Dopamine Transporter Imaging‐Based Radiomics for Predicting 4‐Year Motor Progression (Unified Parkinson's Disease Rating Scale Part Ⅲ) in Parkinson's Disease

Parkinson's disease (PD) shows substantial heterogeneity in motor progression, and reliable early prediction remains challenging. Dopamine transporter (DAT) imaging reflects presynaptic dopaminergic degeneration but is typically analyzed using simplified metrics that may not capture spatial heterogeneity. This study aims to evaluate whether magnetic resonance imaging (MRI)‐guided DAT radiomics combined with machine learning can predict 4‐year motor progression in PD.

In this retrospective study, 120 patients from the Parkinson's Progression Markers Initiative were included. Motor progression was defined using the 4‐year change in Movement Disorder Society–Unified Parkinson's Disease Rating Scale part Ⅲ (UPDRS‐Ⅲ) score. Bilateral caudate nuclei and putamina were manually delineated on T1‐weighted MRI and transferred to co‐registered DAT images. Radiomics features were extracted using PyRadiomics. Twenty base models were constructed by combining two feature‐selection methods with 10 classifiers, and Top‐2 voting fusion models were developed. Clinical‐variable, levodopa equivalent daily dose (LEDD)‐adjusted, and clinical‐radiomics models were also evaluated.

The best‐performing base model achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.736. Fusion models improved discrimination, with the best soft‐voting model achieving an AUROC of 0.920 and balanced accuracy (BACC) of 0.734. A more balanced fusion model achieved an AUROC of 0.902 with higher sensitivity (0.818). The clinical‐radiomics model showed only a slight AUROC increase over the clinical‐only model, and adding Year‐4 LEDD did not materially alter radiomics performance. Stable radiomics features were predominantly texture‐based, particularly gray‐level co‐occurrence matrix (GLCM) and gray‐level run‐length matrix (GLRLM) features derived from original and wavelet‐filtered images.

MRI‐guided DAT radiomics combined with ensemble learning may provide prognostically relevant information for predicting 4‐year motor progression in PD. Fusion strategies improved discrimination, and stable texture features may reflect striatal dopaminergic heterogeneity.