Multimodal MRI Report Findings Supervised Brain Lesion Segmentation with Substructures

A research team has introduced MS-RSuper, a novel AI training framework that significantly improves brain tumor segmentation by learning from the incomplete, qualitative language found in real-world radiology reports. The method, detailed in a paper submitted to IEEE ISBI 2026, addresses a critical bottleneck in medical AI: the scarcity of perfectly labeled, voxel-by-voxel training data. Instead of requiring labor-intensive manual segmentation, MS-RSuper uses the existing textual findings from radiologists—which often describe only the 'largest lesion' or use uncertain terms like 'mild enhancement'—to supervise a deep learning model. This approach, termed report-supervised (RSuper) learning, makes building AI tools for complex, multi-parametric MRI scans far more practical.

The technical innovation lies in MS-RSuper's unified, uncertainty-aware formulation. It explicitly parses both global quantitative findings (e.g., lesion volume) and modality-wise qualitative cues (e.g., 'T1c enhancement' or 'FLAIR edema') from reports. The system aligns these cues with corresponding image substructures using specialized 'existence and absence' losses and enforces one-sided constraints for partial information. Crucially, it incorporates anatomical priors (distinguishing between extra- and intra-axial tumors) and down-weights missing or uncertain report data. Tested on a merged dataset of 1,238 brain tumor scans from the BraTS-MET and BraTS-MEN cohorts, MS-RSuper demonstrated superior performance over both a baseline trained with only sparse labels and a simpler, 'naive' report-supervised method, paving the way for more data-efficient diagnostic AI.