A Brainstem Radiomics Framework to Distinguish Progressive Supranuclear Palsy from Parkinson's Disease.

Area of Science:

• Neuroimaging
• Radiomics
• Differential Diagnosis

Background:

• Distinguishing progressive supranuclear palsy (PSP) from Parkinson's disease (PD) presents clinical challenges.
• Radiomics offers a promising approach to detect subtle image alterations for improved differential diagnoses.

Purpose of the Study:

• To evaluate the diagnostic utility of brainstem radiomic features from T1-weighted MRI in differentiating PSP from PD.
• To assess the performance of various classification models in this distinction.

Main Methods:

• Extracted radiomic features (first-order, shape, texture) from brainstem segmentations using PyRadiomics and AssemblyNet.
• Trained and validated classification models (Random Forest, XGBoost) on independent cohorts (433 participants total).
• Employed nested cross-validation and SHapley Additive exPlanations for model interpretability.

Main Results:

• Radiomics models demonstrated high performance in distinguishing PSP from PD, outperforming brainstem volume.
• Random Forest and XGBoost achieved AUCs of 0.93 and 0.94, respectively, in the validation cohort.
• Texture and intensity features were most predictive, while shape features had lower relevance.

Conclusions:

• Brainstem radiomics from T1-weighted MRI shows excellent and generalizable classification performance for PSP vs. PD.
• Texture-based radiomics captures microstructural changes missed by volumetry, highlighting its value in atypical parkinsonism diagnosis.
• Radiomics holds potential for integration into future multimodal diagnostic frameworks.