YOLO-lung: A Practical Detector Based on Imporved YOLOv4 for Pulmonary Nodule Detection

Automatic pulmonary nodule detection based on CT images plays a key role in the screening of lung cancer. In recent years, deep learning techniques have a significant progress in this field. However, there is often a huge volume of medical data but a limited hardware in most hospitals. It is often necessary to sacrifice accuracy to ensure the infer speed of the detector in practical applications. Therefore, it is meaningful to comprehensively consider the balance between the effectiveness and efficiency of the detector. The aim of this paper is to implement a pulmonary nodule detector with a relative balance of effectiveness and efficiency that can be applied directly in the hospital field, rather than to propose a new model. Our work is completed in two stages. Firstly, we mainly attempts to improve the accuracy of the model by combining various existing techniques: the depthwise over-parameterized convolution layer, the convolutional block attention module and focal loss function. Finally, we perform redundant channel pruning on the designed model to obtain a more efficient pulmonary nodule detector, named YOLO-lung. By combining multiple tricks, extensive experimental results on the LIDC-IDRI dataset show that YOLO-lung can achieve better balance between effectiveness (90.5% AP) and efficiency (25 FPS). Compared with several state-of-the-art detection methods, YOLO-lung has better detection performance. The method presented in this paper is of reference significance for the design of the practical pulmonary nodule detection model.