Diagnosis of Subcortical Ischemic Vascular Cognitive Impairment With No Dementia Using Radiomics of Cerebral Cortex and Subcortical Nuclei in High-Resolution T1-Weighted MR Imaging

To investigate whether the combination of radiomics derived from brain high-resolution T1-weighted imaging and automatic machine learning could diagnose subcortical ischemic vascular cognitive impairment with no dementia (SIVCIND) accurately. A total of 116 right-handed participants involving 40 SIVCIND patients and 76 gender-, age-, and educational experience-matched normal controls (NM) were recruited. A total of 7,106 quantitative features from the bilateral thalamus, hippocampus, globus pallidus, amygdala, nucleus accumbens, putamen, caudate nucleus, and 148 areas of the cerebral cortex were automatically calculated from each subject. Six methods including least absolute shrinkage and selection operator (LASSO) were utilized to lessen the redundancy of features. Three supervised machine learning approaches of logistic regression (LR), random forest (RF), and support vector machine (SVM) employing 5-fold cross-validation were used to train and establish diagnosis models, and 10 times 10-fold cross-validation was used to evaluate the generalization performance of each model. Correlation analysis was performed between the optimal features and the neuropsychological scores of the SIVCIND patients. Thirteen features from the right amygdala, right hippocampus, left caudate nucleus, left putamen, left thalamus, and bilateral nucleus accumbens were included in the optimal subset. Among all the three models, the RF produced the highest diagnostic performance with an area under the receiver operator characteristic curve (AUC) of 0.990 and an accuracy of 0.948. According to the correlation analysis, the radiomics features of the right amygdala, left caudate nucleus, left putamen, and left thalamus were found to be significantly correlated with the neuropsychological scores of the SIVCIND patients. The combination of radiomics derived from brain high-resolution T1-weighted imaging and machine learning could diagnose SIVCIND accurately and automatically. The optimal radiomics features are mostly located in the right amygdala, left caudate nucleus, left putamen, and left thalamus, which might be new biomarkers of SIVCIND.