Enhancing Optical Coherence Tomography Angiography Feature analysis for Detection of Parkinson's Disease by Designing a Computer-Aided Diagnosis System: A Machine Learning-Based Approach
Mohammad Reza Hasanshahi1 *, Dr. Alireza Mehdizade Tzangi10 , Dr. Tahere Mahmoudi10 , Dr. Vahidreza Ostovan10 , Dr. Mohammad Hossein Norozzade10 , Dr. Hossein Parsaee10
- Shiraz University of Medical Sciences - Medical Physics Department
Abstract: Neurological disorders are a major cause of global disability, with Parkinson's disease (PD) being the fastest growing, affecting 6.2 million people in 2015. PD is caused by the early loss of dopamine-producing neurons, leading to motor and non-motor symptoms that can appear years before diagnosis. Since there are no definitive diagnostic tests for PD, identifying early, non-invasive biomarkers is crucial. Optical Coherence Tomography Angiography (OCTA), a retinal imaging technique, shows promise as it reflects changes in the central nervous system. Although OCTA has revealed retinal microvascular changes in PD, its diagnostic potential is still being investigated.
This study used machine learning to analyze OCTA images, aiming to classify PD patients from normal population.
Methods: Data from 92 participants (52 with PD and 40 controls) and 175 eye images were examined. Most PD participants (85%) were in the early stages of the disease (Hoehn and Yahr Stages 1 and 2), making them ideal for studying early non-motor features, including ocular changes. After image pre-processing to enhance the quality of the images, region growing and thresholding algorithms were used to segment the FAZ area and vessels, respectively. Then, related features such as vessel density and FAZ parameters were extracted to learn the Machine Learning (ML) models. Several ML models were trained to distinguish PD patients from healthy controls, with XGBoost (XGB) and Gradient Boosting performing best.
Results: XGB achieved 74% accuracy and 85% sensitivity in independent test dataset. Ensemble methods improved performance further, achieving 80% accuracy and 100% sensitivity, effectively detecting nearly all PD cases. Receiver Operating Characteristic (ROC) analysis confirmed the reliability of XGB and ensemble models, highlighting their potential as diagnostic tools for early PD detection.
Conclusion: This study demonstrates how OCTA and advanced machine learning can provide valuable biomarkers, paving the way for improved diagnosis and management of Parkinson’s disease.
