External Eye Photos for Diabetic Retinopathy Detection with Deep Learning

Abstract

Purpose : To develop a deep learning model to automatically detect vision-threatening diabetic retinopathy using external eye photos.

Methods : Since external eye photos can be taken by mobile devices without requiring the patients to visit a clinical or pharmacy site hosting fundus cameras or OCT devices, we aim to build a deep learning model for diabetic retinopathy (DR) detection using external eye images, which has immense potential to enhance large-scale DR screening. In this study, we used 7,845 samples from 6,451 patients (age: 58.5 ± 20.3 years) from Massachusetts Eye and Ear with external eye photos with corresponding demographic information and DR diagnostic information from electronic medical records. We developed an EfficientNet model to predict the vision-threatening DR (moderate + severe non-proliferative diabetic retinopathy [NPDR] + PDR) vs. the rest (normal + mild NPDR) (Figure 1 [a]). Our models were trained using 4,570 samples from 3,951 patients and tested with 3,275 samples from 2,500 patients. The area under the receiver operating characteristic curve was used to evaluate the overall performance and that between different demographic groups, and Grad-CAM was used to evaluate the model’s Interpretability.

Results : The self-reported patient demographic information is as follows: Gender: female: 53.8% and male: 46.2%; Race: White: 87.6%, Black: 7.3%, Asian: 5.1%; Ethnicity: Hispanic: 6.7%, non-Hispanic: 87.4%. Furthermore, 4.2% of the patients are vision-threatening DR. The performance disparities across races and ethnicities were substantial (White: 0.85, Black: 0.81, Asian: 0.77; Hispanic: 0.86, Non-Hispanic: 0.91). There was also a small disparity of 0.02 between males and females. The p-value between different races was p < 0.001, indicating that the disparity between racial groups is significant. Those results indicate that the development of equitable learning algorithms to improve the accuracy of minority groups is crucial. We have also shown the Grad-CAM images to indicate the important regions for predicting the DR diagnosis in Figure 2.

Conclusions : Our deep learning model can accurately predict vision-threatening DR using only the external eye photos.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1

 

Figure 2