Fundus Photograph-based Deep Learning for Estimation of Blood Bilirubin in a Chinese Population

Abstract

Purpose : To propose a deep learning method to estimate the values of blood total bilirubin and direct bilirubinbased on fundus photography.

Methods : The Kailuan Study is a population-based prospective cohort study, with a total of 101,510 individuals aged 18-98 years recruited. Participants were followed and re-examined every 2 year since 2006-2007. All participants recruited underwent routine history and physical examinations, anthropometry, and laboratory assessment. Ophthalmic examinations were applied, including visual acuity, ocular biometry, fundus photography and optical coherence tomography, in a randomly selected subgroup composed of 14594 participants during 2015 to 2016. The bilirubinblood test was performed within 1 year of the time fundus image was taken (Table 1). A shift group model was applied to group the fundus photos within a specified range of bilirubin (0.5μmol/L), based on which a training set (70%) and a validation set (30%) with comparable distribution of imaging sort were established (Fig 2). A deep learning model of mixture of experts was used to explore the relationship between the bilirubin and the retina imaging (Figure 3).Performance of the model for discrimination of direct and total bilirubinwith fundus photography were characterized by coefficient of determination (R²), mean absolute error (MAE), and root mean absolute error (RMSE).

Results : Our mixture of experts model was able to estimate the value of both direct and total bilirubinfrom fundus photograph, with R²of 0.986 and 0.861, respectively (Table 2). The performance was also demonstrated by the confuse matrix (Figure 4).

Conclusions : With the proposed deep learning model, the bilirubin value can be accurately calculated from fundus photography independently in a large Chinese population. It suggested the association between the eye and the body, and the potential value of retina imaging for systematical disease discrimination in the future.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.