Partially-Independent Framework for Breast Cancer Histopathological Image Classification

The automated classification of histopathology images relives pathologists workload and, hence utilizing the resources to focus more on the most suspicious cases. More recently, inspired by the success of deep learning methods in computer vision application, such frameworks have also been applied in various medical image analysis applications. However, existing approaches showed less interest in exploring multi-layer features for improving the classification. We propose the integration of multi-layer features from a ResNet model for breast cancer histopathology image classification. Specifically, this work focuses on making a framework which considers both independent nature of layers as well as some partial dependency among them. Knowing that, not all the layers learn discriminative features, consideration of layers which learn to negative features will deteriorate the accuracy. Hence, we select the optimal subset of the layers based on an information theoretic measure (ITS). Various experiments are performed on publicly available BreaKHis dataset, and demonstrate that the proposed multi-layer feature fusion yields better performance than the traditional way of using the highest layer features. This indicates that mid- and low-level features also carry useful discriminative information when explicitly considered. We also demonstrate improved performance, in most cases, over various state-of-the-art methods.