Multiple Instance Learning with Mixed Supervision in Gleason Grading

AbstractWith the development of computational pathology, deep learning methods for Gleason grading through whole slide images (WSIs) have excellent prospects. Since size WSIs is extremely large, image label usually contains only slide-level or limited pixel-level labels. The current mainstream approach adopts multi-instance to predict grades. However, some considering ignore labels containing rich local information. Furthermore, method additionally ignores inaccuracy To address these problems, we propose a mixed supervision Transformer based on multiple instance framework. model utilizes both and instance-level achieve more accurate at level. impact inaccurate further reduced by introducing an efficient random masking strategy in training process. We state-of-the-art performance SICAPv2 dataset, visual analysis shows prediction results source code available https://github.com/bianhao123/Mixed_supervision.KeywordsGleason gradingMixed supervisionMultiple