Research On Segmentation Method Of Thyroid Tumor Ultrasound Image Based On Deep Learning

Papillary thyroid carcinoma is a very common malignant tumor,and its incidence has gradually increased in recent years,but its harm to the human body is relatively small,and it is a cancer that can be monitored in real time and treated conservatively.However,the distribution of medical resources in my country is uneven,and the lack of resources such as tertiary hospitals and senior doctors in many areas hinders the diagnosis,monitoring and treatment of papillary thyroid cancer.Ultrasound diagnostic images of patients with papillary thyroid carcinoma contain rich medical information,and there are obvious diagnostic criteria and rules.In addition,in recent years,many progresses have been made in the field of computer vision and deep learning in the direction of image analysis tasks.Therefore,this thesis is dedicated to developing a system that can effectively analyze the ultrasound images of patients with papillary thyroid cancer by means of computer artificial intelligence-assisted medical diagnosis.Or the patient himself provides an important basis for evaluating the condition.The main achievements of the thesis are as follows:Firstly,the thesis constructs an ultrasound dataset of patients with papillary thyroid cancer,including the implementation of annotation tools,the structure design of the dataset,data collection,annotation and desensitization,etc.,and finally sorted out.Two types of ultrasound data sets,with a total of thousands of images;secondly,the thesis designs a multi-level,multi-angle thyroid ultrasound image preprocessing process,through morphological processing,C-MMDetection and data enhancement methods,from the effective image quality Area,image quality and number of available images effectively improve the availability of thyroid ultrasound image datasets,laying the foundation for subsequent image segmentation tasks;thirdly,the thesis introduces a saliency target detection network to solve the problem in this thesis.For the segmentation of thyroid ultrasound images,based on the latest research results of saliency target detection,a refinement module and a result fusion module are introduced and U2-RNet is proposed,which further improves the image segmentation performance of the network.The comparison and segmentation experiments of our U2RNet and other classic image segmentation networks are also carried out,and the experimental results are analyzed and discussed through four evaluation indicators of Dice Similarity coefficient,precision rate,recall rate and F1 score.Finally,the effectiveness and superiority of U2-RNet on the task of thyroid ultrasound image segmentation are strongly demonstrated.