Machine learning algorithm for retinal image analysis

Abstract

Diabetic retinopathy is the most general diabetes complication that affects eyes and results in blindness. It's due to impairment of the arteries a veins located in the fundus of eye (retina) that are composed of light sensitive tissues. The aim of this research work is to design an efficient and sensitive tool for Diabetic Retinopathy using the images acquired from portable fundus camera. The screening tool is based on advanced machine learning and computer vision algorithm which includes patch level prediction. In patch level prediction algorithm will localize the diseased region in the Diabetic Retinopathy image like Hard Exudates and Hemorrhage. The patch level classification uses Support Vector Machine (SVM) machine learning classifier model to predict the potential patch of Hard Exudates and Hemorrhage. In this algorithm, the image is broken into regular rectangular patch. The feature for each patch along with the different class label based on the ground truth is computed and passed to strong classifier SVM. The data sets are split into training dataset and testing dataset. The classifier model is built on training dataset and tested against the test dataset. The performance results of rectangular patch level prediction using SVM the average performance for Hard Exudates was Accuracy 96 %, Sensitivity 94%, Specificity 96%. The average performance for Hemorrhage was Accuracy 85 %, Sensitivity 77%, and Specificity 85%.