Abstract
While (18)F-florzolotau tau PET is an emerging biomarker for progressive supranuclear palsy (PSP), its interpretation has been hindered by a lack of consensus on visual reading and potential biases in conventional semi-quantitative analysis. As clinical manifestations and regions of elevated (18)F-florzolotau binding are highly overlapping in PSP and the Parkinsonian type of multiple system atrophy (MSA-P), developing a reliable discriminative classifier for (18)F-florzolotau PET is urgently needed. Herein, we developed a normalization-free deep-learning (NFDL) model for (18)F-florzolotau PET, which achieved significantly higher accuracy for both PSP and MSA-P compared to semi-quantitative classifiers. Regions driving the NFDL classifier's decision were consistent with disease-specific topographies. NFDL-guided radiomic features correlated with clinical severity of PSP. This suggests that the NFDL model has the potential for early and accurate differentiation of atypical parkinsonism and that it can be applied in various scenarios due to not requiring subjective interpretation, MR-dependent, and reference-based preprocessing.