Multi-view representation for pathological image classification via contrastive learning

Pathological images have become indispensable in clinical practice, but their complexity and blurred tissue structures pose challenges for accurate classification. To overcome this, it is crucial to combine the color view and edge view, as the features extracted from the edge view are sensitive for capturing subtle changes in repetitive patterns of cells and tissues. However, existing multi-view methods for pathological images often overlook the consistency among multiple views. Thus, we propose a multi-view method for pathological image classification using contrastive learning. By combining filtering techniques with the deep network, we enhance the interpretability of models and explicit modeling of features. We utilize Sobel filters to obtain the edge view, which complements the color view. To align the multi-view representations, we employ contrastive learning, encouraging the network to learn the intra-sample consistency and distinguish inter-sample differences. The fusion of features from the multi-view representations further enhances the representation. The experimental results on the BreakHist and CRC-100K datasets demonstrate the effectiveness of the proposed method, and we achieve better results on multiple evaluation metrics.