IGCNN-FC: Boosting interpretability and generalization of convolutional neural networks for few chest X-rays analysis

Abstract

Computer-aided diagnosis (CAD) with convolutional neural networks (CNNs) has been widely applied to assist doctors in medical image analysis. However, most of them often encounter two obstacles: (1) Data scarcity, because the advanced performance of CNNs heavily depends on a large amount of data, especially high-quality annotated ones. (2) Interpretability, CNNs cannot directly provide evidence related to the decision-making process to support their diagnosis results. To overcome these two obstacles, we propose an interpretable deep learning framework based on CNNs. Specifically, we introduce a multi-scale loss-based attention to leverage the mid- and high-level features to mine significant features for decision-making. Additionally, to better explore the semantic knowledge from training data, we utilize the mixup method to produce more annotated training images. Moreover, to boost model generalization capability, we employ the self-distillation to learn the knowledge generated from previous training epochs. Experiments on two benchmark Chest X-ray datasets demonstrate the effectiveness of the proposed framework with superior performance over recent SOTA methods, with boosting model interpretability.