Retinal image segmentation using generative adversarial networks

Description

OCT imaging of the human eye provides a non-invasive technique that allows clinicians to view 2-dimensional “slices” of the nerve layers of the retina. These images provide a window into not just diseases of the eye, such as macular degeneration, but also systemic diseases such as diabetes and hypertension. The following example shows an example OCT scan of a healthy eye (top) along with scans showing macular degeneration (centre) and diabetic macular oedema (bottom) [1].

The retina itself consists of layers of different nerve types. These layers are (manually) highlighted in the following example images [2].

Ideally, clinicians are able to diagnose and track disease by identifying and tracking changes to the layers. Unfortunately this is very time consuming in practice, and this kind of personalised medical care is not widely available.

To address this problem, significant attention has been given to attempting to automatically identify the layers from OCT scans, most recently using AI techniques. This is a challenging task, since the OCT scans have low contrast and contain significant noise.

Conditional generative adversarial networks (cGANs) provide a way of generating images that meet certain constraints. (Generative networks have been prominent in the news recently in the context of large language models, such as ChatGPT.) Recently they have been applied with some success on a range of image-to-image translation tasks [3], and in particular, to noise reduction in retinal OCT scans [4].

This project seeks to further explore the use of cGAN-based image-to-image techniques in the retinal segmentation problem. In particular, the aim is to better understand the role of different sample distributions and loss metrics, and their impact on automatically generated images.

Assumed Skills

Fluency in python. Taking the Machine Learning (CITS5508) and/or Deep Learning (CITS5517) units is recommended.

Sample references

[1] T. Ilginis, J. Clarke, and P. Patel, "Ophthalmic imaging". British Medical Bulletin. 111. 10.1093/bmb/ldu022.

[2] J. Wu, J. Chen, Z. Xiao, and L. Geng, "Automatic Layering of Retinal OCT Images with Dual Attention Mechanism," in 3rd International Conference on Intelligent Medicine and Image Processing, 2021.

[3] P. Isola, J. Zhu, T. Zhou, and A. Efros, "Image-to-Image Translation with Conditional Adversarial Networks", 2018, 1611.07004, arXiv.

[4] M. Mehdizadeh et al., "Texture loss to denoise OCT images using Generative Adversarial Networks", to appear.