Poster R13 - How to Use Poor-Quality Retinal Images to Make Diagnoses of Diabetic Retinopathy: Improving Access to Care with Noisy Images

Introduction: Diabetic retinopathy is characterized by progressive damage to the blood vessels supporting the retina and early access to retinal imaging and appropriate treatment can quickly change an individual’s outcome by significantly reducing the risk of vision loss. According to the Center for Disease Control and Prevention, diabetic retinopathy is the leading cause of blindness in American adults. To make this diagnostic technology more accessible, portable retinal imaging devices have been developed, but have limitations in obtained image quality. This research seeks to develop the software component of an alternative approach: a smartphone-compatible retinal imaging attachment that doesn’t require dilation of the patient's eyes and uses post-processing of the resulting poor-quality image to make diagnoses.

Materials and

Methods: Two publicly available databases were utilized for this research project. The first dataset used is the APTOS 2019 Blindness Detection dataset publicly available by Kaggle and consists of 3662 training images graded on a scale of either no presence, mild, moderate, severe, or proliferate diabetic retinopathy. The second dataset used is the IEEE Multi-Label Retinal Disease (MuReD) dataset publicly available by IEEEDataPort and consists of 2208 images with 20 disease labels. Images were altered with varying amounts of Gaussian White Noise, a blurring-effect, and optical transfer function (OTF) noise to replicate the modulation transfer function. Altered images were evaluated separately and in combination through a modified convolutional neural network.

Results, Conclusions, and Discussions: Performance was evaluated using a receiver operating characteristic (ROC) curve. Performance is defined as the area under the ROC curve (AUC), as a value between 0.00 and 1.00. Overall, an AUC of 0.8910 was obtained for the 5-class determination of diabetic retinopathy. When observing the addition of Gaussian White Noise, a blurring-effect, and OTF noise separately, an AUC of 0.8652, 0.8694, and 0.8750 respectively. In combination, an AUC of 0.8139 was obtained. Through this, a clinician would have the ability to gain insight into a patients specific risk for vision-loss. Future research is needed to reconstruct the retinal image and further improve the accuracy.

Acknowledgements (Optional): I would like to acknowledge Dr. Murray Loew and Dr. Fadi Nasrallah for all of their support throughout this research project!