An integrated framework for automatic clinical assessment of diabetic retinopathy grade using spectral domain OCT images

Abstract

Diabetic retinopathy (DR) is a progressive disease and its detection at an early stage is crucial for saving a patient's vision. In this paper, an enhanced computer-assisting diagnostic (CAD) system is developed for the discovery and grading of non-proliferative DR from optical coherence tomography (OCT) images. The proposed CAD system elaborates three sequential stages. Initially, 12 distinct retina layers are localized using our previously developed segmentation approach based on an integrated joint model that combines shape, intensity, and spatial information. Secondly, three features, namely the reflectivity, curvature, and thickness are quantitatively measured from the segmented layers. Finally, a two-stage deep fusion classification network (DFCN), trained by stacked non-negativity constraint autoencoder (SNCAE), is used first to classify the subject as normal or DR, then assess the grade of DR as either early stage or mild/moderate. Using "leave-one-subject-out" experiments on a dataset of 74 OCT images, the CAD system distinguished between normal and DR subjects with a 93% accuracy (sensitivity =91%, specificity =97%) and achieved a 98% correct classification between early stage and mild/moderate DR. These results confirm the proposed framework as a reliable non-invasive diagnostic tool.