Objective: Our primary goal was to implement an automated measure of the midbrain to pons ratio (MPR) using an artificial intelligence method. Our second objective was to assess the diagnostic accuracy of the MPR for PSP against a range of parkinsonian disorders.

Background: There is a high degree of clinical overlap between atypical parkinsonian disorders, there is a high degree of clinical overlap, particularly in early disease. PSP is an
atypical parkinsonian syndrome causing rapidly progressive neurodegenerative disease with motor and cognitive impairment.
The MPR is a recognised and partially validated imaging biomarker for the disease. It has not been assessed in differentiating PSP from other atypical parkinsonian disorders.
Currently, MPR requires manual assessment by an expert familiar with neuroimaging. Automating it would facilitate large scale analysis for use as a diagnostic biomarker for clinical use and in trials.

Method: Our deep reinforcement learning method, communicative multi-agent reinforcement learning (C-MARL), trains agents to find landmarks. Agents explore 3D volumes and are rewarded based on their distance to the target landmark. They learn to communicate to increase their accuracy.
We used data from 370 participants (HCN=131, CBS=80, PSP=134, MSA=25 MSA) from the Cambridge Centre for Parkinsons Plus Disorders and the PROSPECT-M data split in training, testing and validation sets (70:15:15). Four landmarks from the midbrain and pons were manually recorded, using Massey et al.’s annotation method [1].
ROC analysis was carried out to assess the efficacy of the MPR in discriminating PSP from other APS.

Results: The algorithm was accurate in identifying landmarks (mean error 0.8mm, SD 0.3mm), resulting in an accurate measurement of MPR (mean percent error 3.7, SD 1.7).
The MPR was significantly smaller in PSP patients in comparison to all other groups [Figure 1], with a high accuracy for distinguishing between groups (AUC values: HC
0.98, MSA 0.97, CBD 0.90).

Conclusion: We report a new highly accurate automated measure of MPR predictor using artificial intelligence. We demonstrate that the MPR is a useful diagnostic biomarker for distinguishing PSP from other atypical parkinsonian disorders.

References: [1] Massey, L.A., Jäger, H.R., Paviour, D.C., O’Sullivan, S.S., Ling, H., Williams, D.R., Kallis, C., Holton, J., Revesz, T., Burn, D.J. and Yousry, T., 2013. The midbrain to pons ratio: a simple and specific MRI sign of progressive supranuclear palsy. Neurology, 80(20), pp.1856-1861.

« Back to MDS Virtual Congress 2021

MDS Abstracts - https://www.mdsabstracts.org/abstract/automated-midbrain-pons-ratio-using-artificial-intelligence-differentiates-psp-from-other-atypical-parkinsonian-disorders/