Using automated FreeSurfer subnuclear segmentation and XGBoost on two independent cohorts, the study found marked atrophy in lateral, paraventricular, and pulvinar thalamic nuclei in PSP and achieved AUC 0.96 for distinguishing PSP from PD based on thalamic nuclei volumes.

Identifies a robust, translatable MRI-based subregional thalamic volumetric biomarker that can improve differential diagnosis and trial cohort selection for PSP (and thereby indirectly aid therapeutic development), though it offers limited direct mechanistic or targetable insights for Parkinson's…

BACKGROUND: Several MRI studies have documented thalamic atrophy in Progressive Supranuclear Palsy (PSP), but investigations to date have considered the whole thalami, without evaluating their subcomponents. This study aimed to assess atrophy of thalamic substructures in PSP and investigate their potential to support PSP differential diagnosis. METHODS: A total of 398 subjects were enrolled in the study, including 164 PSP patients, 180 Parkinson's disease (PD) patients and 64 healthy controls (HC), from two cohorts. An automatic probabilistic segmentation of thalamic nuclei was employed on T1-weighted MRI images using FreeSurfer 7.4 to investigate thalamic subregional atrophy. Subsequently, machine learning analysis (XGBoost) was employed to differentiate 97 PSP from 98 PD patients, and the results were validated in an independent international cohort (67 PSP, 82 PD patients). RESULTS: PSP patients showed widespread thalamic atrophy, most prominent in the lateral, paraventricular, and pulvinar nuclei in comparison with HC. Conversely, no significant differences were observed between PD patients and HC. The machine learning model based on the thalamic nuclei volumes achieved excellent performance in distinguishing PSP from PD, and the results were validated in the independent international cohort (AUC: 0.96 in both cohorts), outperforming the volume of the whole thalamus (De Long test, p < 0.05). CONCLUSION: This study investigated subregional thalamic atrophy in PSP patients and demonstrated that a machine learning based on thalamic nuclei volumes was able to accurately distinguish PSP from PD in two independent cohorts, highlighting the potential of subregional thalamic atrophy as diagnostic biomarker for PSP.