Source-linked AI summary
Native-Space 3D CarveMix for Multi-Site T1w Stroke Segmentation
Dexter Wen Jie Teo, Kumaradevan Punithakumar
TL;DR
Native T1w stroke-lesion segmentation remains difficult across scanners because lesions are subtle and acute cases are scarce. The paper combines MedNeXt-L with on-the-fly 3D CarveMix that pastes real lesions into healthy tissue within subject-isolated folds. On 1,453 scans from 55 centers, the method reaches mean 5-fold cross-validation Dice 0.648 at 500 epochs, improving 0.630 by +0.018.
Problem
Multi-site native T1w stroke segmentation is limited by subtle lesion contrast, scanner and protocol variation, and severe underrepresentation of acute cases.
Method
The method combines MedNeXt-L with on-the-fly 3D CarveMix that grafts real lesion patches into healthy tissue within subject-isolated training folds.
Results
0.648 mean 5-fold cross-validation Dice is achieved at 500 epochs, a +0.018 improvement over the matched MedNeXt-L baseline of 0.630.
Takeaways & Limitations
Dynamic lesion placement expands training diversity while preserving fold isolation and improves native-space segmentation across acute, subacute, and chronic stages.
Takeaways & Limitations
Per-stage metrics are inherently noisy because only 92 acute subjects are available, warranting further validation on prospectively collected acute cohorts.
Abstract
from arXiv · showhide
Segmenting ischemic stroke lesions on T1-weighted (T1w) MRI acquired across different scanners and protocols without intensity standardization is difficult because lesions are subtle and share intensity characteristics with cerebrospinal fluid. Standard deep learning architectures trained across multiple centers plateau around Dice 0.66, with acute lesions ($\le 7$ days post-stroke) performing substantially worse due to severe sample scarcity. We combine a MedNeXt-L ($k=5$) backbone with on-the-fly 3D CarveMix augmentation that pastes real lesion patches into healthy brain regions during training. By generating synthetic lesion placements dynamically within each fold with subject-level split isolation, the model sees more diverse lesion patterns without requiring pre-generated copies on disk. We evaluate on 1,453 native T1w scans from 55 clinical centers in the ISLES 2026 challenge. Our method achieves a mean 5-fold cross-validation Dice of 0.648 at 500 epochs, a +0.018 improvement over the MedNeXt-L backbone at a matched training budget (0.630)
1 Introduction
Native T1w MRI is valuable for multi-site stroke research but difficult to segment because lesions are subtle, scanner-dependent, and especially scarce in acute cases. The paper addresses this bottleneck with native-space MedNeXt-L and dynamically generated 3D CarveMix augmentation.
- Motivation: T1w MRI is often the only available modality in large retrospective and observational stroke datasets, motivating direct lesion segmentation.DWI and FLAIR provide higher acute-lesion contrast but are frequently unavailable in these cohorts.
- Technical hurdles: T1w lesions are subtle hypointensities that blend with cerebrospinal fluid, while scanners and protocols introduce substantial inter-site intensity variation.The challenge combines intensity ambiguity with multi-center domain shift.
- Central bottleneck: Acute lesion segmentation remains the main performance bottleneck because acute cases are substantially underrepresented in multi-center training data.The paper identifies sample scarcity, rather than preprocessing alone, as the central limitation.
- Contributions: The study evaluates native raw T1w scans with MedNeXt-L and positions the pipeline as a reproducible benchmark for multi-site stroke lesion segmentation.The contribution includes testing conventional remedies and establishing a native-space evaluation setting.
- Proposed method: 3D CarveMix extracts real lesion patches from donor cases and grafts them into healthy recipient tissue during training with edge blending.Synthetic placement occurs within the active training fold.
- Proposed method: Dynamic lesion placement exposes the network to continuously varying spatial and morphological patterns without cross-fold leakage or extra disk copies.Subject-level isolation is maintained while augmentation expands sample diversity across epochs.
2 Methods
The method processes native-space T1w volumes with MedNeXt-L and implements CarveMix inside nnU-Net’s training loop. Lesion patches are sampled and pasted dynamically while subject-level fold isolation prevents anatomy from crossing validation boundaries.
- Preprocessing: Volumes are processed in native space without registration, using per-volume Z-score normalization over foreground brain voxels.N4 correction was tested but omitted because it made no meaningful difference.
- Architecture: MedNeXt-L uses depthwise separable convolutions and inverted bottleneck blocks, with kernel size k = 5 selected from cross-validation performance.The wider receptive field is intended to help distinguish subtle lesions from adjacent cerebrospinal fluid.
- Augmentation: CarveMix takes a lesion from one subject and pastes it into another to create a synthetic example with a known label.The operation is implemented on the fly within the nnU-Net training loop.
- Augmentation: A lesion bank is built from training-fold patches, and batches apply CarveMix with probability p = 0.4 using inverse-chronicity-frequency donor weighting.Donor lesions are placed in healthy foreground tissue, with acute lesions sampled more often to address rarity.
- Data isolation: Per-patch pasting creates different configurations across epochs without extra disk copies, while subject-level splitting keeps all variants within one fold.This prevents the same brain anatomy from appearing across training and validation.
3 Experimental Setup
Experiments use the ISLES 2026 native T1w benchmark with center-grouped five-fold validation. The cohort is strongly skewed toward subacute and chronic cases, leaving acute lesions underrepresented.
- Cohort composition: Acute cases comprise 92 subjects (6.3%), compared with 712 subacute and 649 chronic subjects.Acute cases are defined as ≤7 days post-onset, whereas chronic cases exceed 180 days.
- Validation: Five-fold validation groups scans by center ID so each hospital belongs entirely to training or validation, testing generalization to unseen acquisition protocols.Folds are also stratified by chronicity.
- Training: Training uses nnU-Net v2 with MedNeXt-L for 500 epochs, Dice plus cross-entropy loss, Nesterov SGD, and 128×128×128 patches.The batch size is 2 and the initial learning rate is 0.01 with polynomial decay.
- Sampling: Each epoch has 250 iterations, and CarveMix further increases diversity by synthesizing new lesion placements during those iterations.Subjects are sampled roughly once every six epochs on average.
4 Results and Ablation
CarveMix improves matched-budget mean Dice over the MedNeXt-L baseline, with similar gains across chronicity stages. Ablations show that N4 correction provides little benefit, while explicit Z-axis coordinates and FiLM conditioning reduce performance.
- Baseline performance: 0.407 acute DSC, 0.666 subacute DSC, and 0.696 chronic DSC are achieved by the baseline at 1,000 epochs.The stage disparity indicates substantially weaker acute-case detection.
- CarveMix result: +0.018 mean Dice is obtained with CarveMix at 500 epochs, reaching 0.648 versus 0.630 for the matched MedNeXt-L baseline.The comparison is explicitly made at the same training budget.
- Chronicity-stratified results: Acute, subacute, and chronic cases each improve by approximately +0.018 with CarveMix at the matched 500-epoch budget.The reported gains are acute +0.019, subacute +0.018, and chronic +0.018.
- Ablation: N4 correction adds only ∆Dice = +0.002, while Z-axis coordinates lower MedNeXt-S Dice from 0.636 to 0.616.FiLM metadata conditioning also decreases ResEnc-XL performance from 0.656 to 0.649.
- Overall result: The full MedNeXt-L plus dynamic CarveMix pipeline reaches 5-fold cross-validation Dice of 0.648, a +0.018 gain over the matched 0.630 backbone.The reported comparison is on native T1w scans.
5 Conclusion
Dynamic 3D CarveMix addresses acute-lesion scarcity in native T1w stroke segmentation, improving mean Dice while showing similar gains across chronicity stages. However, acute-stage metrics remain noisy and warrant prospective validation.
- 0.648 mean 5-fold cross-validation Dice improved by +0.018 over the matched MedNeXt-L baseline of 0.630.The comparison uses a matched 500-epoch training budget.
- Dynamic 3D CarveMix generates synthetic lesion placements on the fly to address acute lesion scarcity.Acute cases comprise 92 of 1,453 subjects, with minimal tissue contrast in the early post-stroke period.
- Approximately +0.018 improvement was consistent across acute, subacute, and chronic cases.The conclusion characterizes this stage-wise consistency as a robust improvement across chronicity stages.
- Per-stage metrics are inherently noisy with only 92 acute subjects, motivating further validation on prospectively collected acute cohorts.