Early Parkinson's Disease Diagnosis using Hand-Drawn Image Patterns using Lightweight and Deep Learning Models with Explainable AI

Parkinson's disease (PD) is a neurological disorder characterized by subtle early-stage motor impairments, drawing considerable attention from the medical community. This research study presents a benchmarking between lightweight and deep learning convolutional neural network (CNN) models for Parkinson’s Disease prediction using the curated handwritten image dataset. It is chosen to assess six models across two categories: lightweight architectures (MobileNetV2, ShuffleNetV2, GhostNet) and deep learning models (InceptionV3, EfficientNetB3, DenseNet121) and analyse their potential on PD diagnosis. These models were subjected to rigorous evaluation using the comprehensive framework metrics, which included classification metrics, computational efficiency metrics (FLOPs (Floating-Point Operations per Second), parameters, inference time), robustness to adversarial perturbations (Gaussian noise and FGSM attacks), and total training duration. The results revealed that DenseNet121 excelled in accuracy (97%) and robustness, while GhostNet, a lightweight model, struck an impressive balance with 96% accuracy, an outstanding F1-Score (0.98), and high noise resilience. This makes GhostNet an ideal choice for environments with limited resources. EfficientNetB3 demonstrated competitive performance (91% accuracy) with rapid training (5.2 minutes) and strong robustness. Additionally, MobileNetV2 and ShuffleNetV2 showcased swift inference and reduced computational overhead, rendering them highly suitable for real-time applications. Our findings underscore the effectiveness of lightweight models enhanced with explainable AI (Grad-CAM) for interpretable and resource-efficient Parkinson's disease detection. This study highlights that judicious model selection, paired with balanced preprocessing strategies, facilitates scalable, transparent, and accurate diagnosis of Parkinson's disease using simple hand-drawn images.