Brain metastases segmentation compared across multiple AI models
Brain Metastases Segmentation for BraTS 2026 Task 1: A Multi-Architecture Comparison
Computer Vision and Pattern Recognition
Summary
Brain metastases are cancerous tumors in the brain that come from other parts of the body and can be very small and different in size. The authors compare several AI models that automatically find and outline these tumors in brain images, which helps doctors plan treatment and monitor tumor size. One transformer-based model called Primus performed best overall in judging the tumor areas, while a variant called ResEncL was better at finding individual tumor spots but worse at outlining them precisely. The authors also share some common issues they found in processing these tumor labels.
What this means in practice
- •For medical imaging teams: Deliver improved automatic delineation of brain metastases in patient MRI scans to assist radiologists in treatment planning and volume assessment.
- •For clinical trial coordinators: Use automated lesion detection to monitor changes in brain metastases volumes over time for evaluating treatment efficacy in oncology studies.
Authors
Mahdi Islam, Musarrat Tabassum
Abstract
Brain metastases are the most common intracranial malignancy, occurring in roughly 30% of patients with primary solid tumors and carrying a median survival near 5.9 months. Automated segmentation is critical for treatment planning and volumetric monitoring, but metastases are frequently small, numerous, and heterogeneous in size within a single patient. We compare a plain nnU-Net baseline, a Residual Encoder Large (ResEncL) variant, region-based training, and a Primus transformer model for BraTS-METS 2026 Task 1, using patient-grouped cross-validation to prevent leakage from the longitudinal UCSD subset. Primus (label-based) is our strongest individual model by aggregate DSC/NSD, achieving 0.710/0.761 (ET), 0.742/0.785 (TC), 0.683/0.689 (WT), and 0.531/0.436 (RC). ResEncL trails Primus on aggregate DSC/NSD but achieves substantially higher lesion-wise F1 (e.g. ET: 0.452 vs. 0.052); a probability-averaging ensemble of the two only partially preserves ResEncL's F1 advantage (ET lesion-wise F1: 0.064). We further report three postprocessing and label-reconstruction pitfalls we believe generalize beyond this challenge. Code is available at https://github.com/mahdiislam79/BraTS_METS_2026.