DEEP LEARNING PREDICTION OF AGE AND SEX FROM OPTICAL COHERENCE TOMOGRAPHY

Convolutional neural networks (CNNs) have achieved remarkable success in predicting clinical information and individuals' characteristics from medical images. Previous ophthalmological studies have suggested that age and sex have retinal manifestations that can be observed in retinal optical coherence tomography (OCT) scans. Following on these studies, we evaluated the use of three-dimensional CNNs for predicting the subject's age and sex directly from 3D retinal OCT scans. We also assessed the effect of the receptive field size on the model performance. In addition, we adopted a robust and simple bias-adjustment scheme for further performance enhancement of eye age prediction. We used a large dataset consisting of 66,767 subjects with OCT scans from the UK Biobank data and evaluated our model on 10% of the dataset (i.e. 6,676 subjects). An accurate prediction was obtained for age (mean absolute error (MAE): 3.3 years, coefficient of determination R-2 : 0.89) while an acceptable performance was achieved for sex (area under the curve (AUC): 0.86).