Does Tumor-Mask Guidance Improve Glioma Grade Classification from T2-Weighted MRI Alone? An Out-of-Fold Evaluation with nnU-Net and Ground-Truth Localization

This is the authors' abstract. We don't add key points for this paper.

Multi-sequence MRI is the standard for glioma assessment, but complete or contrast-enhanced protocols are not always available (withheld gadolinium, low-field MRI, incomplete archives). We asked whether tumor localization improves grade classification from T2-weighted MRI alone. In this retrospective study of the BraTS2020 training cohort (369 patients; legacy labels 293 high-grade [HGG], 76 low-grade [LGG]), 74 patients (15 LGG) were held out at the patient level for internal testing. A 2D U-Net trained in five development folds supplied out-of-fold whole-tumor masks for training 3D CNN classifiers; nnU-Net served as a segmentation benchmark. Classifiers received T2 alone or T2 plus a mask, which at test time came from the 2D U-Net, nnU-Net, the ground truth (oracle) or random noise. Analyses used five seeds, DeLong tests with Holm correction and development-set cross-validation. Whole-tumor Dice was 0.807 (95% CI 0.773–0.840) for the 2D U-Net and 0.877 (0.857–0.895) for nnU-Net. The T2-only CNN reached an ROC-AUC of 0.814 (95% CI 0.702–0.913). Mask-guided variants reached 0.677–0.742, none above T2-only (Holm-adjusted p = 1.00 for all four confirmatory comparisons of informative masks). Cross-validation reproduced this ranking (0.782 vs. 0.707). With LGG as the positive class, PR-AUC was 0.571 (no-skill 0.203). Tumor-mask guidance did not improve T2-only grading at any level of mask quality, including ground truth. The results concern a legacy grade label in one public cohort and require external testing.