Source-linked AI summary
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
TL;DR
Rare, variable resection cavities make post-treatment brain metastasis segmentation difficult, particularly under lesion-wise evaluation. The paper combines RC-aware optimization, cavity augmentation, a residual encoder, and lesion-aware post-processing; its final ensemble reports the listed lesion-wise Dice and NSD scores on BraTS-MET 2026 Task 1. Small lesions, especially small RCs, remain a major error source.
Problem
Rare, variable resection cavities are underrepresented during optimization, while lesion-wise evaluation makes cavity segmentation a challenging part of post-treatment brain metastasis analysis.
Source: §1 · pp. 2–3Method
The framework combines RC-weighted optimization, anatomically consistent CarveMix-RC augmentation, residual encoder nnU-Net, and lesion-aware post-processing.
Source: §1 · p. 3, §2.6 · p. 6Results
The final ensemble 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.
Source: §5.2 · p. 11Takeaways & Limitations
The ensemble provided more balanced performance across four target regions while preserving the RC improvement, although CarveMix-RC primarily improved RC and modest changes in other regions indicated a trade-off.
Source: §5.3 · p. 12No small RC lesions were detected, and geometrically feasible cavity placement without explicit spatial or atlas-based constraints may limit the anatomical realism of some synthetic examples.
Source: §5.2 · p. 11, §5.3 · p. 12
Abstract
from arXiv · showhide
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.
1 Introduction
Post-treatment brain metastasis segmentation is difficult because surgery and treatment alter anatomy, while resection cavities are rare and variable. The paper proposes a unified framework combining RC-aware optimization, cavity augmentation, residual encoder features, and lesion-aware post-processing; it reports improved RC segmentation alongside competitive performance across tumor regions.
1 Introduction
- Post-treatment anatomy is altered by surgery, cavities, blood products, edema, tissue deformation, and treatment-related imaging changes that complicate reliable delineation.These appearances can resemble recurrent disease or radiation-induced effects.Source: §1 · p. 2
- RC occupies little image volume and varies with surgical technique, healing, hemorrhage, fluid accumulation, and surrounding tissue response, making it underrepresented during optimization.Conventional models may consequently show reduced RC sensitivity and inconsistent lesion detection.Source: §1 · p. 2
- Standard nnU-Net optimization favors abundant ET, TC, and WT regions over rare RC, a problem sharpened by lesion-wise evaluation of missed cavities.Even a small missed cavity can reduce overall performance despite accurate segmentation of larger tumor regions.Source: §1 · p. 2
- The framework integrates class-aware optimization, anatomically consistent cavity augmentation, enhanced feature representation, and lesion-aware false-positive suppression within nnU-Net.Its components are intended to improve learning of underrepresented cavities while maintaining anatomical consistency.Source: §1 · p. 3
- On BraTS-MET 2026 Task 1, the proposed framework improved RC segmentation while maintaining competitive ET, TC, and WT performance.The paper identifies lesion-aware augmentation with class-aware optimization as an effective strategy for postoperative brain metastasis segmentation.Source: §1 · p. 3
2 Related Work
Related work describes the added anatomical complexity of post-treatment metastases, the limits of standard nnU-Net under severe class imbalance, and the need to augment rare cavity appearances and account for lesion-wise evaluation. The paper situates its approach among benchmark, optimization, augmentation, and post-processing strategies.
2.1 Brain Metastasis Segmentation
- Post-treatment brain metastases add cavities, hemorrhage, deformation, edema, and heterogeneous enhancement to segmentation, unlike comparatively well-defined untreated or preoperative lesions.BraTS-MET extends benchmark evaluation to pre- and post-treatment cases and four tissue labels: NETC, SNFH, ET, and RC.Source: §2.1 · p. 4
2.2 Deep Learning for Medical Image Segmentation
- nnU-Net is a reference framework with automatic configuration, but its default optimization can favor frequent structures over rare postoperative cavities.Cavities occupy little imaging volume yet contribute substantially to lesion-wise evaluation metrics.Source: §2.2 · pp. 4–5
2.3 Learning Under Severe Class Imbalance
- Weighted losses, adaptive sampling, hard-example mining, and curriculum learning address imbalance but cannot replace limited anatomical diversity in rare postoperative structures.Optimization alone may therefore be insufficient for reliable RC segmentation.Source: §2.3 · p. 5
2.4 Data Augmentation for Rare Anatomical Structures
- Conventional augmentation increases appearance diversity without changing anatomical composition, while region-level methods exchange image regions but are not designed specifically for postoperative anatomy.The paper motivates cavity augmentation that preserves anatomical plausibility while increasing rare RC representation.Source: §2.4 · p. 5
2.5 Lesion-Aware Segmentation and Postprocessing
- Lesion-wise metrics penalize false positives and missed small lesions even when global Dice appears satisfactory, motivating morphology-aware post-processing.Connected-component and morphological operations are described as tools for suppressing spurious predictions while preserving plausible structures.Source: §2.5 · pp. 5–6
2.6 CarveMix Augmentation
- CarveMix transplants resection-cavity regions between anatomically compatible subjects while preserving corresponding labels, augmenting cavity morphology and spatial distribution.The augmented samples are used alongside original training data during nnU-Net optimization.Source: §2.6 · p. 6
3 Methodology
The methodology combines RC-aware optimization with synthetic cavity augmentation, then trains a residual encoder nnU-Net and applies lesion-aware post-processing. CarveMix-RC transfers cavity regions into target cases using feasible native-size placement and modality-wise intensity harmonization.
3 Methodology
- The pipeline augments multi-parametric MRI with cavity synthesis, trains a Residual Encoder nnU-Net using RC-aware optimization, and applies lesion-aware morphological post-processing at inference.The post-processing targets anatomically implausible false-positive cavity predictions while preserving tumor structures.Source: §3 · pp. 6–7
3.1 RC-aware Optimization
- RC-weighted optimization addresses the cavity's small share of foreground voxels, which otherwise leads standard optimization to prioritize larger tumor structures.RC-aware experiments use cross-entropy class weights, equally weighted Dice and cross-entropy losses, and increase foreground oversampling from 0.33 to 0.66.Source: §3.1 · p. 7
3.2 RC-aware Anatomically Consistent Cavity Synthesis
- CarveMix-RC synthesizes cavity-positive cases because limited RC-positive subjects restrict the range of cavity appearances in training.The primary setting generated 300 synthetic cases and the extended setting 450; donors with fewer than 50 RC voxels were excluded, and λ was sampled from U(−3, 5).Source: §3.2 · p. 7
- CarveMix-RC transfers the carved RC at native dimensions into a uniformly sampled target location only when its full bounding box fits, without resizing or interpolation.Source intensities are harmonized per modality to local target mean and standard deviation; RC overrides existing labels within the insertion mask, with no atlas placement or tumor-overlap exclusion.Source: §3.2 · p. 8
3.3 Residual Encoder Segmentation and Lesion-aware Post-processing
- The residual encoder model uses sliding-window inference and lesion-aware component filtering, removing RC components below 50 voxels and ET components below 15 voxels.RC hole filling is applied to reduce internal discontinuities while suppressing small isolated false positives and preserving larger lesion boundaries.Source: §3.3 · p. 9
4 Experimental Setup
Experiments use the BraTS-MET 2026 Task 1 dataset and nnU-Net v2 with a residual encoder backbone. Inference combines sliding-window prediction, test-time augmentation, lesion-aware post-processing, and probability averaging for ensembles.
4.1 Dataset
- The BraTS-MET 2026 Task 1 dataset comprises 1,296 pre- and post-treatment cases across four co-registered MRI modalities, with expert annotations for NETC, SNFH, ET, and RC.Source: §4.1 · p. 9
4.2 Implementation Details
- The implementation uses nnU-Net v2 with a PyTorch Residual Encoder, default preprocessing, weighted Dice and cross-entropy, RC-aware oversampling, and cavity augmentation.Training and inference were performed on NVIDIA H100 GPUs.Source: §4.2 · p. 9
- Inference uses sliding-window prediction with Gaussian weighting and test-time augmentation, followed by lesion-aware filtering and cavity hole filling.Ensembling averages softmax probabilities before lesion-aware post-processing.Source: §4.2 · p. 9
5 Results
On BraTS-MET 2026 Task 1 validation, the final ensemble reports Dice/NSD results across ET, TC, WT, and RC, while lesion-level results show a marked small-lesion sensitivity gap. The discussion reports a more balanced four-region profile with the ensemble, but notes that small lesions and the anatomical realism of synthetic cavity placement remain concerns.
5.1 Experimental Evaluation
- Evaluation used the official BraTS-MET 2026 Task 1 validation server and lesion-wise Dice and NSD for ET, RC, TC, and WT.Lesion-level TP, FP, FN, and F1 were also evaluated for all, large, and small lesions.Source: §5.1 · p. 10
5.2 Quantitative Results
- RC weighting raised ResEnc RC Dice/NSD from 0.404/0.341 (C2) to 0.479/0.418 (C3), while CarveMix-RC raised RC Dice from 0.540 (C4) to 0.575 (C8).CarveMix-RC was accompanied by ET, TC, and WT Dice decreases of 0.017, 0.021, and 0.008; post-processing increased RC Dice from 0.530 (C6) to 0.575 (C8).Source: §5.2 · p. 10
- C9 increased lesion-wise Dice over C2 by +0.171 for RC (95% CI, +0.067 to +0.285), the largest reported regional gain.The corresponding increases were +0.055 for ET, +0.053 for TC, and +0.064 for WT, using paired patient-clustered bootstrap analysis with 10,000 resamples.Source: §5.2 · p. 11
- The final ensemble C9 achieved Dice/NSD scores 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 on the BraTS-MET 2026 validation dataset.Source: §5.2 · p. 11, Table 1 · p. 11
- No small RC lesions were detected, whereas large-RC recall reached 0.833.For small versus large lesions, recall was 0.194 versus 0.881 for ET, 0.198 versus 0.883 for TC, and 0.159 versus 0.858 for WT.Source: §5.2 · p. 11, Table 2 · p. 12
5.3 Discussion
- The ensemble mitigated the trade-off between RC gains and changes in other regions, while small lesions remained a dominant error source and cavity placement lacked explicit spatial or atlas-based constraints.The authors identify spatially informed augmentation and sampling that better represent small postoperative lesions as future work.Source: §5.3 · p. 12