FuzzyMIL: Decoupling Pathological Phenotypes through Deep Fuzzy Clustering for Efficient Whole Slide Image Analysis

In Multiple Instance Learning (MIL) for Whole Slide Image (WSI) analysis, attention mechanisms are often employed to weigh the importance of different instances. However, global attention may lead to feature homogenization and overlook tissue differences. Adding local attention can capture these variations but is more parameter-intensive. To tackle the above challenges, in this work, we propose a deep fuzzy clustering framework (FuzzyMIL) based on a learnable Fuzzy C-means variant named FCM to analyze WSIs in a compact and efficient manner. By iteratively updating the fuzzy clustering centers, FCM decouples the morphological features of WSIs, resulting in more distinct and less correlated phenotypes. In this learning process, FCM compresses the feature representation space of WSI, guiding the features to gradually converge toward the representation prototypes. These prototypes, influenced by the soft assignment mechanism, take into account all updated features, enabling the model to retain both global information and local awareness. We evaluated our approach using three public datasets for diagnosis and sub-typing. Experimental results show that our approach achieves competitive performance while significantly reducing the downstream task framework parameters, striking a good balance between accuracy and model complexity. We release our code at https://github.com/Liuanana/FuzzyMIL.