Design of deep polynomial network-based magnetic resonance imaging-aided diagnostic model for assessing Parkinson's disease in patients

Parkinson's disease (PD) is a common neurodegenerative disorder characterized primarily by the degeneration and loss of dopaminergic neurons in the substantia nigra (SN). To enhance the accuracy of early diagnosis and the objectivity of disease progression assessment in PD, this study proposed a magnetic resonance imaging (MRI) image-aided diagnostic model based on a deep polynomial network (DPN) to investigate its potential as a cost-effective alternative for evaluating the clinical condition of PD patients. This study analyzed retrospective MRI data from 146 participants (86 PD patients, 60 healthy controls) collected between 2021 and 2024. All underwent standard head MRI (T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), fluid-attenuated inversion recovery (FLAIR)). Features were extracted from key regions (SN, globus pallidus, striatum), underwent deep polynomial encoding, and were concatenated for higher-order representations. A multi-modal DPN (MDPN) model was built and evaluated using 5-fold cross-validation. Its performance in PD classification and Hoehn-Yahr stage regression was compared to support vector machine (SVM), ResNet18, and Transformer models. In the 5-fold cross-validation, the proposed MDPN indicated superior performance in PD recognition on the validation sets, achieving an average accuracy of 93.1 % and an area under the curve (AUC) of 0.962, considerably outperforming the comparative models: SVM (accuracy 85.3 %, AUC 0.893), ResNet18 (accuracy 88.5 %, AUC 0.921), and Transformer (accuracy 90.2 %, AUC 0.936). In the Hoehn-Yahr staging prediction regression task, the MDPN model attained the lowest mean squared error ((MSE) = 0.217) and mean absolute error (MAE = 0.371), demonstrating high fitting accuracy across different stages (R 2 = 0.87). Despite the limitations of a single-center retrospective design, the MDPN model has high accuracy and robustness in the auxiliary diagnosis and condition assessment of PD. It outperforms traditional methods in classification and regression tasks and has a promising clinical application prospect. Furthermore, the model identified a high-order interaction between SN texture and striatal volume, aligning with the known nigrostriatal degenerative pathway in PD.