Automated detection of genetic relatedness from fundus photographs using Convolutional Siamese Neural Networks

Abstract

Purpose : To develop an automated algorithm for detection of genetic relatedness from color fundus photographs (FPs)

Methods : The degree of shared ancestry among pairs of UK Biobank participants (identity by descent, IBD) was estimated genome-wide using PLINK software. Pairs with IBD > 0.1875 (halfway between the expected IBD for third- and second-degree relatives) but < 0.98 (presumed duplicates) were considered related, while those with non-calculable IBD were considered unrelated. Unrelated FP pairs were generated using a combination of FPs from the related pairs and FPs randomly sampled from the rest of the data. FP pairs were divided into training and testing sets in an 80:20 ratio. A convolutional Siamese neural network-based algorithm composed of two Densenet-121 networks each pretrained on imageNet was trained to output a measure of genetic relatedness using 5620 pairs (2810 related and 2810 unrelated) of fundus images.

Results : Among the 1380 pairs of FPs in our test set, the model trained assigned each pair of images a Euclidean distance (ED), a measure of genetic relatedness. The average ED was 19.78 ± 10.06 among related pairs and 36.36 ± 12.29 among unrelated pairs, Fig 1. After normalization of EDs and using an optimal threshold for ED=27.95 (Youden’s Y point), the sensitivity and specificity of our model were 78.7% and 74.7% respectively. When the computed EDs were used to determine probability of relatedness, the area under receiver operating curve for identifying related vs unrelated FP pairs reached 0.843. The ROC curve along with the associated confusion matrix using the computed threshold is shown in Fig 2.

Conclusions : A Siamese neural network based Euclidean distance obtained from pairs of FPs is able to accurately predict genetic relatedness.

This is a 2021 ARVO Annual Meeting abstract.

 

Violinplot showing the distribution of Euclidean Distances predicted by the model separated by class

 

left : Reciever Operating Characteristics (ROC) curve of the model; right - Confusion Matrix