The Study Of Classification Method In Conduct Disorder Based On SMRI And Machine Learning

Conduct disorder(CD)is a psychiatric disorder in either childhood or adolescence and is characterized by aggressive and antisocial behavior.Although CD has been shown to be associated with structural abnormalities by structural magnetic resonance imaging(sMRI),whether these structural alterations can distinguish CD from healthy controls(HCs)remains unknown.Here,we quantified these structural differences and explored the classification ability of these quantitative features based on machine learning.High-resolution 3D sMRI was acquired from 60 CD subjects and 60 age-matched HCs.Firstly,the grey matter volumes were extracted as the morphological features,and input into three machine learning classifiers: logistic regression,random forest and support vector machine(SVM).Multivoxel pattern analysis(MVPA)was used to characterize the brain structural differences’ spatial patterns.Finally,we applied three deep convolutional neural network(CNN),frameworks,namely,Alex Net,ResNet and ResNeXt,to automatically extract multi-layer high dimensional features of sMRI images,and to complete the classification of CD.The classification performance of different models were evaluated by comparing the receiver operating characteristic(ROC)curves of traditional machine learning and CNN.We found that these traditional machine learning models,including logistic regression,random forest and SVM,achieved comparable classification performance,with area under the ROC curve(AUC)of 0.76~0.80.The AUC of logistic regression,random forest,SVM with RBF kernel were compared with the AUC of SVM with linear kernel,respectively,and no significant difference was detected.The accuracy were 77.9%~80.4%,specificity were 73.3%~80.4%,sensitivity were 75.4%~87.5% based on the optimal threshold of ROC curves.MVPA detected significant grey matter volume difference in supramarginal gyrus which was not found by traditional VBM,and the highest classification accuracy based on the brain regions detected by MVPA reached 83.0%.The models based on Alex Net achieved better classification results compared to traditional machine learning methods,and the AUC of AlexNet reached 0.88 which is much higher than that of SVM with linear kernel(p = 0.061).To summarize,based on sMRI and machine learning,we successfully established the classification models of CD with high accuracy compared to the similar studies in literature.For traditional machine learning and MVPA methods,feature extraction in requires manual intervention and is based on some assumptions,however the model training and testing is relatively simple.The classification models based on CNN do not require manual intervention and specific assumptions to extract features,and the results of classification are much better than traditional machine learning methods,however the training requires large-scale data,and the time cost of network structure design and model training was higher.In conclusion,the machine learning-based classification models of CD could be a reliable diagnostic tool in assisting and promoting the clinical diagnosis of CD.