Towards Automated Optical Coherence Tomography Based Classification of Diabetic Retinal Tissue

Abstract

 

Purpose:

 

To develop and validate an automated method for prediction and classification of early diabetic retinopathy (DR) using optical coherence tomography (OCT) images and an artificial neural network (ANN).

 

Methods:

 

Structural and optical properties measurements from a total of 930 OCT images were obtained from a total of 155 eligible eyes from 99 participants (see table 1). Intraretinal layers (RNFL, GCL+IPL, INL, OPL, ONL+IS, OS and RPE) were segmented automatically by using a custom-built algorithm (OCTRIMA). Thickness, fractality and total reflectance were calculated for each intraretinal layer. Structure-optical properties relationships were mapped employing an ANN characterized with Bayesian radial basis function (BRBF) using four data sets for training (see table 1). The ANN mapping was performed for layers that showed a maximum diagnostic power:GCL+IPL complex and OPL. This model was tested by using four data sets to predict and classify MDR eyes (see table 1). The ANN used different pairs of input and target features based on structural and optical properties measurements. True-positive, true-negative, false-positive and false-negative results were obtained, allowing calculation of sensitivity, specificity and accuracy.

 

Results:

 

Our results showed that prediction and classification of early DR in diabetic subjects may be possible by using OCT images and an ANN. When the total reflectance and fractality were used as input and target pairs in the BRBF’s ANN (training sets: 20, 30, 40 healthy eyes), 37 MDR eyes with OPL differences as well as 35 MDR eyes with GCL+IPL differences (out of 43 MDR testing eyes) were effectively differentiated. When 20 MDR eyes (out of 43) were selected to train the BRBF’s ANN, 18 MDR eyes with GCL+IPL differences (out of 23 MDR testing eyes used in the 4th testing data set) were classified. Results of sensitivity, specificity and accuracy are shown in table 1.

 

Conclusions:

 

Our results suggest that the BRBF’s ANN could be an effective clinical tool to classify early DR in diabetic subjects based on GCL+IPL and OPL differences on OCT images. Further research using a larger data set is warranted to determine how this approach may be used to improve diagnosis and treatment of diabetes.  

 

Keywords: diabetic retinopathy • discrimination • optical properties