Diagnosis Of Major Depressive Disorder In Adolescents And Prediction Of The Early Response To Antidepressant Medication Based On Radiomics Analysis Of Multiscale Structural MRI

Part 1Normalizing radiomics features from multiscale structural MRI of the adolescent brainObjectives:With the in-depth study of multicenter neuroimaging,there is an urgent need to deal with the abiotic differences introduced by different MRI scanners and acquisition protocols,namely "scanner effect".There are differences in intensity distribution and uneven bias field,which seriously affect the stability of radiomics,and even lead to wrong results.The purpose of this study is to explore the combination of N4 bias field correction,histogram-matching normalization and ComBat harmonization to reduce the "scanner effect" of radiomics features from brain sMRI.Methods:Brain sMR1(3D-T1WI and DTI)was performed in 23 healthy volunteers with three MRI scanners(Philip 1.5T、Philip 3.0T、GE 3.0T).Cat 12 toolkit of SPM12 software and FSL(FMRIB’s software library)software were used for preprocessing respectively and radiomics features of gray matter and white matter were extracted.Then,N4 bias field correction and histogram-matching normalization were performed on the preprocessed T1WI and DTI images.Finally,A.K.software(Artificial intelligence Kit)was used to extract the radiomics features of gray matter and white matter and then Combat harmonization are carried out on the image data.The ShapiroWilk test was used to test the normality of the image feature measurements.The analysis of variance and Tukey HSD test were used to compare the image feature measurements of the three MRI scanners,and the Bartlett spherical test was used to estimate whether the variance between the scanners was uniform.The differences between scanners in the number of radiomics features and numerical statistical distribution in each processing are qualitatively and quantitatively evaluated.Results:There were 10 males and 13 females.There was no significant difference in age(t=1.090,P=0.316),education years(t=-0.638,P=0.574)and CES-D score(t=-0.670,P=0.510)between the sexes(P>0.05).In the original images acquired by the three MRI scanners,the distribution range and peak value of the intensity histogram were not aligned.When N4 bias field correction was performed using 5-level(50 iterations)full mask,the intensity variation coefficient of brain tissue region among the three MRI was the lowest.N4 correction sharpened the intensity peak,HM normalized and aligned each intensity peak,and Combat harmonization to further align the image intensity distribution range and peak of the three MRI.The image calibration process(N4 bias field correction,histogram-matching normalization and ComBat harmonization)has the same influence trend on the calibration of 3D-T1WI and DTI sequences.Through the combination of N4 correction,histogram-matching normalization and Combat harmonization,the percentage of radiomics features with differences between the three scanners was reduced from 88.6%(70/79)before bias field correction to 3.8%(3/79)after ComBat harmonization.Among them,ComBat harmonization is crucial to significantly reduce the number of features with differences.At the same time,the percentage of radiomics features with differences between VOI of gray and white matter increased from 43.0%(34/79)before bias field correction to 84.8%(67/79)after ComBat harmonization.Conclusions:The combination method of N4 bias field correction,histogrammatching normalization and ComBat harmonization can effectively eliminate the"scanner effect" of brain MRI.This could help to incorporate multi-center MRI data across scanners and improve the repeatability of radiomics features.Part 2Diagnosis of MDD and subthreshold depression in adolescents with radiomics analysis based on multiscale structural MRI of brain and comparison with conventional imagingObjectives:At present,the diagnosis of major depressive disorder(MDD)and subthreshold depression(StD)mainly depends on subjective scores and structured interview,which results in risk of misdiagnosis and medical burden.Radiomics is an emerging image analysis framework that provides more details than conventional methods.In present study,we aimed to develop and validate a radiomics classifier for MDD,StD and healthy controls(HC)in adolescents with multiscale brain structural radiomics analysis after normalization,and to compare it with classification model based on the conventional imaging indicators and unnormalized radiomics features.Methods:150 subjects(50 cases of MDD,50 cases of StD and 50 cases of healthy controls matched for sex,age and education)were recruited for brain structural MRI including the three dimensional T1 weighted image(3D-T1WI)and the diffusion tensor imaging(DTI).FreeSurfer/Ants hybrid segmentation algorithm toolkit of SPM12 software and FSL(FMRIB’s software library)software were used for preprocessing respectively.Then,N4 bias field correction and histogram-matching normalization were performed on the preprocessed T1WI and DTI images.Finally,A.K.software(Artificial intelligence Kit)was used to extract the radiomics features of gray matter(GM)and white matter(WM)based on the voxel-based morphometry(VBM)and surfaced-based morphometry(SBM),and then ComBat was used to harmonize the radiomics data.The two-level screening strategy of analysis of variance(ANOVA)and recursive feature elimination was used to reduce the dimension of radiomics features.The classification models of MDD,StD and HC were constructed by using the conventional imaging indicators,unnormalized radiomics features and normalized radiomics features with the support vector mechanism(SVM).The performance and generalization ability of the classification models were evaluated by using the Leaveone-out cross-validation(LOO-CV)and permutation tests.The receiver operator characteristic curve(ROC curve),area under curve(AUC),accuracy,accuracy,recall,specificity,and F1 score based on the LOO-CV results were used to assess the performance of the classifier.The Delong nonparametric test was used for AUC multiple and pairwise comparisons.Results:142 subjects were finally enrolled(43 cases of MDD,49 cases of StD,50 cases of healthy controls).There was no significant difference in age(F=0.430,P=0.674),sex(x2=1.178,P=0.989)and education years(F=0.689,P=0.278)among the three groups(P>0.05).A total of 742 conventional image indicators including 344 features for SBM,206 features for VBM,192 diffusion features for DTI)and 13150 image omics features including 2338 features for SBM,10044 features for VBM,768 diffusion features for DTI were extracted after MRI preprocessing.After two-level dimensionality reduction,only 7 conventional image indicators and 63 image omics features were selected.The overall accuracy of RBF-SVM classifier was 78.30%and 89.21%respectively.Among them,SVM classifier based on radiomics features after combat harmonization had the best performance.The AUC,sensitivity,specificity,and accuracy of discriminating MDD and HC,MDD and StD,StD and HC were 0.928,89.2%,93.2%and 90.5%,0.821,73.0%,85.0%and 80.8%,0.836%,82.4%,77.1%and 79.7%respectively.Permutation tests result was P<0.001.High discriminant radiomics features to distinguish MDD patients and HC subjects were mainly located in the right middle temporal gyrus,bilateral medial orbitofrontal cortex,left superior temporal gyrus,bilateral hippocampus,right inferior temporal gyrus,left posterior cingulate(cingulate)and so on.High discriminant radiomics features distinguishing StD from HC subjects were mainly located in bilateral hippocampus,bilateral inferior temporal gyrus,left superior temporal gyrus,right medial orbitofrontal cortex,right anterior cingulate(cingulate),left corpus callosum,left internal capsule forelimb,right anterior cingulate(cingulate)and so on.High discriminant radiomics features to distinguish MDD and StD subjects were mainly located in the right superior temporal gyrus,bilateral hippocampus,bilateral medial orbitofrontal cortex,left inferior temporal gyrus,left anterior cingulate(cingulate)and so on.Conclusions:These findings provide preliminary evidence that radiomics analysis based on brain multiscale sMRI can effectively distinguish MDD,StD subjects and healthy controls and is superior to the classification model based on conventional image indicators and unnormalized radiomics features.The radiomics features of cuneiform lobe and cerebellum(lobule vi,vii-b and x,4/5 area of cerebellar vermis)have a high weight in diagnosis,indicating that the structural abnormalities of these regions play a key role in the pathophysiology of MDD and StD,and requiring further research.Part 3Prediction of early response of MDD to antidepressant medication in adolescents with radiomics analysis based on brain multiscale structural MRI and comparison with conventional imaging indicatorsObjectives:Due to biological heterogeneity,60%～70%patients with MDD have no response to first-line antidepressant medication(ADM).Neuroimaging biomarkers that can predict the efficacy of early response may be helpful to the personalized selection of initial antidepressant drugs.We aimed to use radiomics strategy to predict the early response to ADM in adolescents MDD by using multiscale structural brain MRI after normalization and to compare it with the prediction model based on conventional imaging indicators and unnormalized radiomics features.Methods:139 hospitalized patients with MDD were recruited for baseline brain structural MRI(3D-T1WI and DTI).After receiving antidepressant selective serotonin reuptake inhibitors(SSRIs)or serotonin norepinephrine reuptake inhibitors(SNRIs)for 2 weeks,the subjects were divided into improved group(SSRIs improvement or SNRIs improvement)and non-improved group according to the early reduction rate of HAMD17 score.N4 bias field correction and histogram-matching normalization were performed on the preprocessed T1WI and DTI images.Then,A.K.software(Artificial intelligence Kit)was used to extract the radiomics features of gray matter and white matter based on the voxel-based morphometry(VBM)and surfaced-based morphometry(SBM),and ComBat was used to harmonize the data.The two-level screening strategy of analysis of variance(ANOVA)and recursive feature elimination was used to reduce the dimension of radiomics features.Support vector machine(SVM)was used to integrate the multiscale sMRI of brain to construct an early efficacy prediction model.The relative importance of each feature to the prediction model was compared and the performance was evaluated by recursive feature elimination algorithm and the Leave-one-out cross-validation(LOO-CV).The ROC curve,AUC,accuracy,sensitivity and specificity based on the LOO-CV results were used to evaluate the performance of the prediction model.The Delong nonparametric test was used for AUC multiple comparisons,and the permutation test was used to evaluate the generalization ability of the model.Results:122 patients with MDD were finally enrolled,including 67 in the ADM improvement group(31 in the SSRIs improved group and 36 in the SNRIs improved group)and 54 in the non-improved group.There was no significant difference in age,sex,education years,family history of mental illness and baseline HAMD-17 score between ADM improved group and non-improved group,SSRIs improved group and SNRIs improved group(P>0.05).After 2 weeks of ADM,the score of HAMD-17 in the ADM improved group(score reduction rate≧20%)was significantly lower than that in the noo-improved group(score reduction rate<20%)(P<0.05).There was no significant difference in HAMD-17 score between SSRIs improved group and SNRIs improved group(P>0.05).A total of 398 conventional image indicators including 35 features for SBM,102 features for VBM,261 diffusion features for DTI)and 2990 radiomics features including 667 features for SBM,1410 features for VBM,913 diffusion features for DTI were extracted after preprocessing.After two-level dimensionality reduction,only 8 conventional image indicators and 49 radiomics features were selected.The overall accuracy of the early efficacy prediction SVM model based on conventional image indicators and radiomics features after normalization was 74.80%and 88.19%respectively.The prediction model based on the radiomics features after normalization had the best performance.The radiomics features that predicted early remission of ADM were mainly located in the anterior medial orbitofrontal cortex,superior frontal gyrus,hippocampus,cingulate gyrus,amygdala,cerebellum,corpus callosum,internal capsule and superior corona radiata.The radiomics features predicting SSRIs remission were mainly located in hippocampus,amygdala,putamen,cerebellum,fornix,and cerebellar peduncle.The radiomics features predicting SNRIs remission were mainly located in the anterior medial orbitofrontal cortex,anterior cingulate gyrus,ventral striatum,accumbency area,knee of corpus callosum,internal capsule and anterior corona radiata.Conclusions:The results of this study suggest that the radiomics prediction model based on multiscale structural brain MRI after normalization at baseline could effectively predict the early treatment response of ADM in MDD patients,and the radiomics features that are predominant and discriminant for SSRIs/SNRIs selection may help for the clinical treatment and drug selection of MDD at the individual level.