Research On Segmentation Of Synovial Mri Images Of Joints Based On Dense-UNet++ Deep Learning Network

Rheumatoid arthritis is a serious multi-system immune disease that cannot be cured and has multiple complications.Synovial hyperplasia is one of the pathological changes of rheumatoid arthritis.Synovial hyperplasia not only invades the synovial tissue of human limbs and joints,but also destroys human nerves,blood vessels and other important organs or tissues.It usually manifests as synovial hyperemia,edema,and exudation.The synovium thickens and erodes the tissues,and is accompanied by the spread of effusion,which can cause nerve necrosis in the joints.Therefore,it is of great medical value to be able to diagnose the disease in the early stage and take corresponding treatment measures to improve the degree of bone and joint invasion and reduce or delay the disability rate.Compared with CT images and color ultrasound images,MRI images have high definition of soft tissue imaging and can be imaged in any direction,so synovium’s MRI images can more clearly show the changes of joints and synovial hyperplasia to a certain extent.In order to accurately segment the synovial lesion area and formulate an effective treatment plan,the doctor needs to manually mark the MRI images.This method has two problems.First,it takes too long and is inefficient.The segmentation of the synovium requires careful judgment of the boundary,and manual adjustment of the segmentation of the synovial boundary is required,which needs high energy and attention.Secondly,different doctors will make different judgments on the same MRI image due to different experiences.This method is highly subjective and cannot accurately analyze the specific conditions of the synovium.In recent years,many researches had made major breakthroughs in network architecture,and the accuracy of deep learning segmentation algorithms were getting higher and higher.At the same time,it had also been widely used in the field of medical imaging.In the past,the deep learning models such as UNet,Res UNet,etc.,were more efficient in detecting hyperplastic synovium than doctors’ manually delineating.However,since the synovial lesion area has no fixed rules and the sizes of the lesion are different,those models have poor accuracy,high training costs and poor robustness.In order to obtain a network model with high accuracy,good robustness and low training cost,this paper proposed a new network called Dense-UNet++,inserting the Dense modules into the UNet++ network,and using the Swish activation function for training.The network was trained using 14512 synovial images augmented by 1104 synovial magnetic resonance image data,which were collected from two different hospitals’ 77 patients.Finally,the average segmentation accuracy of the model can reach 99.31% of the DSC coefficient,which was greatly improved compared with the Res UNet algorithm.When compared with Res UNet,the result was increased by29.33%,the training time of this model was reduced by 47.3%,and the stability variance was reduced by 93.75%.The results showed that the algorithm in this paper can segment the area of ??synovial hyperplasia and obtain the occupied area and assist doctors in calculating the volume of the synovial membrane and formulating related treatment plans.