Lung nodule detection method based on improved YOLOv4

Abstract

Aiming at the disadvantages of low accuracy and slow speed in current lung nodule detection methods, this paper proposes a lightweight lung nodule detection algorithm based on improved YOLOv4. The first is data preprocessing, such as resampling of the open-source data set LUNA16, lung parenchymal segmentation, etc. Then, based on the YOLOv4 network model, replace the original backbone feature extraction network CSPDarknet53 network with the Mobilenet-v3 network to speed up the detection of lung nodules. Use deep separable convolution block to replace the $3\times 3$ ordinary convolution in the enhanced feature extraction network structure, which greatly reduces the number of model training parameters. The K-means++ clustering method is used to replace the original K-means clustering method to perform cluster analysis on the data set, and 9 anchor boxes of different sizes suitable for lung nodule detection are obtained again. Finally, the improved YOLOv4 detection algorithm is used to detect lung nodules. The experimental results on the LUNA16 data set prove that the improved network model mAP value can reach 96.71%, and the detection speed can reach 41.99FPS. Compared with other network models, the experimental results prove that the improved network has better and faster detection results and is feasible.