Varun Ranganathan, MCOptom

Clinical Optometrist

An OCULAR Interface Exclusive

Keywords: Generative Adversarial Networks, Artificial Intelligence, Eye-care, Deep Learning, Ocular Imaging.

Introduction

In the field of Ophthalmology, recent advances using Artificial Intelligence (AI) algorithms is gradually gaining momentum and is being used extensively in diagnosing certain common eye conditions such as Age-related Macular Degeneration (AMD) and Diabetic Retinopathy (DR). Deep Learning (DL), a subset of AI and Machine Learning (ML), are used to create these algorithms because they are programmed to think and make decisions like the human brain ¹.

Training the algorithms requires large datasets which have their own disadvantages such as patient consent, data protection, and data classification ². To overcome this major hurdle, Generative Adversarial Networks (GAN) were developed which consists of two neural networks, Discriminator and Generator ³. These neural networks work in equilibrium to create artificially generated images which are like the original images. These images can be used to train computers to identify different eye conditions.

GANs in Fundus Imaging

Apart from generating normal fundus images, GANs have been used to generate fundus images on AMD, DR, Glaucoma, and Retinopathy of Prematurity (ROP) with the data obtained from the international cohort studies. The algorithms have not only been able to create clinically accurate images but also label the different landmarks in the pathophysiology of the eye conditions ⁴, which will also be helpful in classifying the eye conditions based on the severity and progression. Such algorithms will be immensely useful in screening programs and regular eye examinations, resulting in appropriate and timely referrals to Ophthalmologists for treatment.

GANs in Optical Coherence Tomography (OCT)

OCT is regarded highly among eye-care professionals for its ability to analyse and produce high-quality images of both anterior and posterior segments of the eye. It creates a three-dimensional structure of the tissue of interest by using a combination of both A and B-scans ⁵. GANs are used to synthesise and classify images based on the urgency of referrals and have been successfully tried on Diabetic Macular Oedema (DMO) and Choroidal Neovascularisation (CNV) ⁶. GANs have also been used to reduce speckle-noise, which is often experienced in OCT imaging, to enhance the image quality ⁷.

OCT-Angiography or OCT-A was first introduced in 2014 and has been used to detect and monitor vascular pathologies of the retina and choroid. It is preferred over Fundus Angiogram (FA) due to being a non-invasive technique and reducing allergic reactions to dyes. Due to its versatility and ease of use, it has been employed in monitoring pre and post therapeutic changes of AMD ⁸ and has enormous potential to be used in clinical trials featuring retinal drug implants.

Drawbacks of GANs

As discussed above, two neural networks work in equilibrium to synthesise retinal images. The generator produces synthetic images, and the discriminator tries to distinguish the fake from the real images. The main drawback of GAN is mode collapse ⁹, where the generator fails to fool the discriminator or in other words, it falls into a loop of producing limited images without generating diverse outputs, thus not utilising the entire dataset. To overcome this, several theories have been put forward; from using multiple generators and discriminators to manually monitoring and training the GAN. These strengthened GAN models can drastically change the current screening programs and aid in early detection of eye conditions.

References

1. Choudhary, K., DeCost, B., Chen, C. et al. Recent advances and applications of deep learning methods in materials science. npj Comput Mater 8, 59 (2022). https://doi.org/10.1038/s41524-022-00734-6.
2. Office for Civil Rights (OCR), HIPAA for Professionals. U.S. Department of Health & Human Services. Available at: https://www.hhs.gov/hipaa/for-professionals/index.html.
3. Alankrita Aggarwal, Mamta Mittal, Gopi Battineni, Generative adversarial network: An overview of theory and applications, International Journal of Information Management Data Insights, Volume 1, Issue 1, 2021, 100004, ISSN 2667-0968, https://doi.org/10.1016/j.jjimei.2020.100004.
4. Beers A, Brown J, Chang K, et al. High-resolution medical image synthesis using progressively grown generative adversarial networks. ArXiv 2018; 1805.03144.
5. Aumann S, Donner S, Fischer J, et al. Optical Coherence Tomography (OCT): Principle and Technical Realization. 2019 Aug 14. In: Bille JF, editor. High Resolution Imaging in Microscopy and Ophthalmology: New Frontiers in Biomedical Optics [Internet]. Cham (CH): Springer; 2019. Chapter 3. Available from: https://www.ncbi.nlm.nih.gov/books/NBK554044/ doi: 10.1007/978-3-030-16638-0_3.
6. Zheng C, Xie X, Zhou K, et al. Assessment of generative adversarial networks model for synthetic optical coherence tomography images of retinal disorders. Transl Vis Sci Technol 2020; 9:29.
7. Kande NA, Dakhane R, Dukkipati A, Yalavarthy PK. SiameseGAN: A generative model for denoising of spectral domain optical coherence tomography images. IEEE Trans Med Imaging 2021; 40:180–192.
8. Lee H, Kim S, Kim MA, et al. Post-treatment prediction of optical coherence tomography using a conditional generative adversarial network in age-related macular degeneration. Retina 2021; 41:572–580.
9. Kossale, M. Airaj and A. Darouichi, Mode Collapse in Generative Adversarial Networks: An Overview, 2022 8th International Conference on Optimization and Applications (ICOA), Genoa, Italy, 2022, pp. 1-6, doi: 10.1109/ICOA55659.2022.9934291.