Research On Segmentation Of Synovial MRI Images Of Joints Based On 2D ResU-net Deep Learning Network

Synovitis is a chronic and persistent disease,it is an important pathological manifestation of many osteoarthritis,especially rheumatoid arthritis,early diagnosis of the disease and finding the indicators of bone destruction are of great significance for the assessment of the disease and the formulation of effective treatment plans.In order to accurately determine the patient's condition and develop effective treatment plan,doctors often extract synovial hyperplasia from MRI images by hands and annotation it.However,this method has some problems,such as time consuming,low efficiency and subjectivity.Moreover,the judgment of the severity of the disease and its therapeutic effect often lacks reliability and consistency.In order to solve the above problems,this paper tries to explore the use of AI training to realize the automatic detection of synovial hyperplasia.In the process of automatic detection of synovial hyperplasia,rapid and accurate extraction and segmentation of synovial hyperplasia tissue area from the image is a key step in the whole detection process,which plays a crucial role in assisting clinicians to make a correct diagnosis of the disease in time.In recent years,due to the rapid development of deep learning,image segmentation technology based on deep learning has been widely used in the field of medical image segmentation,among which the U-net algorithm is the most widely used in the field of medical image segmentation.However,due to the wide distribution of synovial lesion areas and the different lesion shapes in each part,the U-net model did not achieve the expected effect in MRI image segmentation of synovial lesion.In order to obtain a more accurate segmentation effect,a new network structure,namely 2D ResU-net network structure is proposed in this paper,in view of the slight shortage of U-net network depth.This Network combined the U-net Network with ResNet(Residual Network)proposed by He et al in the ILSVRC competition in 2015,and carried out batch standardization and other related operations on the Network.With the addition of residual network,the depth of U-net network structure becomes deeper,so as to enhance the feature extraction and classification ability of the network,and at the same time,the gradient vanishing problem of the deep network is effectively solved,thus improving the segmentation effect.In addition,in the process of network training,it is usually necessary to define a loss function to estimate the loss between the predicted value of the model and the real output.Considering that some synovial images have only one or two targets,and the proportion of target pixels is small,which makes network training more difficult,Dice loss was selected as the loss function in this paper.Compared with U-net algorithm,2D ResU-net algorithm DSC coefficient increased by 10.72%,Intersection over Union(IOU)index increased by 4.24% and VOE coefficient of volume overlap error decreased by 11.57%.The results of the experiments show that the proposed algorithm can achieve better automatic segmentation effect for synovial hyperplasia area in MRI images and has a better clinical application prospect.