Imbalanced Histopathology Image Classification Using Deep Feature Graph Attention Network

Recent deep learning studies have shown great progress in metastasis detection over histopathology whole slide images (WSIs). As WSIs are extremely large, most of the existing studies adopt patch level analysis, leveraging spatial context to enhance patch-wise classification. However, class imbalance in patch distribution may result in an exceedingly large number of false negatives, thereby worsening the WSI level classification performance. In this paper, we propose a novel framework for classification in class imbalanced datasets, which adopts a graph attention network to capture feature dependent interactions and a minority preferred inference mechanism for patch-level classification. Our experiments on CAMELYON16 show that the proposed method substantially improves the detection of the minority class (tumour) under a highly imbalanced class distribution.