Diagnose Chest Pathology in X-ray Images by Learning Multi-Attention Convolutional Neural Network

Automatically diagnosing diseases from X-ray images has been drawing great attention from the community of computer vision and medical image computing in recent years. Existing works predominantly formulate this task as multi-label image classification by deep neural networks, which have demonstrated the effectiveness for classifying general image categories (e.g., in ImageNet). Although promising results are reported, there are still grand challenges to model large intra-class variances and subtle inter-class distinctions existed in those fine-grained categories (e.g., two similar chest pathologies). In this paper, we propose a novel approach for automatic disease diagnosis by computer vision techniques, in which a multi-attention convolutional neural network is designed to learn a set of discriminative features for each category. In particular, we first extract convolutional feature representations from an X-ray image, and further propose to learn multi-attention maps by category. As each attention map consists of multiple types of discriminative features to each category, which are generated by a group of convolutional channels and are optimized by corresponding category labels, the subtle visual representation for fine-grained categories can be captured. To make the final prediction, we generate a probability score for each category by top-k pooling from attention maps, which ensures the most discriminative features can be preserved. Extensive experiments on chest X-ray dataset (namely ChestX-ray14) the largest ChestX-ray 14 dataset show the superior results of the proposed network against the CheXNet, with 1.8% relative increases in terms of the AUC metric.