Automatic classification of MSA subtypes using Whole-brain gray matter function and Structure-Based radiomics approach

Abstract

Background: This study aims to develop a radiomics method based on the function and structure of whole-brain gray matter to accurately classify multiple system atrophy with predominant Parkinsonism (MSA-P) or predominant cerebellar ataxia (MSA-C).

Methods: We enrolled 30 MSA-C and 41 MSA-P cases for the internal cohort and 11 MSA-C and 10 MSA-P cases for the external test cohort. We extracted 7,308 features, including gray matter volume (GMV), mean amplitude of low-frequency fluctuation (mALFF), mean regional homogeneity (mReHo), degree of centrality (DC), voxel-mirrored homotopic connectivity (VMHC), and resting-state functional connectivity (RSFC) from 3D-T1 and Rs-fMR data. Feature selection was conducted with t-test and least absolute shrinkage and selection operator (Lasso). Classification was performed using the support vector machine with linear and RBF kernel (SVM-linear/SVM-RBF), random forest and logistic regression. Model performance was assessed via receiver operating characteristic (ROC) curve and compared with DeLong's test.

Results: Feature selection resulted in 12 features, including 1 ALFF, 1 DC and 10 RSFC. All the classifiers showed remarkable classification performance, especially the RF model which exhibited AUC values of 0.91 and 0.80 in the validation and test datasets, respectively. The brain functional activity and connectivity in the cerebellum, orbitofrontal lobe and limbic system were important features to distinguish MSA subtypes with the same disease severity and duration.

Conclusion: Radiomics approach has the potential to support clinical diagnostic systems and to achieve high classification accuracy for distinguishing between MSA-C and MSA-P patients at the individual level.