Abstract
Purpose :
Optical Coherence Tomography (OCT) has become the most common imaging modality in Ophthalmology due to its versatility and non-invasive approach. The abundance of OCT images has made it a strong candidate for training deep learning (DL) models which require substantial amounts of diverse, labeled data. However, training state-of-the-art neural networks through supervised learning requires two components: the original OCT image and the corresponding retinal tissue segmentation. Currently, OCT segmentation is done manually, which is a laborious and unfeasible process due to the sheer volume of data. Our solution to this problem utilizes deep learning techniques to generate large volumes of high-quality, labeled synthetic ophthalmic images. We aim for future researchers to utilize synthetic data to accelerate the transition of AI to the clinic setting.
Methods :
We employ a Generative Adversarial Network (GAN) trained on 87 real OCT scans obtained from openICPSR, which learns the underlying properties of the real data. Upon convergence, the network has learned a distribution that we can sample from to produce unique, synthetic OCT scans. We then use a Convolutional Neural Network (CNN) to produce segmentation masks that are paired with their corresponding synthetic OCT scan, a two-fold hierarchical generation process that was pioneered by Guibas and Virdi et al.
Results :
After sampling a sufficient number of synthetic pairs of OCT scans and producing their corresponding segmentations, we use this data to train a third neural network that serves as the control variable. By comparing the accuracy of the third segmentation network when trained on synthetic images versus real images alone, we found that synthetic images achieved a similar Area Under Curve (AUC) score.
Conclusions :
We have shown equal accuracy in training artificial intelligence (AI) models using synthetic ophthalmic images compared to real images alone. As ocular imaging requirements increase and new technology develops, a growing clinical need emerges for autonomous diagnostic methods, specifically for early detection and prevention of disease. Synthetic data has strong potential in Ophthalmology to increase efficiency in training diagnostic AI models and image segmentation, while concurrently minimizing human error, sidestepping accessibility concerns, and reducing data collection costs.
This is a 2021 Imaging in the Eye Conference abstract.