Purchase this article with an account.
Nergis C Khan, Chandrashan Perera, Eliot R Dow, Theodore Leng, Vinit B Mahajan, Prithvi Mruthyunjaya, Diana V Do, David Myung; Predicting systemic health features from retinal fundus images using transfer-learning based AI models. Invest. Ophthalmol. Vis. Sci. 2022;63(7):3006 – F0276.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Whilst color fundus photos are used in routine clinical practice to diagnose ophthalmic conditions, a growing body of evidence suggests that ocular imaging contains valuable information regarding systemic clinical features of patients. These features can be identified through a variety of computer vision techniques including deep learning (DL) AI models. Predictions regarding coronary heart disease, chronic kidney disease, hypertension, current smoking status, and gender have been demonstrated in prior studies. In this study, we aim to both construct a DL model that can predict a variety of systemic features from fundus images and determine the optimal method of DL model construction for this specific task.
Data were collected from a cohort of patients who had routine fundus imaging for diabetic retinopathy screening between March 2020 and March 2021. This data consisted of matched fundus images and clinical data from chart review. A series of DL models were then created and trained based on the DenseNet201 architecture to predict each of the recorded clinical features. In order to ascertain the optimal method of creating these models, two models were created for each clinical feature - one utilizing transfer learning with images from the ImageNet database, and one without transfer learning.
A total of 1277 fundus images were used to train the DL models to predict various systemic clinical features. Area Under the Receiver Operating Characteristics (AUROC) scores were used to compare the performance of each model. We found that models utilizing transfer learning were superior to those that did not (mean AUROC 0.78 vs 0.63, p < 0.05). Models using transfer learning were able to predict the following systemic features Age > 70 (AUROC 0.95), Ethnicity (AUROC 0.93), Gender (AUROC 0.85), ARB Medication use (AUROC 0.81), and ACE Medication use (AUROC 0.81).
Fundus images contain valuable information about the systemic characteristics of a patient. A DL model can predict Age, Ethnicity, Gender and ARB/ACE medication usage with a good degree of accuracy. To optimize DL model performance, we recommend that even domain specific models utilize transfer learning techniques to enhance performance.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only