Automated Decision Support System for Pathology of Diabetic Retinopathy from OCT

Diabetic retinopathy is a complication of diabetes causing progressive damage to the retina, located at the back of the eye, potentially leading to clouded vision or blindness. Disease signs may be visualized by Optical Coherence Tomography (OCT) and include formation of new and weaker blood vessels, fluid accumulation, exudates and changes to Retinal Vascular Geometry (RVG). Presence of these indicators can provide information as to the stage of the disease. Image-processing strategies are applied for the automated detection, segmentation, extraction, classification toward likelihood estimation of progression of diabetic retinopathy to visual biomarkers present in OCT, using time-sequenced data in the early stages of the disease. Gabor and Savitsky-Golay filtering enables extraction of the vessel map and fuzzy control for segmentation of hard exudates. Feature data are extracted using bounding boxes, vector map and connected component methodology for binary decision tree classifier construction, training and testing. Feature values comprising classifier nodes include: exudate features of compactness, area, convexity and form factor, in addition to vessel features: width, elongation, bifurcation angles, form factor and solidity. Classifier accuracy is 93.3%, with 6.7% misclassification and 0% false-negative classification. Automated image processing of diabetic retinopathy is achieved with high classification accuracy for the extraction of vessel map and hard exudate biomarkers from OCT. Application of smoothing algorithms and removal of vessel map shadows may further improve classification accuracy.