Generation of Synthetic Images of Retinal Scans
Data scarcity and class imbalance are issues regularly encountered by AI practitioners when dealing with medical imaging and clinical data pertaining to rare diseases. Especially with IRD genetic diagnosis, the rarity of certain conditions creates a natural imbalance that makes diagnostic models slightly biased and hence challenging to train. Our study aimed to address this issue by implementing GANs to create synthetic data samples that could augment the dataset downstream. Specifically, we implemented MSGGANs in order to generate high-quality high-resolution FAF images of 36 IRD gene classes. With the trained generator of the GAN model, new latent vectors were sampled for each class and inputted to produce new images. These images were evaluated using subjective visual inspection as well as quantitative metrics.
A challenge often encountered in the development of deep convolutional neural networks for image segmentation is a lack of representative data. Manual delineation of features is tedious and time-consuming, and availability of training data is therefore often limited. Additionally, some features are much less common than others, resulting in additional challenges during model development. Finally, differences between populations and cameras often negatively affect generalizability of developed models.
A promising approach to address these issues could be to use synthetically generated data for model development. Recently, the use of generative adversarial networks (GAN's) has received a lot of attention. With this approach, a model that generates synthetic images (generator) is trained simultaneously with a model that distinguishes synthetic from real images (discriminator). The feedback of the discriminator, while getting better at identifying synthetically generated images, is used to update the generator. Ultimately, the generator is able to generate realistic looking images. These images are generated either from random noise as input, or may use another image as input.
In this project, we aim to generate realistically looking retinal optical coherence tomography (OCT) images, using a target segmentation mask as input. This is the opposite of what a normal segmentation model does. The ultimate goal is to obtain a model that can generate realistic OCT images, modeled to contain specific abnormalities in certain areas.
- Model a retina picture from first principles or images from literature or UK Biobank
- Myopia oct transformation (morphing of retina).
- Age augmentation, can we age a retina synthetically?
- Ethnicity augmentation, can we transform across ethnicities?
- Generate disease images (e.g introduce edemas)
- Synthetic nevus generator
- Standardisation of Optos (e.g dealing with the artifactual colouring)
- Colour fundus to Optos or vice-versa