Age Prediction from Retinal Fundus Images and Segmented Vessel Images using Deep Learning

Abstract

Purpose : Establishing the structure-function relationship between ophthalmic images and systematic features is challenging, but recent advancements in deep learning and high-resolution imaging have made it possible. This study investigated whether the incorporation of the vascular segmentation of retinal images conferred an improvement in age prediction over the raw color fundus image alone.

Methods : The study analyzed the data from the United Kingdom Biobank (UKBB), a population-based cohort of 67,311 individuals with retinal images. After data exclusion based on image quality, 29,190 individuals were kept in the dataset. Retinal color photographs were segmented by using the open-source deep learning model, AutoMorph. We developed Convolutional Neural Network models by using retinal fundus images and vessel images as combined data input. The model was evaluated on the test set of 7,298 patients. The performance was specifically assessed in clinically relevant subgroups stratified by smoking status, hypertension, and diabetes mellitus performance. Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) with 95% confidence intervals were used to evaluate the model’s performance.

Results : The developed model, using combined vessel and raw images, achieved an MAE of 3.05 (95% CI 3.01, 3.09) and an RMSE of 3.78 (95% CI 3.72, 3.84). This performance was higher than using retinal fundus images alone (MAE: 3.31 (95% CI 3.25, 3.37), RMSE: 4.18 (95% 4.11, 4.25)). When the model of vessel and raw images was evaluated on predicting the age range ± 5.0 years of the actual age, the accuracy on the standard test set was 81.27%, while that of retinal images alone is 77.56%.

Conclusions : Our approach of combining segmented vessel images and retinal fundus images shows significant improvement in predicting patients' actual age as compared to using vessel features or retinal images alone. Future studies can evaluate the potential of utilizing structural annotations on predicting additional systemic traits.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Comparing the predicted age from Fundus and Vessel Images with chronologic age.

 

Prediction model schematics of 4-channel inception v3 with Retinal Fundus Images and Vessel Images as combined input.