Research And Application Of Auxiliary Diagnosis Of MS And NMOSD Based On Joint Deep Learning

Multiple sclerosis(MS)and neuromyelitis optic spectrum disorder(NMOSD)are demyelinating diseases of the central nervous system,autoimmune diseases,and have a very high disability rate.The clinical symptoms and imaging manifestations of these two diseases have overlapping characteristics,making it difficult to diagnose and differentiate.However,the prognosis and optimal treatment plan of the two are different,and even the drugs for treating MS such as interferon can promote the exacerbation of NMOSD.Therefore,early accurate diagnosis and timely intervention and treatment are the key to improving the prognosis and quality of life of patients.Magnetic Resonance Imaging(MRI)is an important method for diagnosing neurological diseases because of its high soft tissue resolution.Due to the multifocal and similar imaging features of MS and NMOSD,it is time-consuming and challenging to manually screen the lesion area and make an accurate diagnosis using MRI images.In recent years,the vigorous development of deep learning has brought great changes to medical image analysis and image-based computer-aided diagnosis.Although there are related studies on lesion segmentation and disease classification in MS and NMOSD,the existing models are only designed for a single task,ignoring the potential correlation between the two tasks.In response to the above problems,this paper proposes a joint model for MS and NMOSD lesion segmentation and disease classification,which uses the correlation between the two tasks to make full use of the information in the data in a mutually guided manner,thereby improving each task.performance.The joint model mainly includes three parts: information sharing subnetwork,lesion segmentation subnetwork and disease classification subnetwork.Among them,the information sharing subnetwork adopts a dual-branch structure composed of convolution module and Swin Transformer module to extract local and global features respectively.Then input the lesion segmentation subnetwork and disease classification subnetwork,and obtain the results of the two tasks at the same time.In addition,in order to further enhance the mutual guidance between the two tasks,this paper proposes two information interaction methods: one is to combine the lesion probability map obtained by the lesion segmentation subnetwork as prior information into the disease classification through the lesion guidance module(LGM).sub-network to help the classification task focus on the lesion area and reduce the interference of other brain tissues;the second is to propose a cross-task loss function.In the segmentation sub-network,a multi-layer lesion localization map is generated according to the category visualization mechanism,and it is separated from the lesion The multi-scale feature maps obtained by the subnets are mutually supervised by cross-task losses to improve the localization ability of the segmentation subnetwork for discrete point-like lesions and the accuracy of the classification subnetwork.At the same time,the lesion localization map provides interpretability for the diagnosis process of the deep learning model.This paper compares the proposed joint model with the advanced classification and segmentation model.The results show that the joint model can effectively improve the performance of the two tasks.By setting up ablation experiments,the effectiveness of information sharing and interaction between tasks is verified.In order to apply the joint model to clinical work,this paper designs and implements a medical aided diagnosis platform for MS and NMOSD.The platform includes basic interaction module,lesion segmentation module,MS/NMOSD classification module and joint diagnosis module,so as to provide effective assistance for doctors.