Prediction of gender from macular optical coherence tomography using deep learning

Abstract

Purpose : Differentiation of gender from color fundus photography using artificial intelligence has been reported. However, the actual structural differences in retina between male and female have not been well known. This study was to use deep learning algorithms to identify the gender with macular optical coherence tomography and to explore the structural differences in retina between male and female.

Methods : A total of 5,024 healthy eyes receiving 6x6-mm² macular volume scanning were enrolled in this study, and 120,576 B-scan OCT images were obtained. We randomly classified these images into the training dataset, validation dataset and testing dataset with a ratio of 8:1:1. A convolutional neuronal network (modified VGG19) was used for feature extraction and classification, and the gradient-weighted class activation mapping (Grad-CAM) was used to produce localization maps highlighting the important regions in the images for predicting the gender.

Results : After the training and validation, the model could predict the gender in the testing dataset with an accuracy of 91.4% (AUC=0.908). The Grad-CAM showed that the model focused mostly on the retinal nerve fiber layer, ganglion cell layer and the retinal pigment epithelium layer for the prediction.

Conclusions : The deep learning algorithm we developed could predict the gender efficiently with a high accuracy. There might be differences in the retinal nerve fiber layer, ganglion cell layer and retinal pigment epithelium layer between male and female.

This is a 2020 ARVO Annual Meeting abstract.

 

Receiver operating characteristic (ROC) curve and area under the curve (AUC).

 

Gradient-weighted class activation mapping (Grad-CAM).