Deep Learning-Based Detection of Lung Lesions in Medical Images

This paper designs a target detection algorithm based on Transformer network architecture. Swin Transformer was used to extract features of lesion areas in lung medical images based on zero basic training, and the extracted feature images were embedded into the detection segmentation network Mask R-CNN to accurately locate and classify all suspected lesion areas in lung medical images. Multi-scale feature extraction and feature fusion are carried out through network, which solves the problems of focus region recognition caused by single resolution, weak communication between different layers and lack of attention fusion. After submitting the test on the open data set, the mAP accuracy can reach 0.118, which is better than the mainstream target detection algorithm and provides a reference scheme for the application of computer-aided diagnosis system.