# Augmenting a prognostic deep learning system for referable diabetic retinopathy and maculopathy with synthetic retinal images

## TL;DR

This study shows that adding synthetic retinal images improves AI models for predicting diabetic eye disease, but the benefits don't consistently apply to new data.

## Contribution

The novel use of synthetic retinal images generated by a conditional cascaded diffusion model to enhance deep learning system performance for diabetic retinopathy prediction.

## Key findings

- Synthetic retinal images generated by CCDM are realistic and comparable to real images.
- Augmenting with ×2 synthetic positive cases improved internal test AUROC but not external test performance.
- Resampling real positive cases alone did not improve model performance.

## Abstract

Labelled data scarcity and class imbalance are common deep learning system (DLS) development challenges. We investigated if synthetic retinal images from a conditional cascaded diffusion model (CCDM) improves prognostic DLS (pDLS) performance for 2-year incident referable diabetic retinopathy or maculopathy (rDR/rM) prediction.

Macula images from 72,559 eyes (September 2013 to December 2019) from the UK South-East London Diabetic Eye Screening Programme (SEL-DESP) formed the development dataset, whilst 9,071 eyes were used for internal testing. Images from 2,842 eyes from Birmingham DESP were used for external testing. Prognostic DLS were augmented with ×1, ×2, and ×4 additional synthetic positive cases (pDLS-G) and compared to unaugmented (pDLS-N) and ×1 positive-case resampled pDLS (pDLS-R) using the Area-Under-the Receiver Operating Characteristic curve (AUROC).

Here we show that CCDM generate realistic synthetic retinal images that are comparable to real images and demonstrate the utility of synthetic retinal images in augmenting the development of a pDLS. The internal and external test AUROC for the pDLS are 0.827 (95% CI: 0.794–0.861) and 0.756 (0.680–0.831), respectively. Augmentation with ×2 additional synthetic positive cases (pDLS-G ×2) significantly improves the internal test AUROC to 0.845 (95% CI: 0.812–0.877, p = 0.044) but does not improve the external test AUROC 0.717 (0.633–0.828, p = 0.243). Resampling positive real cases alone does not improve pDLS-R performance.

Augmenting pDLS with synthetic retinal images significantly improves pDLS performance on internal testing but not external testing suggesting further research is required to enhance the generalisability of synthetic retinal image augmentation.

Artificial Intelligence (AI) can be applied to images of the eyes to predict the presence of diabetic eye complications. However, setting up AI systems that enable this requires labelled images. This study explored whether adding images from the back of the eye (retina) could improve predictive AI model accuracy in identifying diabetic eye disease. We compared whether AI models trained using real images and images generated by computers could be used to predict future eye complications. Images generated by computers had similar features to real images and improved predictive AI model accuracy. The key finding is that synthetic retinal images can enhance predictive AI model performance. This could mean that AI created with the help of synthetic retinal images would be better at predicting diabetic eye disease, and this could be used as an accurate tool to assist clinicians to better monitor and treat the disease.

Nderitu et al. develop a deep learning system to predict two-year referable diabetic retinopathy and maculopathy, augmented with synthetic retinal images generated by a conditional cascaded diffusion model. Augmenting training data with synthetic images improves internal test prognostic model performance but did not generalise on external testing.