Objective: Compare the performance of deep learning and machine learning models trained on quantitative diffusion measurements in the prediction of visual dysfunction in Parkinson’s Disease cohort.

Background: Parkinson’s Disease (PD) is a primary progressive neurodegenerative disorder characterized by debilitating motor and non-motor symptoms such as visual dysfunction.¹

Diffusion tensor imaging (DTI) is a post-processing neuroimaging modality derived from DWI.

Method: This study includes 43 patients with PD clinically evaluated for visual dysfunction. Quantitative measures at 37 brain ROIs were obtained from 7 DWI maps. Deep learning (DL) models included a 5 multi-layer perceptron (MLP) and 1-dimension convolutional neural network (1D-CNN). DL model performance was evaluated using train, validation, and test set accuracy (%) and AUC-ROC score. Comparison to classical ML models with hyperparameter tuning and cross validation were performed using 5 K-folds.

Results: DL models were trained for 50 epochs in which the 1D-CNN achieved a higher average train and comparable best validation accuracy compared to the MLP model (96.7% vs 60% train; 83.0% vs 83.0% validation). Test set evaluation of deep learning models revealed 1D-CNN outperformed MLP on the test dataset in accuracy (80% vs 43%) and AUC-ROC (0.72 vs 0.42).

Conclusion: We demonstrate the ability of DL models to classify visual dysfunction in PD using quantitative measures derived from DWI imaging. Due to the relationship between visual dysfunction and poor prognostic outcomes in PD, our study indicates machine learning based tools may aid in early disease detection and clinical management.

« Back to 2024 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/deep-learning-based-prediction-of-visual-dysfunction-in-parkinsons-disease-from-31-brain-regions-of-interest-using-quantitative-diffusion-mri/