Deep learning approach for automated detection of retinal pathology from ultra-widefield retinal images

Abstract

Purpose : This study aims to develop a deep learning approach that can automatically detect retinopathy in ultra-widefield (UWF) retinal images with limited training data. It can serve as the steppingstone for large scale deployment of real-time automated retinopathy detection tools using UWF images where no similar approach has been developed.

Methods : The dataset used to train and validate the approach includes 324 3-channel images obtained using the ultra-widefield fundus camera (UWF Primary, Optos) on 162 diabetic patients at Duke Endocrinology. An image preprocessing method that did affine transformation and added noise to raw images was developed to increase data amount. A Residual Neural Network (ResNet) was trained to detect retinal pathology. Specifically, 75% of ultra-widefield retinal images were used to train the ResNet, and the rest 25% of the images were used for the test. Figure 1 shows an overview of our approach.

Results : Our approach achieved 87.97% (95% CI [85.88%, 90.05%]) accuracy in detecting referrable retinopathy, along with the False Negative Rate of 13.23% (95% CI [11.06%, 15.41%]), Recall of 86.77% (95% CI [84.59%, 88.94%]), Precision of 86.82%(95% CI [84.65%, 88.99%]), F1 Score of 86.79% (95% CI [84.62, 88.96]), and AUC-ROC of 86.77% (95% CI [84.59%, 88.94%]). The confusion matrix is shown in Figure 2.

Conclusions : Our model, trained with limited data, can successfully detect retinopathy in ultra-widefield retinal images. It could be used as a useful real-time automated tool in clinical practice. Future studies will improve detection performance by collecting more UWF images and combining demographics information as model input.

This is a 2021 ARVO Annual Meeting abstract.

 

Figure 1. Overview of our proposed approach

 

Figure 2. Confusion matrix of the test result. 0 represents normal, and 1 represents retinopathy.