Abstract
OBJECTIVE: This study investigates deep convolutional neural networks (CNNs) for automated detection of early to mid-stage Parkinson's disease (PD) from static facial images, aiming to explore non-invasive, cost-effective approaches for early diagnosis and remote monitoring. METHODS: 2,000 facial images were collected from PD patients and healthy controls, followed by data augmentation to expand the dataset to 6,000 images. After randomly dividing the dataset into training and test sets according to 8:2, five CNN architectures were fine-tuned and assessed. Model performance was assessed by accuracy, precision, recall, specificity, F1 score, and area under the ROC and PR curve (AUC). Grad-CAM visualization techniques were applied to identify the discriminative facial regions associated with PD. RESULTS: ResNet18 achieved the best overall performance, yielding an F1 score of 99.67% across metrics. MobileNetV3 also performed robustly, particularly excelling in recall (99.00%), suggesting its suitability for high-sensitivity screening applications. EfficientNetV2 demonstrated stable convergence and competitive classification performance (F1 score: 96.30%), while VGG16 exhibited balanced performance with rapid convergence. Inception-v4 showed relatively lower accuracy and greater variability, indicating a potential risk of overfitting. Grad-CAM heatmaps revealed that the most predictive facial regions across models were concentrated around the eyes, lips, and nose, consistent with PD-related hypomimia. CONCLUSION: CNNs, particularly ResNet18 and MobileNetV3, exhibit significant potential for the automated identification of PD from facial imagery. These models offer promising avenues for developing scalable, non-invasive screening tools suitable for early detection and remote healthcare delivery, providing significant clinical and social value in the context of aging populations.