Diabetic Retinopathy Diagnostic CAD System Using 3D-Oct Higher Order Spatial Appearance Model

Abstract

Diagnoses of Diabetic Retinopathy (DR) at an early stage are of extreme importance so that the retina can be preserved and the risk of substantial damage to the retina or loss of vision is reduced. A new Computer-Aided Diagnosis (CAD) method based on Optical Coherence Tomography (OCT) scans of the retina is presented here for the detection of DR at an early stage. Utilizing an adaptive appearance-based approach that uses prior shape information, the system segments the retinal layers from the 3D-OCT scans. From the layers segmented from the B-scans volume of the OCT, novel texture features are extracted for DR diagnosis. In particular, a 2nd-order reflectivity value is calculated for each individual layer using the 2D Markov-Gibbs Random Field (2D-MGRF) model. Then, Cumulative Distribution Function (CDF) descriptors are used to represent the extracted image-derived feature using CDF’s percentiles. A feed-forward neural network is used for layer-by-layer classification of 3D volume using Gibbs energy features extracted from each individual layer. In the final stage, all twelve layers are fused with a global subject diagnosis based on a majority voting method. We evaluated a 3D-OCT system using 180 subjects using a combination of different k-fold validation techniques. The system performance for this CAD system using 4-, 5-, and 10-fold cross validation achieved accuracies of 89.4%, 91.5%, and 95.7%, respectively. In addition, our system’s ability to detect the DR early has been validated by further comparisons with the state-of-the-art deep learning networks.