Deep Learning to Identify Parkinson’s Disease Biomarkers in Age-related Macular Degeneration Fundus Image

Parkinson's disease is a psychological disorder that leads to neurodegenerative effects on the normal physical movement of the patients. The cure or reversal of the symptoms of any Parkinson's patient can only be achieved only through early detection. This research explores a novel approach for Parkinson's detection from the Age-related Macular Degeneration (AMD) fundus images of the Parlinson's affected patients since researchers suggested that there may be shared genetic predispositions and environmental risk factors (such as smoking and diet) that could contribute to the development of both conditions. The presence of one condition may sometimes mask or complicate the diagnosis of the other, especially in early stages. For example, visual problems in a patient with PD might initially be attributed solely to normal aging or PD-related issues, potentially delaying the diagnosis of AMD. The features extracted from fundus images and training classification models like Random Forest, Gradient Boosting model, AdaBoost model, XGBoost model, Multi-Layer Perceptron [MLP] model, Support Vector Classifier [SVC], Decision Tree, K-Nearest Neighbors [KNN] classifier along with Stacked, Ensemble and Bagging Classifier using combination of the mentioned models to evaluate the feasibility of the proposed method. Fundus images primarily capture the blood vessels at the back of the eye and can potentially reveal patterns of neurological conditions. Various image pre-processing algorithms like green channel extraction, Contrast Limited Adaptive Histogram Equalization [CLAHE], Adaptative Thresholding followed by Morphological Operations like TopHat filtering, morphological opening along with Sobel and Canny Edge Detector, Hough Line Detector, Graycomatrix method for Texture Properties, are used to extract the Region features from fundus images. The best features for training the classification models are found using methods such as Random Forest importance score modelling. Gradient Boosting frequency modeling, Recursive Feature Elimination with Cross-Validation (RFECV) technique and Univariate Feature Selection technique. The performance of the proposed approach is evaluated using metrics such as Accuracy, Precision Score, Recall Score, F1-Score, Sensitivity, Specificity, Positive Prediction Accuracy [PPA], Negative Prediction Accuracy [NPA], Area under the Receiver operating characteristic [ROC] Curve {AUC-ROC}. The best performing model is the Stacked Classifier with accuracy of 97.17 %, Fl score of 0.9716, Sensitivity of 0.9529, Specificity of 0.9824, AUCROC of 0.9937, PPA of 0.9687, NPA of 0.9733, Precision of 0.9687 and recall of 0.9529. The utilization of morphological features from fundus images for Parkinson's detection offers a non-intrusive, cost-effective solution, ultimately improving the quality of diagnosis of Parkinson's disease.