Post-Operative Glioma Segmentation via Loss Stabilization, Normalization and Subspace Attention

This paper studies why post-operative glioma segmentation models struggle across clinical protocols. Using the MU-GLIOMA-POST and UCSF-ALPTDG datasets, the authors report that standard Generalized Dice Loss is unstable under domain shift: Whole Lesion (WL) Dice falls from 0.88 on internal validation to 0.73 on the external UCSF test set. They combine brain-masked percentile normalization, voxel-level contrastive learning, and a Subspace-Aware Class Attention (SACA) module. On internal validation, SACA increases Enhancing Tumor sensitivity by 8%, or 9.1% relatively. Ensembles with nnU-Net reach WL Dice 0.94, while the SACA ensemble obtains HD95 of 2.92 mm on MU-GLIOMA-POST.