Patch-Slide Discriminative Joint Learning for Weakly-Supervised Whole Slide Image Representation and Classification

In computational pathology, Multiple Instance Learning (MIL) is widely applied for classifying Giga-pixel whole slide images (WSIs) with only image-level labels. Due to the size and prominence of positive areas varying significantly across different WSIs, it is difficult for existing methods to learn task-specific features accurately. Additionally, subjective label noise usually affects deep learning frameworks, further hindering the mining of discriminative features. To address this problem, we propose an effective theory that optimizes patch and WSI feature extraction jointly, enhancing feature discriminability. Powered by this theory, we develop an angle-guided MIL framework called PSJA-MIL, effectively leveraging features at both levels. We also focus on eliminating noise between instances and emphasizing feature enhancement within WSIs. We evaluate our approach on Camelyon17 and TCGA-Liver datasets, comparing it against state-of-the-art methods. The experimental results show significant improvements in accuracy and generalizability, surpassing the latest methods by more than 2%. Code will be available at: https://github.com/sm8754/PSJAMIL .