Selective Scale Cascade Attention Network for Breast Cancer Histopathology Image Classification

Convolutional Neural Networks (CNNs) approaches are widely applied to histopathological image analysis due to the breakthrough performance achieved. However, it remains challenging because complex backgrounds obscure the most discriminative region. In this paper, we propose selective scale cascade attention network (SSCA) to learning discriminative features for breast histopathological image classification. Based on the backbone, SSCA mainly consists of three modules: 1) cross scale attention module leverage deeper level features to generate an attention map that locates the discriminative part in high-resolution features and conduct multi-scale classification. 2) cascade attention modules gradually identify fine-grained cues and increase their weight through a cascade of attention. 3) selective scale fusion module dynamically adjusts the weight of each scale feature, i.e., the select scale depends on the characteristics of the cancer subtypes. Extensive experiments show that our proposed consistently outperforms the existing state-of-the-art methods on the public BreakHis dataset.