Clinical Staging Of Thyroid Associated Ophthalmopathy Based On Convolution Neural Network

Objective:This study aimed to establish a deep learning system for detecting the active and inactive phases of thyroid-associated ophthalmopathy(TAO)using magnetic resonance imaging(MRI).Background and content:TAO is an orbital autoimmune disease.The stage of disease course and the severity classification are the basis of treatment.Clinical activity score(CAS),the commonly used method for staging,has some shortcomings,such as strong subjectivity,low sensitivity and low accuracy,while the indicators used for severity classification are always tedious and difficult to use.Because of the advantages of no radiation damage,multi-parameter imaging and high soft tissue resolution,MRI is used to assist the grading and staging of TAO.However,due to the limitation of professional knowledge and cognition,the reduced-dimensional interpretation of orbital MRI leads to the loss of depth information.With the development of hardware and deep learning algorithm,convolution neural network(CNN)is widely used in medical image recognition.This study systematically investigated the progress of orbital MRI,established a medical image database of TAO,and proposed an intelligent staging system based on CNN.Methods and results:A retrospective analysis was conducted on 108 patients with TAO in the ophthalmology clinic of the Ninth People's Hospital affiliated to Shanghai Jiao Tong University School of Medicine from May 1^st,2018 to July 1^st,2019.The database was established by integrating the basic information,medical status and medical imaging data of the patients.160 MRI images of these patients and their corresponding CAS were obtained from the database.Network A based on CNN was established to distinguish active and inactive patients,the accuracy,specificity and sensitivity of which were 86.3%,0.896 and 0.750 respectively.Because of the phenomenon of vanishing gradient in the process of training,we introduced the structure of residual neural network to establish network B,the accuracy of which was 85.5%.Network B could improve the sensitivity(0.821)while maintaining a good specificity(0.865).In addition,this paper also evaluated the performance through the confusion matrix and the receiver operating characteristic curve,explained the networks by visualization methods,and explored its potential of therapeutic effect evaluation.Conclusion:The depth characteristics for staging were included in the orbital MRI images.CNN could mine the depth features to assist the staging of TAO and doesn't rely on subjective judgment,with small measurement error and strong robustness.The network established in this study performed well in terms of accuracy,specificity and sensitivity.Its establishment and promotion could standardize the diagnosis of thyroid-associated ophthalmopathy and speed up the decision-making process.