A Coarse Feature Reuse Deep Neural Network for CXR Lesion Detection

Abstract

Lung disease screening using Chest x-ray (CXR) radiographs can obviously decrease the incidence of lung cancer. Using computer-aided diagnosis system to assist doctors in lung disease screening can greatly improve the diagnosis efficiency. In this paper, a coarse feature reuse deep neural network for CXR lesion detection is proposed. Firstly, we design a coarse feature reuse (CFR) block that can reuse low-level semantic features and extract high-level semantic information, which is used to replace the max-pooling layer in the shallow part of the network to achieve better feature extraction. A novel backbone network - RRCNet, which combines RepVGG block and Resblock, is proposed. The RepVggblock is used for better feature extraction at shallow layers and the Resblock is used for better feature fusion at deep layers. Extensive experiments on VinDr-CXR dataset demonstrate that our RRCNet-based detection network outperformes other classic detectors on both mAP (17.67%) and inference speed (0.1426s).