# Automated MRI‐Based Classification of Parkinsonism: A Deep Learning Approach to Distinguish PD From PSP

## TL;DR

This study uses deep learning to automatically calculate MRI-based biomarkers to distinguish Parkinson's disease from progressive supranuclear palsy, improving diagnostic accuracy.

## Contribution

An automated deep learning method for MRPI calculation from thick-slice MRI, achieving high diagnostic accuracy in differentiating PD and PSP.

## Key findings

- MRPI 2.0 showed higher diagnostic accuracy than MRPI 1.0 with an AUC of 0.78.
- The automated method achieved an average AUC of 0.85 in distinguishing PSP from PD.
- The method demonstrated strong linear correlation with manual assessments, validating its reliability.

## Abstract

Differentiating Parkinson's disease (PD) from progressive supranuclear palsy (PSP) is crucial for appropriate treatment, as each disease has distinct therapeutic requirements. The Magnetic Resonance Parkinsonism Index (MRPI) has shown promise as a diagnostic biomarker, yet manual methods introduce variability and limit its applicability. In this study, we aim to develop a fully automated algorithm for MRPI 1.0 and 2.0 calculation, and assess its ability to distinguish PD from PSP in two cohorts from different regions of China.

A total of 75 PD patients and 29 PSP patients from two hospitals were enrolled. All participants underwent neurological examinations, including the MDS‐UPDRS‐III and H‐Y scale, as well as brain MRI scans. Additionally, tissue‐intensity images derived from 3D isotropic T1WI images from 2D thick slices using a deep learning (DL)‐based super‐resolution (SR) technique were aligned to a standard template followed by corresponding structural mask parcellation for measurement of MRPI 1.0 and MRPI 2.0. Subsequently, a logistic regression model was constructed to identify PD patients from PSP based on these indexes.

MRPI 2.0 demonstrated higher diagnostic accuracy than MRPI 1.0, with an AUC of 0.78. Additionally, the automated method showed strong linear correlations with manual assessments from an experienced radiologist, validating its reliability, and identification of PSP from PD with the average AUC of 0.85.

The automated MRPI method improves diagnostic accuracy for differentiating PD from PSP, providing a reliable and clinically applicable tool. The integration of a super‐resolution technique to convert 2D MRI data into high‐resolution images expands the potential of MRPI as a neuroimaging biomarker.

Using a deep learning‐based super‐resolution technique, we developed an automated method for calculating MRPI 1.0 and 2.0 from thick‐slice MRI, achieving an AUC of 0.85 in distinguishing PSP from PD. This approach enhances diagnostic accuracy, offering a reliable tool for the clinical differentiation of these parkinsonian disorders.