Histopathologic Cancer Detection by Dense-Attention Network with Incorporation of Prior Knowledge

To identify the cancerous region in histology Whole-Slide Images (WSI), the common approach is to apply a patch-level classifier. Appending surrounding tissues could improve the accuracy of patch-wise classification and maintain consistency of WSI. However, the rule that surrounding tissues play a supporting role rather than a decisive one is difficult to be learned directly by a Convolutional Neural Networks (CNN). In this paper, we propose Dense-Attention Network (DAN) for cancerous patch classification, where the attention mechanism is further developed to incorporate prior knowledge about the surrounding tissue. Moreover, the effectiveness of Data Augmentation in Inference stage (DAI) is further validated. The proposed method is evaluated on the PatchCamelyon dataset, where images with tumor tissues in the center are labeled positive, and those in the outer regions do not influence the label. Compared with other competitive deep-learning methods the proposed method has achieved better performance in terms of AUC.