Purpose

To quantify the performance of a new Deep Learning (DL)-based algorithm in differentiating Progressive Supranuclear Palsy (PSP) patients from healthy individuals and patients with Parkinson’s disease (PD).

Materials & Methods

Quantitative brain volumetry was carried out with our new DL-based algorithm. It was tested on 248 patients that consist of 221 healthy patients (age 61.3y±10.1y, source: Parkinson’s Progression Markers Initiative, PPMI), 46 patients with confirmed PD (age 64.2y±8.7y, source: PPMI) and 11 patients with confirmed PSP (age 68.8y±8.4y, source: internal). Images were acquired on 1.5T/3.0T MR-scanner using 3D-T1w images. Brain volumetry and quantitative comparison against a normal model was performed for: whole brain, grey&white matter, frontal, parietal, occipital, temporal lobe, hippocampus, mesencephalon, pons and all ventricles. Furthermore, the midbrain-to-pons-ratio (MtPR) was calculated in 3D. Performance was tested with respect to the algorithm’s ability to correctly identify (i) PSP against healthy patients and (ii) PSP against those with confirmed PD based on percentiles. For that, ROC was used to calculate the corresponding AUC and sensitivity/specificity on the best performing regions.

Results

In the case of (i), calculations yielded highest AUC for: mesencephalon (0.944+/-0.049) and 3D MtPR (0.896+/-0.129) while for (ii), calculations yielded highest AUC for: mesencephalon (0.915+/-0.084) and the occipital lobe (0.893+/-0.085). Resulting sensitivity/specificity calculations on the best performing region yielded values of 0.91/0.89 for (i) and of 0.91/0.87 for (ii).