Search papers, labs, and topics across Lattice.
This study addresses the BraTS-GoAT challenge by introducing a novel segmentation method for brain tumor sub-regions that combines a semi-supervised learning approach with tumor-aware deformations. Utilizing the nnU-Net framework, the method generates pseudo-labels from unlabeled data and applies localized deformations to enhance segmentation accuracy while preserving anatomical integrity. The results demonstrate significant improvements over baseline models, achieving high Dice and NSD scores across various tumor regions, confirming the effectiveness of self-training and augmentation strategies in enhancing generalizability.
Self-training and tumor-aware deformations together boost brain tumor segmentation accuracy, outperforming traditional labeled-only approaches.
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/.
