An advanced deep learning model for iridology based disease diagnosis using Pyramid network driven iris segmentation

Diabetes is a condition that commonly affects human beings. Diabetes can be diagnosed using a variety of approaches, one of which is blood pressure monitoring. However, this method is inefficient because it requires a blood sample and might be time-consuming. Iridology is a health assessment technique that examines the iris of the eye. As a result, many diabetes cases remain undiagnosed in their early stages. This paper presents a diabetes prediction system through the analysis of iris images based on machine learning techniques. Initially, iris images were collected from the diabetes iridology database and IIT Delhi iris dataset. Next, pre-processing is performed by Gaussian Amended Wiener Filter (GAWF) to reduce noise and enhance contrast in the images. Then, an improved Pyramid Scene Parsing Network (PSPNet) model is employed to segment the Inner Iris Zone from the iris images. After that, Depthwise Separable Convolutional Dense Net (DSC-DenseNet) is utilized to extract relevant features from these segmented images. Finally, extreme gradient boosting (XGBoost) with Bayesian optimization (BO) is used to classify diabetes disease. The proposed model attains an accuracy of 98.42%, precision of 97.82%, recall of 97.98%, and F1-score of 98.23%. In addition, the IIT Delhi iris dataset achieves 98.25% accuracy, 97.32% precision, 97.58% recall, and an F1-score of 98.02%. The proposed approach correctly classifies diabetes from these input samples and enhances the accuracy of the model.