Abstract
Automated brain tumor segmentation from multi-parametric magnetic resonance imaging (mpMRI) is a fundamental component of computer-assisted diagnosis, treatment planning, and disease monitoring. The BraTS 2026 Generalizability Across Tumors (BraTS-GoAT) challenge extends conventional brain tumor segmentation by evaluating a single model across heterogeneous tumor populations, including adult glioma, sub-Saharan African glio-ma, meningioma, brain metastasis, and pediatric brain tumors. The resulting diversity in tumor morphology, imaging characteristics, and tumor sub-region composition poses signif-icant challenges for developing robust, generalizable segmentation algorithms. In this work, we investigate multiple deep learning strategies to improve segmentation generalization using Residual-Encoder nnU-Net, self-supervised SwinUNETR, and a neural-ODE network. We further evaluate the contributions of self-supervised pre-training, enhancing tumor (ET)-weighted optimization, lesion-aware post-processing, probability-level model fusion, and confidence-filtered semi-supervised learning using challenge-provided unlabeled data. All methods are developed exclusively using the BraTS-GoAT training data and evaluated on the official validation set. Experimental results show that self-supervised pre-training pro-vides region-dependent improvements in transformer performance, while the ET-weighted Residual-Encoder nnU-Net provides stronger supervised performance. Probability-level fusion yields configuration-dependent gains, whereas confidence-filtered semi-supervised learning achieves the best overall online validation performance. The final model achieves lesion-wise Dice scores of 0.8004, 0.8427, and 0.8970 for enhancing tumor, tumor core, and whole tumor, respectively, with corresponding HD95 values of 38.07, 16.10, and 10.85 mm. These results highlight the importance of systematically evaluating optimization, post-processing, model fusion, and unlabeled-data utilization for segmentation generalization across heterogeneous brain tumor populations. Code and implementation details are public-ly available at https://github.com/shiblyg/generalizable-brain-tumor-seg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_GoAT_014.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/BraTS_GoAT_014_Supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/profile?id=%7EMD_SHIBLY_SADIQUE1
BibTex
@InProceedings{SadM._ASystematic_MICCAISAT2026,
author = { Sadique, M. S. AND Hossen, MD Fayaz Bin AND Evans, Michael L. AND Farzana, Walia AND Rahman, Asfaqur AND Temtam, Ahmed AND Iftekharuddin, Khan M.},
title = { { A Systematic Evaluation of Learning Strategies for Generalizable Brain Tumor Segmentation Across Heterogeneous Tumor Populations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}
