Abstract
Purpose :
Machine Learning (ML) models have shown above-human performance in the classification of ophthalmic imaging. However, this is often only achieved in limited ways and in unrealistic environments. In addition, many methods use Artificial Neural Networks (ANNs), the workings of which are uninterpretable to human users. These 'black box' models can lead to unintended consequences, including misdiagnosis. As ML models find use in the clinical environment, these errors could be perpetuated at scale and jeopardise patient safety. There is an urgent need for interpretable models, since this will enable users to identify biases, train more accurate models and even find use in machine teaching. Using a new method, counterfactual visual explanations (CVEs), we explore a novel way to allow humans to understand model outputs. This allows clinicians to answer 'What if?' questions like 'How would this OCT showing AMD need to change in order for it to be classified as a healthy scan?'.
Methods :
Generative Adversarial Networks (GANs) are a gold standard ML framework for producing synthetic images. We train a state-of-the-art model, StyleGAN, using 1,000 central B-scans from patients with no pathology, and with AMD. In addition, we use a cyclical GAN, cycleGAN, that takes in a healthy scan and outputs a new image with signs of AMD. In an exploratory qualitative investigation, we show our output images to several consultant ophthalmologists for review.
Results :
After training, we are able to produce synthetic OCT B-scans at native resolution. We show for the first time that it is possible to create synthetic optical coherence tomography scans of sufficient quality that world leading retinal experts are unable to differentiate between real images and our synthetic ones. We also successfully demonstrate the first example of counterfactual image generation using GANs and provide a framework for further exploration of this method in the literature.
Conclusions :
GANs can be used to generate realistic OCT B-scans and generate counterfactual examples. Along with improved interpretability, synthetic images can be produced at scale to help with augmenting datasets and in addition obviate many of the ethical hurdles that currently exist around distributing medical data.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.