Mine local homogeneous representation by interaction information clustering with unsupervised learning in histopathology images

Background and objective: The success of data-driven deep learning for histopathology images often depends on high-quality training sets and fine-grained annotations. However, as tumors are heterogeneous and annotations are expensive, unsupervised learning approaches are desirable to obtain full automation. Methods: In this paper, an Interaction Information Clustering (IIC) method is proposed to extract locally homogeneous features in mutually exclusive clusters. Trained in an unsupervised paradigm, the framework learns invariant information from multiple spatially adjacent regions for improved classification. Additionally, an adaptive Conditional Random Field (CRF) model is incorporated to detect spatially adjacent image patches of high morphological homogeneity in an offset-constraint free manner. Results: Empirically, the proposed model achieves an observable improvement of 11.4% on the downstream patch-level classification accuracy, compared with state-of-the-art unsupervised learning approaches. Conclusion: Furthermore, evaluated with our clinically collected histopathology whole-slide images, the proposed model shows high consistency in tissue distribution compared with well-trained supervised learning, which is of important diagnostic significance in clinical practice.