What question did this study set out to answer?

This study aims to evaluate the use of cerebellar radiomics features from MRI to differentiate patients with levodopa-induced dyskinesia from those without.

April 26, 2026Frontiers in Aging Neuroscience0 citationsOpen Access

Feasibility study of differentiating patients with levodopa-induced dyskinesia using cerebellar gray and white matter radiomics features from 3DT1WI images

YCYi ChenTaizhou Municipal Hospital YCYini ChenTaizhou Municipal Hospital ALAndong LinTaizhou Municipal Hospital

Key Points

This study aims to evaluate the use of cerebellar radiomics features from MRI to differentiate patients with levodopa-induced dyskinesia from those without.
Retrospective analysis of 3D T1WI data from the PPMI database, including 69 LID and 142 N-LID patients.
Data split into training and testing sets at an 8:2 ratio, with radiomic features extracted from four cerebellar ROIs.
Model developed using XGBoost algorithm, followed by performance evaluation and validation.
Model achieved an AUC of 0.962 on the training set and 0.849 on the testing set.
Thirteen radiomic features were identified as most representative for modeling.
The model effectively distinguishes between LID and N-LID patients, enhancing diagnostic performance.

Abstract

Background Levodopa therapy effectively treats Parkinson’s disease (PD) motor symptoms but causes Levodopa-Induced Dyskinesia (LID) in some patients long-term. Cerebellar changes exist in LID cases; however, radiomics models based on this region haven’t been evaluated for diagnostic use. We built diagnostic model using cerebellar structural radiomics from 3D T1WI to non-invasively diagnose LID. Methods In this study, we retrospectively collected 3D T1WI data from the Parkinson’s Progression Markers Initiative (PPMI) database, including data from 69 LID patients and 142 non-LID (N-LID) patients. These data were randomly split into a training set and a testing set at an 8:2 ratio. Using Fastsurfer segmentation, we identified four regions of interest (ROIs) corresponding to the left and right cerebellar gray matter and white matter. Python scripts were employed to independently extract radiomic features from each ROI. Subsequent steps involved feature selection and model construction. After selecting the optimal model, its performance was evaluated and validated. Finally, the SHAP method was used for model visualization. Results Ultimately, the most representative 13 radiomic features were used for modeling. The model built based on the XGBoost algorithm achieved an AUC value of 0.962 on the training set and 0.849 on the testing set. Conclusion The radiomic model extracted from the cerebellar gray and white matter effectively distinguishes between LID and N-LID patients. It offers a novel perspective on the heterogeneous characteristics of LID patients, significantly enhancing diagnostic performance and providing auxiliary support for clinical diagnosis.

KI fragen

Bookmark

View Full Paper