A Cerebellar Radiomics Model Based on Automatic Segmentation and SHAP-Interpretable Machine Learning for Parkinson's Disease Discrimination
A Cerebellar Radiomics Model Based on Automatic Segmentation and SHAP-Interpretable Machine Learning for Parkinson's Disease Discrimination
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Publication status: preprint
A plain-language reading has not been prepared for this paper yet.
Original abstract
Abstract Background Parkinson's disease (PD) is among the most common progressive neurodegenerative disorders, and early and accurate identification is vital for effective subsequent treatment interventions. Recent pathological and neurophysiological research has clearly demonstrated the cerebellum's significant role in PD's pathological progression. This study utilizes automatic segmentation techniques to investigate the potential of radiomics methods, based on cerebellar MRI images, for distinguishing between PD patients and healthy control (HC) individuals. Methods This study retrospectively gathered and analyzed 3DT1WI images from 377 subjects in the PPMI database. Fastsurfer segmentation was employed to identify four cerebellar regions of interest (ROIs): left and right gray matter, and left and right white matter. From each ROI, 833 radiomic features were independently extracted. Subjects were randomly assigned to training and testing sets in a 7:3 ratio. Following preprocessing and feature selection, various machine learning algorithms were used for modeling and evaluation, with the best-performing model selected for SHAP visualization. Results Nine radiomic features were ultimately chosen for modeling. The best model, developed using the logistic regression (LR) algorithm, achieved an area under the curve (AUC) of 0.774 in the training set and 0.714 in the testing set. SHAP visualization indicated that the feature 47_wavelet_LHH_glcm_Imc2 contributed most significantly to differentiating PD. Conclusion This study, based on 3DT1WI imaging data, extracted radiomic features from the left and right gray and white matter of the cerebellum and successfully constructed a radiomic model to differentiate between PD patients and HC individuals. The model effectively captures the heterogeneous characteristics of cerebellar gray and white matter, highlighting potential pathological changes in PD patients and offering valuable support for PD diagnosis.