Dual selective gleason pattern-aware multiple instance learning with uncertainty regularization for grade group prediction in histopathology images.

Accurate prediction of Gleason Grade Group (GG) is of great importance for prostate cancer risk stratification and treatment planning. Although multiple instance learning (MIL) methods have advanced Gleason grading, most existing studies overlook the domain knowledge that GG is determined by the joint contribution of different Gleason Patterns, thereby limiting both accuracy and interpretability. In this study, we propose DSPA-U-MIL, an uncertainty-driven dual-selective Gleason Pattern-aware MIL model for patient-level GG prediction. Our method learns representative features by integrating learnable pattern aggregation tokens with expert concept-guided patch-level aggregation, and incorporates a teacher-student knowledge distillation framework to simulate cooperative prediction among different Gleason Patterns. In addition, we introduce an uncertainty constraint to mitigate the impact of noisy labels and enhance prediction robustness. Extensive experiments on five datasets, comprising 10,809 whole slide images (WSIs) and 1133 tissue microarray (TMA) images, demonstrate that DSPA-U-MIL consistently outperforms existing MIL approaches in Gleason GG prediction. Among these datasets, our method achieves up to a 6.7% improvement in quadratic weighted kappa (QWK) score over the strong baseline CLAM, with the largest gain observed on TCGA-PRAD. Furthermore, the gating weight distributions of the student model are well aligned with pathologists' Gleason Pattern annotations, reinforcing the interpretability of our approach. Our source code is available at https://github.com/AlexNmSED/DSPA-MIL.