Source-linked AI summary
Generalizable Brain Tumor Segmentation with Self-Training and Tumor-Aware Deformations
Henrique Zan Grande, Jeovane Honorio Alves, Rayson Laroca, Andre Gustavo Hochuli
TL;DR
The paper tackles brain-tumor segmentation that must generalize across heterogeneous tumor types, acquisition sites, and patient populations. It extends nnU-Net with confidence-filtered self-training and tumor-aware deformable augmentation, and the combined approach improves validation performance over the labeled-only student. The authors report complementary contributions from pseudo-labeling and localized morphological augmentation.
Problem
BraTS-GoAT tests whether brain-tumor segmentation can generalize across heterogeneous tumor types, acquisition sites, and patient populations rather than relying on narrow training distributions.
Method
The method combines a Large Residual Encoder teacher, confidence-filtered pseudo-label self-training, and tumor-aware deformable augmentation within nnU-Net.
Results
82.4% average Dice was achieved by the combined approach, compared with 81.2% for the labeled-only student, while average HD95 decreased from 30.6 to 25.4 mm.
Takeaways & Limitations
Self-training and tumor-aware augmentation are complementary, with augmentation providing clearest additional gains for Whole Tumor and Tumor Core.
Takeaways & Limitations
Qualitative validation examples cannot be directly compared with ground truth because validation annotations are unavailable.
Abstract
from arXiv · showhide
This work presents an approach to the Generalizability Across Tumors (BraTS-GoAT) task of the BraTS 2026 Challenge, which focuses on robust segmentation of brain tumor sub-regions across a heterogeneous patient population. The proposed method employs the nnU-Net framework with a large residual encoder architecture, integrating a semi-supervised learning technique with pseudo-labels generated from the unlabeled training data and a tumor-aware deformable augmentation that locally deforms the lesion while preserving the surrounding anatomy. We evaluate the individual contributions of each component, as well as their combination, using varying proportions of the most confident pseudo-labeled cases. The submitted configuration for the generalization task achieves Dice and NSD scores of 0.881 and 0.473 for Whole Tumor, 0.817 and 0.490 for Tumor Core, and 0.775 and 0.533 for Enhancing Tumor on the BraTS-GoAT validation set, improving over the labeled-only baselines across all tumor regions and confirming that self-training and the proposed augmentation are complementary. Our source code is publicly available at https://github.com/Henrique-zan/brats-goat-2026/.
1 Introduction
The paper addresses reliable brain-tumor sub-region segmentation across heterogeneous tumors, sites, and patient populations. It combines self-training with tumor-aware deformation to improve generalization while preserving surrounding anatomy.
- Motivation: Brain tumors vary in appearance, size, and location, complicating automatic MRI segmentation of Whole Tumor, Tumor Core, and Enhancing Tumor.Accurate delineation supports diagnosis, treatment planning, and therapy assessment.
- Motivation: nnU-Net adapts preprocessing, network design, training, and inference to each dataset, but reliable performance across tumor types, protocols, sites, and populations remains challenging.
- BraTS-GoAT: BraTS-GoAT evaluates generalization across heterogeneous tumor distributions, including multiple tumor types, acquisition sites, and patient populations.The task penalizes overfitting to site-specific characteristics or narrow tumor morphologies.
- Approach: The proposed approach extends nnU-Net with a large-residual-encoder teacher, confidence-filtered self-training, and tumor-aware deformable augmentation.The augmentation increases tumor-shape variability while preserving surrounding anatomy.
- Evaluation: A systematic ablation evaluates self-training, augmentation, and their combination using a patient-level holdout and official BraTS-GoAT validation selection.
2 Dataset Specification
The BraTS-GoAT dataset is a multi-institutional MRI collection designed to test segmentation generalizability across tumor types, sites, and patient populations. It provides labeled and unlabeled training data alongside an official validation set.
- Imaging data: Each case contains four MRI input channels: T1n, T1c, T2w, and T2f.Volumes are skull-stripped, registered to an anatomical template, and resampled to 1mm^3 at 240 × 240 × 155 voxels.
- Dataset splits: The dataset includes labeled training data, an unlabeled training pool, and an official validation set evaluated through the challenge platform.
- Tumor populations: Labeled cases include adult gliomas, partially annotated meningiomas, and brain metastases, while the validation set adds BraTS-Africa and pediatric tumors.These additions introduce previously unseen tumor types and patient populations.
3 Method
The method uses a high-capacity teacher to generate confident pseudo-labels, then trains an nnU-Net student on labeled and selected unlabeled cases. Tumor-localized deformation expands plausible lesion morphology without globally altering anatomy.
- 3.2 Semi-Supervised Model (Student): A supervised teacher generates pseudo-labels for unlabeled cases, whose most confident predictions are merged with labeled data for student training.Cases with empty Whole Tumor predictions are discarded before confidence ranking.
- 3.1 Supervised Model (Teacher): The teacher uses nnU-Net v2 with a Large Residual Encoder, group normalization, axial decoder attention, a Hausdorff-distance loss term, and five-fold patient-level cross-validation.It predicts overlapping WT, TC, and ET regions using sigmoid outputs.
- 3.2 Semi-Supervised Model (Student): 75% of confidence-ranked pseudo-labeled cases are retained because this setting provided the best trade-off between retained-case quantity and label reliability.The compared retention fractions were 25%, 50%, 75%, and 100%.
- 3.2 Semi-Supervised Model (Student): The student uses an unmodified large-residual-encoder nnU-Net and trains on combined labeled and pseudo-labeled data with tumor-aware deformable augmentation.A patient-level 95%/5% holdout supports model selection and leakage prevention.
- 3 Method: The student benefits from the teacher’s knowledge while avoiding the teacher’s full model complexity.
- 3.3 Tumor-Aware Deformable Augmentation: Tumor-aware augmentation applies localized elastic deformation or isotropic scaling to sufficiently large lesions while smoothly preserving unchanged surrounding tissue.A Gaussian-smoothed weighting map derived from a dilated tumor mask restricts the transform to the lesion and nearby region.
4 Results
The evaluation isolates each pipeline component, reports official challenge-platform metrics, and compares configurations under distinct validation protocols. Internal validation results are restricted to within-block comparisons and do not estimate generalization to the official validation set.
- Experimental protocol: The protocol evaluates the teacher, labeled-only student, self-training, augmentation alone, and their combined configuration.Each component is isolated before evaluating the complete approach.
- Evaluation metrics: Dice and NSD are reported as percentages, HD95 in millimeters, and all Table 3 results as mean ± standard deviation over validation cases.The official metrics are returned by the challenge platform.
- Interpretation: Internal validation results should be interpreted only within their respective blocks because the blocks use different validation protocols.The left and right blocks use 95%/5% and 80%/20% holdouts, respectively.
- Interpretation: Scores from training-derived splits reflect in-distribution performance and should not be treated as estimates of generalization to the official challenge validation set.This limitation applies to the complementary internal analysis.
- Internal validation: Table 4 compares internal mean Dice scores across models and pseudo-label retention fractions using validation splits derived from the challenge-released training data.The table reports scores averaged across validation cases.
5 Discussion
The higher-capacity teacher contributes mainly through pseudo-label generation rather than direct deployment. Self-training improves overall student performance, while tumor-aware augmentation adds complementary boundary and region-wise gains.
- Teacher and student: The higher-capacity teacher differs only marginally from the labeled-only student despite substantially greater computational complexity.Its main value is producing reliable pseudo-labels for a simpler student.
- Self-training: 81.2% to 81.9% average Dice and 30.6 to 26.0 mm average HD95 follow from adding pseudo-labels through self-training.This result addresses RQ1 and indicates effective use of unlabeled data.
- Combined approach: 82.4% average Dice is achieved by the submitted model combining self-training with augmentation, with average NSD of 49.9% and HD95 of 25.4 mm.The combined model has the best average Dice among the student configurations.
- Interpretation: Self-training and tumor-aware augmentation are complementary: pseudo-labels expand the effective training distribution, while augmentation improves robustness to morphological variability.The augmentation is especially beneficial for larger tumor regions, while self-training provides the largest Enhancing Tumor improvement.
- Qualitative analysis: Qualitative validation examples support inspection of predicted regions but cannot be directly compared with ground truth because validation annotations are unavailable.Figure 4 therefore provides qualitative rather than quantitative validation evidence.
6 Conclusions
The submitted approach combines confidence-filtered self-training with tumor-aware deformable augmentation for BraTS-GoAT segmentation. On the official validation set, the combined approach improves overlap and boundary metrics over the labeled-only student, while further analyses and extensions remain future work.
- The approach combines confidence-filtered self-training with tumor-aware deformable augmentation.Self-training uses high-confidence pseudo-labels, while augmentation increases plausible tumor-shape and size diversity without globally deforming surrounding anatomy.
- 82.4% average Dice and 25.4 mm average HD95 improve over 81.2% and 30.6 mm for the labeled-only student.These results are reported on the official validation set and cover both overlap and boundary metrics.
- The two components are complementary, with augmentation providing its clearest additional gains for WT and TC.
- Iterative self-training, teacher ensembles, improved pseudo-label ranking, post-processing, and broader domain analyses are identified as future work.