Interpretable hybrid deep learning model for Parkinson's disease screening using hand-drawn spiral and waveform images.

Abstract

Purpose: It is still hard to diagnose Parkinson's disease (PD) early and correctly, since the motor symptoms are often relatively mild and there are no unique diagnostic tools. The goal of this project was to create and test a hybrid deep learning model that can accurately classify PD by using hand-drawn spiral and waveform pictures.

Material and methods: A novel two-channel hybrid model was created, where the input representation combines normalised greyscale features and Canny edge features to capture both spatial and structural stroke patterns from patient drawings. The model combines a convolutional neural network (CNN) with hand-crafted grey-level co-occurrence matrix (GLCM) features to enhance its performance and make it easier to understand. We trained and evaluated three different models: a baseline CNN, a fusion CNN + GLCM, and a fine-tuned ResNet-50. We did this on both the original and pre-processed datasets.

Results: The hybrid CNN + GLCM model that used pre-processing had the best classification accuracy at 97.02% and worked well on datasets that were not used to train it. Statistical studies validated the importance of enhancements in performance relative to baseline models.

Conclusions: The suggested technique provides a straightforward, comprehensible, and efficient approach for PD screening using easily administered drawing exercises. Its great precision and low equipment needs make it a good candidate for use in real-world clinical settings.