSD) for ET, RC, TC, and WT. Lesion-detection performance was further eval- uated using lesion-wise true positives (TP), false positives (FP), false negatives (FN), and F1-Scores for all, large, and small lesions [24,25]. Since the postoperative resection cavity is the primary focus of this work, particular emphasis is placed on RC segmen- tation performance while monitoring its effects on the remaining tumor subregions. 5.2 Quantitative Results Table 1 summarizes nine configurations evaluating RC-weighted optimization, CarveMix-RC augmentation, lesion-aware post-processing, and ensembling. On the ResEnc backbone, RC weighting increased RC Dice/NSD from 0.404/0.341 (C2) to 0.479/0.418 (C3), with RC Dice reaching 0.540 after post-processing (C4). CarveMix- RC further increased RC Dice from 0.540 (C4) to 0.575 (C8), although ET, TC, and WT Dice decreased modestly by 0.017, 0.021, and 0.008, respectively; RC NSD re- mained similar (0.449 vs. 0.450). The benefit of augmentation was therefore concen- trated primarily in RC overlap, with a modest trade-off in the more prevalent tumor regions. Post-processing also improved cavity prediction, increasing RC Dice from 0.530 (C6) to 0.575 (C8) in the ResEnc CarveMix configuration. An equal-size ensemble without CarveMix does not improve upon the corresponding CarveMix ensemble, Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation 11 providing additional control for the effect of model ensembling; detailed results are reported in the Supplementary Material. The final ensemble (C9) achieved Dice/NSD of 0.732/0.794 for ET, 0.752/0.798 for TC, 0.708/0.727 for WT, and 0.575/0.474 for RC. Table 2 summarizes Lesion-level performance, which showed a pronounced depend- ence on lesion size. Recall for large ET, TC, and WT lesions was 0.881, 0.883, and 0.858, respectively, compared with 0.194, 0.198, and 0.159 for small lesions. No small RC lesions were detected, whereas large-RC recall reached 0.833, identifying small- lesion sensitivity as the principal remaining limitation. Detailed TP, FP, FN, precision, recall, and F1 analyses are provided in the Supplementary Material. Paired patient-clustered bootstrap analysis with 10,000 resamples was used to quan- tify uncertainty. Relative to C2, C9 increased lesion-wise Dice by +0.055 for ET (95% CI, +0.035 to +0.078), +0.053 for TC (+0.033 to +0.076), +0.064 for WT (+0.039 to +0.092), and +0.171 for RC (+0.067 to +0.285). Component-level comparisons and region-specific uncertainty estimates are reported in the Supplementary Material. Table 1. Quantitative comparison of representative model configurations on the BraTS-MET 2026 validation dataset across tumor sub-regions: enhanced (ET), resection cavity (RC), tumor core (TC), whole tumor (WT) Configu- ration Model ET TC WT RC Dice NSD Dice NSD Dice NSD Dice NSD Baseline Std 0.665 0.729 0.688 0.738 0.646 0.67 0.368 0.247 Baseline Res 0.677 0.738 0.699 0.747 0.643 0.666 0.404 0.341 + RC weighting Res 0.662 0.725 0.686 0.736 0.647 0.667 0.479 0.418 + RC weighting + P ResEncL 0.728 0.789 0.748 0.796 0.693 0.711 0.54 0.449 + RC-w + Carve- Mix Std 0.657 0.717 0.676 0.723 0.645 0.668 0.434 0.355 + RC-w + Carve- Mix Res 0.715 0.779 0.736 0.786 0.688 0.709 0.53 0.427 + RC-w + Carve- Mix + P Std 0.677 0.743 0.705 0.758 0.67 0.694 0.481 0.4 + RC-w + Carve- Mix + P Res 0.711 0.773 0.727 0.774 0.685 0.707 0.575 0.45 Ensem- ble (Std + Res) Both 0.732 0.794 0.752 0.798 0.708 0.727 0.575 0.474 12 Sadique et al. Table 2. Lesion-level detection performance of the final ensemble stratified by lesion size. TP/FP/FN and F1 are from the official BraTS-METS evaluation; precision and recall are derived from the reported aggregate counts. Region Lesion size TP FP FN Preci- sion Recall F1 ET Small 1.31 0.57 5.41 0.695 0.194 0.347 Large 3.57 0.35 0.48 0.91 0.881 0.781 All 4.13 0.35 2.74 0.921 0.601 0.811 TC Small 1.32 0.53 5.37 0.715 0.198 0.354 Large 3.58 0.33 0.47 0.916 0.883 0.784 All 4.14 0.33 2.68 0.926 0.607 0.818 WT Small 0.96 0.79 5.07 0.549 0.159 0.251 Large 3.41 0.47 0.56 0.879 0.858 0.793 All 3.87 0.47 2.66 0.892 0.593 0.805 RC Small 0 0 1.17 -- 0 0 Large 0.11 0.14 0.02 0.444 0.833 0.1 All 0.12 0.14 0.06 0.457 0.677 0.476 5.3 Discussion The experimental results demonstrate that RC remains the most challenging region in post-treatment brain metastasis segmentation, reflecting its limited representation and substantial variability in postoperative appearance. The proposed combination of RC-weighted optimization, CarveMix-RC augmentation, and lesion-aware post-pro- cessing specifically addresses this imbalance during training and inference. CarveMix- RC primarily improved RC segmentation, although modest changes in ET, TC, and WT indicate a trade-off when increasing emphasis on the rare cavity class. The ensemble mitigated this trade-off and provided a more balanced performance across the four tar- get regions while preserving the improvement in RC. Lesion-level analysis further showed that performance was strongly dependent on lesion size. Larger lesions were detected reliably, whereas small lesions, particularly small RCs, remained the dominant source of error. This suggests that increasing the representation of RC examples alone is insufficient to resolve the small-lesion problem. In addition, CarveMix-RC uses geometrically feasible cavity placement without ex- plicit spatial or atlas-based constraints, which may limit the anatomical realism of some synthetic examples. Future work will investigate spatially informed cavity augmenta- tion and sampling strategies that better represent small postoperative lesions. Acknowledgments. The authors would like to acknowledge partial support for this work by the National Institute of Health grant #R01 EB020683. Lesion-Aware Synthetic Augmentation for Brain Metastasis Segmentation 13 References 1. Suh, J.H., Kotecha, R., Chao, S.T., Ahluwalia, M.S., Sahgal, A. and Chang, E.L., 2020. Current approaches to the management of brain metastases. Nature reviews Clinical oncol- ogy, 17(5), p.279-299. 2. Aizer, A.A., Lamba, N., Ahluwalia, M.S., Aldape, K., Boire, A., Brastianos, P.K., Brown, P.D., Camidge, D.R., Chiang, V.L., Davies, M.A. and Hu, L.S., 2022. Brain metastases: A Society for Neuro-Oncology (SNO) consensus review on current management and future directions. Neuro-oncology, 24(10), p.1613-1646. 3. Aizer, A.A., Shin, K.Y., Catalano, P.J., Ricca, I., Johnson, M., Benham, G., Roper, K., Spicer, B., Mann, E., Nosker, J. and Parsons, M.W., 2026. Treatment for brain metastases with stereotactic radiation vs hippocampal-avoidance whole brain radiation: a randomized clinical trial. Jama, 335(13), p.1127-1136. 4. Sadique, M.S., Farzana, W., Temtam, A., Lappinen, E., Vossough, A. and Iftekharuddin, K.M., 2024. Brain tumor recurrence vs. radiation necrosis classification and patient surviv- ability prediction. IEEE journal of biomedical and health informatics, 28(10), p.5685- 5695. 5. Moawad, A.W., Janas, A., Baid, U., Ramakrishnan, D., Saluja, R., Ashraf, N., Maleki, N., Jekel, L., Yordanov, N., Fehringer, P. and Gkampenis, A., 2024. The brain tumor segmen- tation-metastases (brats-mets) challenge 2023: Brain metastasis segmentation on pre-treat- ment mri. arxiv, p.arXiv-2306. DOI: https://doi.org/10.48550/arXiv.2306.00838 6. Karargyris, A., Umeton, R., Sheller, M.J., Aristizabal, A., George, J., Wuest, A., Pati, S., Kassem, H., Zenk, M., Baid, U. and Narayana Moorthy, P., 2023. Federated benchmarking of medical artificial intelligence with MedPerf. Nature machine intelligence, 5(7), p.799- 810. DOI: https://doi.org/10.1038/s42256-023-00652-2 7. Topff, L., Petrychenko, L., Jain, N., Lingier, S., Bertels, J., Astudillo, P., Prosec, M., Men- Γ©ndez FernΓ‘ndez-Miranda, P., Gevaert, O., Smits, M. and Derks, S., 2025. A data-centric approach to deep learning for brain metastasis analysis at MRI. Radiology, 315(3), p.e242416. 8. B. H. Menze, A. Jakab, S. Bauer, et al., βThe Multimodal Brain Tumor Image Segmentation Benchmark (BRATS),β IEEE Transactions on Medical Imaging, vol. 34, no. 10, p. 1993β 2024, 2015. 9. Lin, N.U., Lee, E.Q., Aoyama, H., Barani, I.J., Barboriak, D.P., Baumert, B.G., Bendszus, M., Brown, P.D., Camidge, D.R., Chang, S.M. and Dancey, J., 2015. Response assessment criteria for brain metastases: proposal from the RANO group. The lancet oncology, 16(6), p.e270-e278. 10. 11. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J. and Maier-Hein, K.H., 2021. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), p.203-211. 12. Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K. and Jaeger, P.F., 2024, October. nnu-net revisited: A call for rigorous validation in 3d medical image seg- mentation. In International Conference on Medical Image Computing and Computer-As- sisted Intervention (p. 488-498). Cham: Springer Nature Switzerland. 13. Ronneberger, O., Fischer, P. and Brox, T., 2015, October. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image 14 Sadique et al. computing and computer-assisted intervention (p. 234-241). Cham: Springer international publishing. 14. Myronenko, A., 2018, September. 3D MRI brain tumor segmentation using autoencoder regularization. In International MICCAI brainlesion workshop (p. 311-320). Cham: Springer International Publishing. 15. F. Milletari, N. Navab, and S.-A. Ahmadi, βV-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation,β in 3DV, 2016. 16. Evans, M.L., Hossen, M.D., Sadique, M.D., Farzana, W. and Iftekharuddin, K.M., 2026. CoMNeT: A MedNeXt-CorrDiff Framework for Volumetric Brain Tumor Segmenta- tion. arXiv preprint arXiv:2606.15305. 17. Sadique, M.S., Rahman, M.M., Farzana, W., Temtam, A. and Iftekharuddin, K.M., 2022, September. Brain tumor segmentation using neural ordinary differential equations with UNet-context encoding network. In International MICCAI Brainlesion Workshop (p. 205- 215). Cham: Springer Nature Switzerland. 18. Rahman, M.M., Sadique, M.S., Temtam, A.G., Farzana, W., Vidyaratne, L. and Iftekharud- din, K.M., 2021, September. Brain tumor segmentation using UNet-context encoding net- work. In International MICCAI brainlesion workshop (p. 463-472). Cham: Springer Inter- national Publishing. 19. Temtam, A., Sadique, M.S., Rahman, M.M., Farzana, W. and Iftekharuddin, K.M., 2023. Pediatric brain tumor segmentation using multiresolution fractal deep neural network. In International Challenge on Cross-Modality Domain Adaptation for Medical Image Segmen- tation (p. 332-340). Cham: Springer Nature Switzerland. 20. Lin, T.Y., Goyal, P., Girshick, R., He, K. and DollΓ‘r, P., 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (p. 2980-2988). 21. Zhang, X., Liu, C., Ou, N., Zeng, X., Zhuo, Z., Duan, Y., Xiong, X., Yu, Y., Liu, Z., Liu, Y. and Ye, C., 2023. CarveMix: a simple data augmentation method for brain lesion segmen- tation. NeuroImage, 271, p.120041. 22. Zhang, H., Cisse, M., Dauphin, Y.N. and Lopez-Paz, D., 2018, February. mixup: Beyond Empirical Risk Minimization. In International Conference on Learning Representations. 23. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J. and Yoo, Y., 2019. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision (p. 6023-6032). 24. Olsson, V., Tranheden, W., Pinto, J. and Svensson, L., 2021. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision (p. 1369-1378). 25. Reinke et al. Understanding metric-related pitfalls in image analysis validation. Nat Meth- ods. 2024 Feb;21(2):182-194. 26. Maier-Hein et al. Metrics reloaded: recommendations for image analysis validation. Nat Methods. 2024 Feb;21(2):195-212