Research On Adenomyoma Ultrasound Image Segmentation Method Based On Deep Learning

Adenomyoma is a disease that exists in the muscular layer of the uterus.Its clinical symptoms and the pain have seriously troubled women’s physical and mental health.Ultrasound examination is used as the main imaging method for the diagnosis of clinical gynecological diseases due to its advantages of non-destructive and low price.It is often used in the initial screening of uterine diseases and postoperative diagnosis.With the continuous development of deep learning,cross-research with the medical field is also booming,and computer-aided diagnosis systems will gradually be widely used in clinical medicine.Due to the shortcomings of the noise and artifacts of the ultrasound image,its development in the field of artificial intelligence image segmentation is severely limited.Different from the superficial diseases in the body,the boundary of abdominal ultrasound tumor image is not clear and the noise interference is large,which also makes the segmentation difficult.In order to solve the above problems,at the same time reduce the workload of physicians and meet the needs of rapid and efficient clinical diagnosis,this topic attempts to segment the adenomyoma ultrasound image to fill the gap of deep learning in this field.Nearly 2000 ultrasound images of more than 200 adenomyoma patients were collected from Beijing Maternity Hospital,and the lesion images were annotated and preprocessed to construct an ultrasound data set of adenomyoma.The following studies were carried out:First of all,in view of the shortcomings of traditional image segmentation methods,this thesis designs two different deep learning methods to achieve the segmentation of adenomyoma.Using deeplab model,we optimize the edge details of the lesions by dilated convolution algorithm and fully connected CRF,and compare with the U-Net and FCN-8s network to choose better results.Then we use the Mask RCNN instance segmentation model,effectively extracts features through the ResNet structure,and combine with RPN network to realize effective use of features,and constantly optimize network training to realize the fine segmentation of the focus area,proving the accuracy and feasibility of the two models in clinical application.Secondly,based on the above model algorithm,GUI interface design of PyQt framework,a uterine adenomyoma segmentation system is built.The interactive and convenient medical image auxiliary segmentation system interface is designed to assist doctors to quickly segment the lesion,providing numerical basis for the effect evaluation before and after the operation.Finally,this paper proposes to introduce the uterine adenomyoma ultrasound image segmentation algorithm based on deep learning into the evaluation of the efficacy of HIFU ablation technology.Since the lesion area will not completely disappear after HIFU treatment,the ablation rate needs to be calculated through the image.Therefore,based on the test set image results of the previous model,the ablation rate is tracked and calculated according to the clinical diagnosis requirements,and then the practical value of the model is verified from the perspective of clinical application,so as to achieve the cross application of deep learning and medical image segmentation,which proves its wide application prospect.