PhotoAgeClock: deep learning algorithms for development of non-invasive visual biomarkers of aging

Abstract

Aging biomarkers are the qualitative and quantitative indicators of the aging processes of the human body. Estimation of biological age is important for assessing the physiological state of an organism. The advent of machine learning lead to the development of the many age predictors commonly referred to as the "aging clocks" varying in biological relevance, ease of use, cost, actionability, interpretability, and applications. Here we present and investigate a novel non-invasive class of visual photographic biomarkers of aging. We developed a simple and accurate predictor of chronological age using just the anonymized images of eye corners called the PhotoAgeClock. Deep neural networks were trained on 8414 anonymized high-resolution images of eye corners labeled with the correct chronological age. For people within the age range of 20 to 80 in a specific population, the model was able to achieve a mean absolute error of 2.3 years and 95% Pearson and Spearman correlation.

Keywords: age prediction; biomedical imaging; computer vision; deep learning; photographic aging biomarker; photographic aging clock.

Figure 1

Prediction error (predicted age minus true age) for the same 25 images with various resolutions. Images were passed through the developed neural network, with kernels trained for 299 x 299 pixels resolution.

Figure 2

Predicted age vs. the extent of occlusion for two persons. Picture order (up to bottom): original, covered eye area, eyelid and corner covered, and half image area covered. See text for clarifications. Real chronological age for the left subject is 50 years, for the right subject is 62 years.

Figure 3

Estimated age vs. the occlusion step for two persons. The first plot represents the results for the younger-aged person (50 years). The second plot represents the results for the older-aged person (62 years). Blue points correspond to the age produced by zeros tensor. This age reflects the initial step of age estimation by neural network model when it was fed an all-black image. This happened because of learned biases parameters.

Figure 4

Estimation error for several significant steps of occlusion. Mean and standard deviation of the error over 165 pairs of validation images (left and right eye) is reported.

Figure 5

PhotoAgeClock predicted age error for the test set within different ages.

Figure 6

Distribution of actual age in the dataset and predicted age (PhotoAgeClock) labels in the validation set.

Figure 7

Correlation between predicted age and actual age on validation dataset.

Figure 8

Algorithm performance on images obtained with professional cameras and mobile devices. (left) Algorithm performance on a high resolution photo of a celebrity (George Clooney). Chronological age of the person for the time when the picture was taken was 53 years, predicted age by two eye corner areas is 54.2 years. Editorial credit: Denis Makarenko / Shutterstock.com. (right) Algorithm performance on photo obtained with frontal camera of mobile device (selfie). Chronological age of the person is 22, predicted age by two eye corner areas is 23.5. The skin of eye area is smooth enough and young age is recognized despite the strong face expression.

Figure 9

Examples of PhotoAgeClock performance. (A) Cases when the trained model produced the lowest errors on the test set. (B) Cases when the trained model overestimated age the most on the test set. (C) Cases when the trained model underestimated the age the most on the test set. True vs. predicted age is labeled. Eye areas were erased for anonymity purposes but were present in the actual dataset pictures.