Research On Intelligent Medical Image Aided Diagnosis System Based On Deep Learning

As the diet level increases,the incidence of various intestinal diseases continues to increase.Colon polyps and diverticulum are common diseases in colonoscopy,which can cause tumors and even cancer in severe cases.Target detection plays an important role in image recognition and classification tasks.It is a very important research topic in the field of computer vision.In recent years,the application of target detection in the medical field has become more and more in-depth.In this paper,a computer-aided diagnosis system(CAD)based on deep learning for polyp detection was proposed to reduce the possibility of colorectal cancer.The system construction mainly consists of five steps: constructing a deep learning framework,preprocessing of colon lesion images,feature extraction training model,performance comparison of lesion detection models,and front-end interface design and development.In order to accurately and quickly identify lesions such as colon polyps and diverticulum,Faster R-CNN and YOLOv3 were used to train the lesion detection models.The best lesion detection model was selected by comparing the two models.Experiments show that the YOLOv3-based lesion detection model has better accuracy and higher speed than the Faster R-CNN-based model.The model obtained using YOLOv3 can achieve an average precision of 90.43% and a recall rate of 98.36% for the testing set.And the detecting speed can reach 31 FPS or more.We further explored the influences of the number of iterations and the learning rate to the performance of the models.Experiments show that in the case of ensuring sufficient data volume,when the model is trained using the YOLOv3 algorithm,the optimal performance lesion recognition model can be obtained quickly when the number of iterations is 30,000 and the learning rate is 0.01.Because the deep learning in the research process requires a large number of data,and the data annotation process is often time-consuming and laborious,we designed an automatic image annotation system,which can be used for labeling large amounts of image data,greatly reducing the workload.Experiments show that the system can complete the labeling work of more than 86% of the data by automatically identifying the images containing the lesion.Based on the lesion detection model established by YOLOv3 algorithm,Vue.js was used to build front-end user interfaces.The front-end communication was completed through the Web Socket protocol.Finally,the development of intelligent medical image-assisted diagnosis system was completed,which provided a method for real-time and high-efficiency auxiliary diagnosis of clinical colonoscopy.