Conditional Diffusion Models and Retinal Image Synthesis in Diabetic Retinopathy

Abstract

Purpose : A common limitation when developing deep learning models (DLM) is the paucity of labelled samples and class imbalance. Conditional diffusion models (CDM) have recently demonstrated high-quality image generation capabilities. We investigated if CDM-generated retinal images could improve the performance of a DLM that predicts incident referable diabetic retinopathy (DR) or maculopathy.

Methods : From 203,983 eyes of 102,601 patients from the Southeast London diabetic eye screening programme (Sept 2013 to Dec 2019), 90,476 eyes with two visits 2 years apart were included. A development set of 72,559 baseline macula images (512px) were used to train a native DLM (DLM-n) to predict incident referable DR or maculopathy within 2 years. An additional DLM (DLM-r) was then trained using x1 resampled incident referable DR or maculopathy cases from the development set. CDM conditioned on age group, sex, ethnicity, diabetes duration group, baseline DR and incident referable DR or maculopathy was subsequently trained using the development set. The trained CDM generated x1, x2, and x4 additional positive-case images with similar characteristics to development set counterparts. Finally, DLMs were trained using real and x1, x2, or x4 generated positive-case images (DLM-g). The area-under-the receiver operating characteristic (AUROC) summarised native, resampling, and generated DLM performance using one real test set eye per patient chosen at random (n=9,071).

Results : No and incident referable DR or maculopathy case ratios were 52:1 for the native development set, 26:1 plus x1 resampled positive cases, 26:1 plus x1 generated positive cases, 18:1 plus x2 generated positive cases, and 11:1 plus x4 generated positive-cases. The AUROC for predicting 2-year incident referable DR or maculopathy were 0.827 (0.794-0.861) for DLM-n, 0.823 (0.788-0.857, p=0.671 vs DL-n) for DLM-r, 0.847 (0.816-0.877, p=0.079) for x1 DLM-g, 0.851 (0.820-0.882, p=0.044) for x2 DLM-g, and 0.844 (0.812-0.875, p=0.145) for x4 DLM-g. Test set Fréchet Inception Distance was 9.25 suggesting generated and real test images had similar feature distributions.

Conclusions : CDM can generate high-quality synthetic retinal images with realistic clinicodemographic features. Synthetic macula images improved the performance of a DLM for predicting incident referable DR or maculopathy by augmenting the positive-case limited, imbalanced development dataset.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.