MRI-based Brain Structural Changes And Deep-Learning Model For Differential Diagnosis Of Neuromyelitis Optica Spectrum Disorder

Objective:Neuromyelitis optica spectrum disorder(NMOSD)is an autoimmune inflammatory demyelinating disease of the central nervous system.The neuropathological mechanisms of NMOSD are not fully understood.We used neurite orientation dispersion and density imaging(NODDI)and high-resolution T1-weighted imaging(T1WI)to investigate changes of brain structures in patients with NMOSD,aiming to provide imaging information for neuropathological mechanisms.Moreover,the differential diagnosis of NMOSD from multiple sclerosis(MS)and myelin oligodendrocyte glycoprotein antibody-associated disease(MOGAD)is of great significance for precise treatment and prognosis.Thus,we proposed an advanced deep-learning model for classification of NMOSD,MS and MOGAD based on brain and spinal cord MRI,aiming to assist in differential diagnosis in clinical practice.Methods:(1)Study 1(White matter structural changes in NMOSD evaluated by NODDI model):Forty-three patients with NMOSD and 39 healthy controls(HC)matched for age,sex,and educational background were included.Individual maps of intracellular volume fraction,isotropic volume fraction,and orientation dispersion index(ODI)were compared between groups using tract-based spatial statistics analyses.Correlation analyses between NODDI parameters and scores on the Digit Symbol Substitution Test(DSST),Trail Making Test(TMT),and the Extended Disability Status Scale(EDSS)and disease durations were evaluated.Moreover,NODDI parameters were compared between NMOSD subgroups based on presence or absence of optic neuritis episodes.(2)Study 2(Cerebral cortex structural changes in NMOSD evaluated by surface-based morphometry analysis):Forty-three patients with NMOSD and 45 HCs matched for age,sex,and education background were included.Surface-based morphometry analyses of high-resolution T1WI were used to calculate the cortical thickness,sulcal depth,and gyrification index between groups.Correlation analyses were conducted to determine whether any surface parameters correlate with clinical characteristics including DSST,TMT,and EDSS scores and disease durations.Surface parameters were compared between NMOSD subgroups based on presence or absence of optic neuritis episodes.(3)Study 3(MRI-based deep learning for classification of demyelinating diseases of the central nervous system):A total of 290 patients diagnosed with demyelinating disease of the central nervous system from August 2013 to October 2021 were included,including 67 patients with MS,162 patients with AQP4 antibody-positive(AQP4+)NMOSD and 61 patients with MOGAD.We established a multi-location MRI-based deep-learning model based on the state-of-art transformer network and multi-instance learning method.The performance of the model was evaluated by the area under the receiver operating characteristic curve(AUC),accuracy,Cohen’s Kappa(Kappa),and Matthews correlation coefficient(MCC).The classification performance of the deep-learning fusion model based on multi-location(head and spinal cord)MRI,the model based on single-location(head or spinal cord)MRI,and two radiologists were compared.Results:(1)Study 1(White matter structural changes in NMOSD evaluated by NODDI model):①Patients with NMOSD exhibited a higher ODI in the corpus callosum,left corona radiata,left superior longitudinal fasciculus,and right superior corona radiata than HCs(P<0.05).However,the intracellular volume fraction and isotropic volume fraction of the NODDI model showed no significant difference between groups(P>0.05).②Performance on the DSST and TMT was significantly worse in patients with NMOSD than in the HCs(P<0.001).Correlation analysis indicated that the mean ODI in the cluster that tested significant was negatively correlated with DSST scores,and positively correlated with TMT and EDSS scores(P<0.05).However,there was no significant correlation between the mean ODI in the cluster and disease durations(P>0.05).③ Compared with NMOSD patients without any optic neuritis episodes,those who had clinical episodes of optic neuritis showed a higher ODI in the left retrolenticular part of the internal capsule,posterior thalamic radiation,sagittal stratum,superior longitudinal fasciculus,and right anterior corona radiata(P<0.05).(2)Study 2(Cerebral cortex structural changes in NMOSD evaluated by surface-based morphometry analysis):①Cortical thickness in the bilateral rostral middle frontal gyrus and left superior frontal gyrus was lower in the patients with NMOSD than in the HCs(P<0.001).However,no significant differences of sulcal depth and gyrification index was found between groups(P>0.05).②Cortical thickness in the bilateral rostral middle frontal gyrus was positively correlated with scores on the DSST and negatively correlated with scores on the TMT and the EDSS(P<0.05).However,mean cortical thickness in the left superior frontal gyrus did not correlate with clinical parameters(P>0.05),and no significant correlations were observed between mean cortical thickness in the identified brain regions and disease duration(P>0.05).③Subgroup analysis of the patients with NMOSD indicated that compared with those who did not have any optic neuritis episodes,those who did have such episodes exhibited noticeably thinner cortex in the bilateral cuneus,superior parietal cortex,and pericalcarine cortex(P<0.01).(3)Study 3(MRI-based deep learning for classification of demyelinating diseases of the central nervous system):①In the binary classification task,the AUCs of the deep-learning fusion model based on both head and spinal MRI were 0.942(95%CI:0.879-0.987)and 0.803(95%CI:0.629-0.949)when identifying AQP4+NMOSD and MOGAD,respectively,which were higher than that of models based on brain or spinal cord MRI alone.In the identification of MS,the AUC of fusion model was 0.933(95%CI:0.848-0.991),which was slightly lower than that of the model based on brain MRI alone,but higher than that of the model based on spinal cord MRI alone.②In the multi-category classification,the accuracy of fusion model was 81.4%(Kappa 0.666,MCC 0.682),which was higher than that of the models based on brain MRI(accuracy 75.9%,Kappa 0.605,MCC 0.623)or spinal MRI(accuracy 62.7%,kappa 0.202,MCC 0.215)alone.③The accuracy of the fusion model was significantly higher than that of the junior radiologist(81.4%vs.64.4%,P<0.05),and slightly higher than that of the senior radiologist(81.4%vs.74.6%,P>0.05).The fusion model achieved a Kappa of 0.666 and MCC of 0.682,which were higher than those of the junior radiologist(Kappa 0.426,MCC 0.431)and the senior radiologist(Kappa 0.576,MCC 0.578).Conclusions:(1)Abnormal orientation dispersion in corpus callosum,left corona radiata,left superior longitudinal fasciculus,and right superior corona radiata,and cortical atrophy in the bilateral middle frontal gyrus and left superior frontal gyrus were found in patients with NMOSD.Some of these changes may be structural basis of clinicopathological mechanisms of cognitive decline.(2)Compared with those who did not have optic neuritis episodes,NMOSD patients with optic neuritis episodes exhibited abnormal orientation dispersion in several white matter tracts(including optic radiation)and cortical atrophy in the bilateral cuneus,superior parietal cortex,and pericalcarine cortex.These may be secondary injury to the posterior optic pathway after the onset of optic neuritis.(3)The Transformer-based deep-learning model based on brain and spinal cord MRI for classification of AQP4+NMOSD,MS,and MOGAD,may assist in differential diagnosis of demyelinating diseases.