MRI-based machine learning models for differentiating Parkinson’s disease and Progressive Supranuclear Palsy

Background Parkinson’s disease (PD) and Progressive Supranuclear Palsy (PSP) are neurodegenerative disorder which have Parkinsonism as core clinical feature. PD and PSP are often misdiagnosed in early phases by conventional magnetic resonance imaging (MRI). Hence the objective of the study is to evaluate accuracy of MRI based machine learning (ML) models for differentiating PD and PSP. Methods This is a retrospective study including total of 96 subjects (32 PD, 32 PSP and 32 healthy controls). The basal nuclei region such as substantia nigra (SN), caudate nucleus (CN), putamen, globus pallidus (GP), red nucleus (RN), and subthalamic nucleus (STN) were manually segmented in T2 weighted fluid attenuation inversion recovery (T2W-FLAIR) images and radiomics features (RF) were extracted. The top 20 RF were used to train and test ML models. Performance measures of ML models for differentiating PD, PSP and HC were evaluated. Results For differentiating PD from PSP, SN consistently showed highest performance achieving accuracy, F1score and AUC values of 0.95–1.00 across all the tested ML models. RN and STN also demonstrated near perfect classification with several ML models with accuracy and AUC values of 1.00. CN, putamen and GP showed moderate performance (accuracy 38% - 85%). Among classifiers, AdaBoost provided robust performance achieving AUC up to 0.89 and MCC up to 0.72 in test data sets. Conclusion MRI based radiomics models from SN, STN and RN provide the most accurate classification of PD and PSP. It can also help in diagnosis of PD and PSP which are undifferentiable in early stages of disease.