Deep learning for Parkinson's disease classification using multimodal and multi-sequences PET/MR images

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Parkinson's disease (PD) and multiple system atrophy (MSA) exhibit overlapping clinical symptoms, complicating differential diagnosis.
• Deep learning (DL) offers a promising approach for distinguishing between PD and MSA.

Purpose of the Study:

• To develop and evaluate a DL model for accurate classification of PD, MSA, and normal controls (NC).
• To assess the performance of multi-modal imaging in differentiating PD from MSA.

Main Methods:

• Retrospective analysis of 206 patients (PD/MSA) and 38 NC using PET/MR imaging.
• Training a modified 18-layer Residual Block Network (ResNet18) on 2D multi-modal image slices.
• Utilizing a four-fold cross-validation and evaluating performance with accuracy, precision, recall, F1 score, ROC, and AUC.

Main Results:

• Multi-modal models, particularly those combining PET and MRI data, outperformed single-modal approaches.
• The ¹¹C-CFT and Apparent Diffusion Coefficient (ADC) multi-modal model demonstrated superior classification performance.
• The best-performing model achieved 0.97 accuracy, 0.93 precision, 0.95 recall, 0.92 F1, and 0.96 AUC in the training set.

Conclusions:

• The developed DL method shows significant potential as an assistive tool for accurate PD and MSA diagnosis.
• Multi-modal and multi-sequence DL models can enhance the classification accuracy for PD and related disorders.