JMIR Publications

ABSTRACT

Background:

Deep Learning (DL) algorithms have been built for detection of systemic and eye diseases from retinal photographs. The retina possesses features which can be affected by gender differences, and the extent to which these features are captured upon photography differs depending on the retinal image field.

Objective:

To compare DL algorithms’ performance in predicting gender when using different fields of retinal photographs (disc-centered, macula-centered, peripheral).

Methods:

This retrospective cross-sectional study included 172,170 retinal photographs from 9956 adults aged ≥ 40 years from the Singapore Epidemiology of Eye Diseases (SEED) Study. Optic disc-centered, macula-centered and peripheral field retinal fundus images were included in this study as input to a DL model for gender prediction. Performance was estimated at individual level and image level. Receiver operating characteristic (ROC) curves for binary classification were calculated.

Results:

The DL algorithms predicted gender with area under the ROC (AUC) of 0.94 at individual-level and AUC of 0.87 at image-level. Across the three image fields, the best performance was seen in disc-centered (AUC: 0.91 in younger and 0.86 in older age subgroups), and peripheral field images showed the lowest performance (AUC: 0.85 in younger and 0.76 in older subgroups). Between the three ethnic subgroups, performance was lowest in the Indian subgroup (AUC: 0.88) compared to Malay (AUC: 0.91) and Chinese (AUC: 0.91) when tested on disc-centered images. The performance of gender prediction at the image level was better in younger age subgroups of < 65 years (AUC: 0.89) than in older age subgroups of ≥ 65 years (AUC: 0.82).

Conclusions:

We confirmed that gender can be predicted from retinal photographs using DL in Asian population, and the performance of gender prediction differ according to field of retinal photographs, age-subgroups, and ethnic groups. Our work provides a further understanding of using DL models for prediction of gender-related diseases. Further validation of our findings is still needed.