Deep learning model shows promise for DME detection

In this study, researchers developed and tested the accuracy of a deep learning model to detect diabetic macular edema (DME) from color fundus images.

Study design

This retrospective study included 17,997 color fundus photos from patients in the phase 3 RISE/RIDE clinical trials. The collected data was used for the training (80%), testing (10%) and validating (10%) the deep learning model. Researchers used the models to predict central subfield thickness (CST) and central foveal thickness (CFT), as measured by time-domain OCT. Two clinically relevant thresholds for CST (250 µm) and CFT (400 µm) were used as cutoffs.

Outcomes

The deep learning model was able to predict CST and CFT cutoffs of greater than 250 microns with an area under the curve of 0.97 and 0.91, respectively. The best deep learning model that accurately quantified the measurements had an R² of 0.74 and 0.54, respectively. Poor image quality and the presence of laser scars lowered the model’s performance.

Limitations

The macular thickness data available was based on time-domain OCT. Since only 10% of the data was used for validation, future analyses on larger validation datasets will be necessary. These findings are only applicable to DME.

Clinical significance

The authors demonstrate that deep learning models can accurately predict quantitative macular thickness measures using color fundus photography alone. This has important implications for telemedicine and may facilitate earlier detection of vision-threatening complications of diabetic retinopathy.