Automatic Differentiation of Parkinson's Disease Motor Subtypes Based on Deep Learning and Radiomics

Dongming Hui^1,2Xia Wang²Lu Xie³Fengxi Chen⁴Guo Yi^5*YaXi Luo^6*Xiao-Jing He^1*

• ¹Department of Radiology, The Second Affiliated Hospital of Chongqing Medical University, Chongqing, China
• ²Department of Radiology, Chongqing Western Hospital, Chongqing, China
• ³Medical technology, Community Health Center of Yubei in Sha Ping Ba, Chongqing, China
• ⁴7T Magnetic Resonance Translational Medicine Research Center, Department of Radiology, Southwest Hospital, Army Medical University (Third Military Medical University), Chongqing, China
• ⁵Medical Imaging Department, The Fourth People's Hospital of Chongqing, Chongqing, China
• ⁶Department of Radiology, Renji Hospital,School of Medicine, Chongqing University （the Fifth People’s Hospital of Chongqing）, Chongqing, China

Objective: Parkinson's disease (PD) is a common neurodegenerative disorder, and the early and accurate differentiation of its motor subtypes is of significant importance for clinical diagnosis and treatment planning. Research has shown that deep brain nuclei such as the thalamus, caudate nucleus, putamen, and globus pallidus play a critical role in the pathogenesis of different motor subtypes of Parkinson's disease. This study aims to utilize deep learning and radiomics technology to establish an automated method for differentiating motor subtypes of Parkinson's disease. Methods: The data for this study were obtained from the Parkinson's Progression Markers Initiative (PPMI) database, including a total of 135 Parkinson's disease patients, comprising 43 cases of the Postural Instability/Gait Difficulty (PIGD) subtype and 92 cases of the Tremor Dominant (TD) subtype. High-resolution MRI scans were used to extract 2,264 radiomics features from 8 deep brain nuclei, including bilateral thalamus, caudate nucleus, putamen, and globus pallidus. After dimensionality reduction, five independent machine learning classifiers (AdaBoost, Bagging Decision Tree [BDT], Gaussian Process, Logistic Regression, and Random Forest) were trained on the training set and validated on the test set. Model performance was evaluated using the Area Under the Curve (AUC) metric. Results: After feature selection, 17 most discriminative radiomics features were retained. Among the models, the BDT-based diagnostic model demonstrated the best performance, achieving AUC values of 1.000 and 0.962 on the training and test sets, respectively. DeLong's test results indicated that the BDT model significantly outperformed other models. Calibration curve analysis showed that the Parkinson's disease subtype classification model based on MRI radiomics features exhibited good calibration and stability. Clinical decision curve analysis revealed that the BDT model demonstrated significant clinical net benefits across a wide probability range, indicating high clinical utility. Conclusion: The BDT model based on MRI radiomics features constructed in this study exhibited excellent performance in differentiating motor subtypes of Parkinson's disease. This fully automated model is capable of processing MRI data and providing results within 3 minutes, offering an efficient and reliable solution for the early differentiation of Parkinson's disease motor subtypes, with significant clinical application value.