Fig. 1: Overview of the proposed workflow for retinal aging biomarker development and evaluation. | npj Digital Medicine

Fig. 1: Overview of the proposed workflow for retinal aging biomarker development and evaluation.

From: A cross population study of retinal aging biomarkers with longitudinal pre-training and label distribution learning

Fig. 1

a Data collection and preparation: The training dataset comprises retinal images from the UK Biobank (UKB) and a Chinese (CHN) cohort, selected to represent a generally healthy population. These images undergo automated quality control through an image grading model. b Framework for estimating retinal age: In Stage I, snapshot images are used to train a masked auto-encoder. This is followed by a partial fine-tuning of the encoder with longitudinal images, enhancing its ability to model the temporal evolution for retinal aging. Stage II involves a sophisticated two-stage label distribution learning process, where a regression head is fused with the pre-trained encoder, enabling accurate prediction of retinal age. c Evaluation of the retinal aging biomarker: The model’s performance is tested across a varied demographic, utilizing the difference between calculated retinal age (RA) and chronological age (CA) as an innovative biomarker to predict susceptibility to age-related diseases.

Back to article page