The Magnetic Resonance Parkinsonism Index (MRPI) is a neuroimaging biomarker designed to differentiate Progressive Supranuclear Palsy (PSP) from Parkinson’s disease (PD). Despite its strong diagnostic accuracy, its clinical use is limited by the need for manual segmentation of brain structures, a time-consuming and rater-dependent process. This study introduces an automated workflow to compute MRPI, optimizing its diagnostic performance and reducing variability.

A total of 497 participants (135 PSP, 218 PD, 28 MSA, 116 HC) underwent MRI on a 3T GE MR750 scanner. Brain regions (midbrain, pons, superior and middle cerebellar peduncles) were automatically segmented and refined using atlases (Figure 1). Two MRPI indices were computed: a raw index derived directly from atlas-based segmentation and a refined index, manually adjusted by an expert to retain only the regions most affected in PSP. To assess diagnostic performance, we evaluated both indices using machine learning models (Random Forest, Logistic Regression) with 10-fold cross-validation.

Comparing PSP against other groups, the refined MRPI index demonstrated superior diagnostic accuracy. In distinguishing PSP from PD, it achieved an AUC of 0.91, sensitivity of 0.96, specificity of 0.62, and accuracy of 0.83 (Figure 2). The confusion matrices for Logistic Regression and Random Forest further illustrate the classification performance (Figure 3). These results confirm that the refined index provides more reliable classification than the raw MRPI.

By automating MRPI calculation, this approach minimizes observer variability and ensures greater standardization across clinical settings. The integration of refined indices enhances diagnostic accuracy, offering a scalable and reproducible tool for PSP identification.

1- Segmentation masks highlighting four distinct brain regions: (A) Midbrain, (B) Pons, (C) Middle Cerebellar Peduncles, and (D) Superior Cerebellar Peduncles. Raw masks are shown in yellow, refined masks in other colours for comparison.

2- ROC curve analysis to compare indices’ performances in distinguish PSP from PD patients.

3- Confusion Matrices: Logistic Regression and Random Forest algorithms to distinguish PSP from PD patients