An X-ray image classification method with fine-grained features for explainable diagnosis of pneumoconiosis.

Medical image classification has become popular in computer-aided diagnosis (CAD) of pneumoconiosis. However, most current work focuses on improving the accuracy of classification results and has overlooked the corresponding medical explanations. With the expectation to achieve these two sub-goals simultaneously, we propose an explainable X-ray image classification method with fine-grained features to diagnose pneumoconiosis. The proposed method consists of three consecutive stages. First, we generate a highlighted discriminative region by gradient-weighted class activation mapping (Grad-CAM) for each sample. Thus, we can give a visual explanation for the basis of classification. Then, we utilize selective convolutional descriptor aggregation (SCDA) to extract fine-grained features from the obtained discriminative region. After dimension reduction of obtained fine-grained features, we finally make a classification with these features to discover which samples are diseased. Extensive experiments on actual pneumoconiosis X-ray image datasets have shown the validity and superiority of our method. [ABSTRACT FROM AUTHOR]