Purchase this article with an account.
Henrik Hee Seung Yang, Tyler Hyungtaek Rim, Yih Chung Tham, Tae Keun Yoo, Geunyoung Lee, Youngnam Kim, Tien Yin Wong, Ching-Yu Cheng; Deep learning system differentiates ethnicities from fundus photographs of a multi-ethnic Asian population.. Invest. Ophthalmol. Vis. Sci. 2020;61(7):5248.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
There is increasing research using deep learning (DL) algorithms for detection of major retinal diseases such as diabetic retinopathy from color fundus photographs (CFP). Because eye diseases vary substantially by ethnicity, it is important to determine if DL can also adequately differentiate ethnicity based on CFP alone. Hence, in this study, we aim to construct and evaluate performance of a CFP based DL system capable of distinguishing Chinese, Malays and Indian ethnicities in the Singaporean population.
We used CFPs from the Singapore Epidemiology of Eye Diseases Study (SEED), a population-based study consisting of Chinese (patients/image (P/I) n = 3,304/30,508), Malays (P/I n= 3,135/29,205), and Indians (P/I n= 3,309/50,386) conducted in 2004-2015. For each patient, CFPs of left and right disc-centred and macular-centred fields were taken. Poor quality CFPs were excluded. CFPs were pre-processed, data augmented, and divided into 4:1 ratio for training and testing a DL convolutional neural network (CNN) with VGG16 architecture and global average pooling with batch normalization with batch size 64 and training over 200 epochs. Accuracy was evaluated using NumPy, SciPy, matplotlib, and scikit-learn. Afterwards, Guided Grad-CAM saliency maps were generated to highlight CNN focused regions used for ethnicity prediction.
In the testing set of 2012 patients (22,412 images), the overall accuracy in differentiating the 3 ethnicities was 92.1% based on single-field CFP performance, 95.1% on combined disc and macula-centred CFPs, and 97.8% for patient ethnicity classification. Generated saliency maps highlighted optic disc, vasculature, and retinal pigment areas as DL-focused regions (Figure 1).
DL can distinguish ethnicity from CFP with high accuracy among Chinese, Malay, and Indians. Further development and external validation in other ethnicities (whites, African-Americans and Hispanics) will potentially improve the progress of DL application in retinal disease detection using CFP.
This is a 2020 ARVO Annual Meeting abstract.
Figure 1. Saliency maps highlight DL-focused regions for ethnicity prediction. The pattern and distribution vary along the retinal vessel contour, optic disc margins and pigmented retinal areas, across the 3 ethnic groups.
Figure 2. DL model. Convolutional feature encoder with VGG16 architecture and a fully connected layer after compressing feature vectors by global average pooling.
This PDF is available to Subscribers Only