Research And Application Of Medical Image Lesion Region Detection Based On YOLOv5

The rapid development of computer vision has made target detection one of the tasks of great interest.In the medical field,automated lesion detection can significantly improve the diagnostic efficiency of physicians.The accuracy of lesion detection plays a crucial role in computer-aided diagnosis,and lesion detection results directly affect subsequent disease diagnosis and treatment.More and more scholars are focusing on lesion region detection methods using deep learning techniques.In order to make lesion detection more accurate,research on YOLOV5-based lesion region detection methods has been carried out,and the main research contents are as follows:(1)Aiming at the problems of complex background and inadequate feature extraction for lesion detection in medical images,a lesion region detection method based on the dual architecture of Transfomer and CNN is proposed.Combining the coding features of convolutional neural network and Transformer,a new multi-level feature fusion module is designed in this paper.By continuously downsampling in the CNN branch,the perceptual field is gradually increased,and the features are encoded from local to global.In the Transformer branch,the local detail information of the image is gradually recovered by implementing a global self-focus mechanism.The multi-level feature fusion module is used to fuse the same resolution features extracted from each branch to achieve better feature extraction and effectively avoid the drawback of building a very deep network to retain low-level background information.Moreover,by increasing the prediction layer of the network and adding more anchor frames on the feature map with higher resolution,the capture capability of small targets can be improved.Experiments show that this method maintains an advanced level of detection accuracy and speed compared with the current mainstream networks.(2)Based on the first research content,it is found that the independence between feature maps before each prediction output is strong,the feature interaction is not sufficient,and the model cannot be flexibly adapted to the prediction problem based on the corresponding preset anchor frame size under multiple feature scales,and a lesion region detection method based on pairwise attention and feature enhancement is proposed.This method fuses multi-scale features corresponding to residual networks and is able to adaptively fuse local features and their global dependencies,which not only makes the features diverse but also improves the amount of information in the deep network.The effectiveness of the method is demonstrated experimentally,compared with other competitive methods in the same field,73.2% can be achieved under AP@50 index.(3)The lesion area detection system is designed and developed based on the previous research content.The system includes three major modules,namely,login module,detection module and management module.The system uses the VUE framework to build the front-end view and the Spring Boot framework to build the back-end server,which is accessed through URL addresses.The system aims to assist doctors to diagnose patients accurately and quickly,and to help promote the research and application of YOLOv5-based medical image lesion area detection technology.