Establishment And Effectiveness Of An Automatic Scoring Model For Medial Temporal Lobe Atrophy Based On Deep Learning

Objective Alzheimer’s disease(AD)is a common neurodegenerative disease among the elderly.Currently,medical personnel in primary care institutions in China are not sufficiently aware of AD and lack the appropriate diagnostic skills to make a timely diagnosis of AD.The purpose of this study was to develop a deep learning-based automatic scoring model for medial temporal lobe atrophy(MTA)and to confirm its validity,aiming to eliminate differences between scorers and within the scorers themselves,and to eliminate the time-consuming scoring process,in order to help promote the application of MTA scoring for AD diagnosis and to assist physicians in carrying out a more accurate and faster diagnostic process.Methods First,530 coronal hippocampal MRI images were selected,and the MTA scores were given by experienced neurologists and radiologists,and the hippocampal regions were annotated using ITK-SNAP software,and the annotated hippocampal images were submitted to a software development company for training and modeling.The annotated hippocampal images were sent to a software development company for training and modeling,and then 360 coronal hippocampal MRI images of first-time patients were collected and scored by the trained MTA automatic scoring model and physicians manually separately,and the consistency of the two scores was compared.Results 1.Among the 360 included visits,including 67 patients with subjective cognitive disorder(SCD),104 patients with mild cognitive impairment(MCI),142 patients with AD,26 patients with epilepsy,and 21 patients with other types of dementia.There were 160 cases(44.4%)of males(160/360)and 200 cases(55.6%)of females(200/360)with a mean age of(71.12±10.37)years,ranging from 33 to 93 years.The cases were divided into 5 groups according to MTA score scores:14 cases in MTA0 grouping,58 cases in MTA1 grouping,140 cases in MTA2 grouping,121 cases in MTA3 grouping,and 27 cases in MTA4 grouping.2.The agreement rate between model scores and manual scores was 77.8%on the left and 79.4%on the right.Weighted Kappa analysis of model scores and manual scores yielded a weighted Kappa value of 0.776(95%CI:0.731-0.820,p<0.001)for the left side and 0.787(95%CI:0.741-0.832,p<0.001)for the right side.3.The ROC curve was used to predict the cognitive function assessment scale’s screening efficacy,the diagnostic efficacy of MoCA was better than that of MMSE for MCI,while the diagnostic efficacy of MMSE was better than that of MoCA for AD;the diagnostic efficacy of combining both MMSE and MoCA scores was better than that of individual scores for both MCI and AD.4.MMSE scores were positively correlated with MoCA scores,with a Pearson coefficient of(r=0.957,p<0.001,95%CI:0.946-0.965).Conclusion Our developed deep learning-based automatic scoring model for medial temporal lobe atrophy can achieve a high degree of consistency with the scores of experienced clinicians and imaging physicians,is simple and fast to operate,mimics the scoring procedures of clinicians,and assists primary care workers in deriving timely and accurate patient MTA scores,which has great value for clinical promotion and application.