Machine learning for retinal image analysis

Abstract

Retinal images, images of the retina at the back of our eyes, are an important part of modern ophthalmology and further capture the retinal vasculature and nerves, which could allow insight into cardio- and neurovascular disease. This is especially promising as retinal images are non-invasive, fast-to-acquire and low-cost compared to other types of medical imaging such as brain magnetic resonance imaging. A variety of retinal imaging modalities exist, most importantly traditional colour fundus photography (CFP) and optical coherence tomography (OCT). CFP is the most widespread type of retinal imaging and captures a true colour en-face image of the retina, typically with a field of view of around 45 degrees. OCT imaging captures the retina in depth and thus allows assessment of individual layers of the retina and – with modern methods such as Enhanced Depth Imaging – even captures the choroid, a dense vascular tissue beneath the retina. More recent modalities include OCT angiography which uses repeated OCT images to estimate blood flow and ultra-widefield fundus imaging which captures most of the retina with a field of view of around 200 degrees. Retinal imaging is already widespread and continuously proliferating: lower-cost handheld devices or smartphone addons make CFP available in lower resource settings, while once cutting-edge OCT can now be found at high-street opticians in the UK.

Retinal images provide a wealth of information but are complex to analyse, in part due to variations in image quality, anatomy, and retinal pathology that make traditional development of handcrafted analysis pipelines challenging. The recent decade saw great advances in machine learning methods, particularly deep learning for computer vision. Instead of manually designing a pipeline, a machine learning model is a parameterised pipeline that can be fit to training data to approximate the mapping from inputs to outputs. This approach is highly effective for many vision tasks, including classification, regression and segmentation.

In this thesis, I present three themes of work using machine learning for retinal image analysis. First, using machine learning for retinal disease detection. Second, using machine learning for developing efficient and robust automated analysis pipelines for retinal imaging. And third, validating and applying these tools.

For the first theme, I developed a deep learning model that can detect seven key retinal diseases in ultra-widefield pseudo-colour retinal images with very promising performance and investigate which regions of the ultra-widefield images are important for automated disease detection in a data driven way.

For the second theme, I developed three tools. First, deep approximation of retinal traits, or DART for short, that computes retinal fractal dimension (FD), a metric relating to the complexity of the blood vessels in CFPs, orders of magnitudes faster and more robustly than traditional methods. Second, jointly with a colleague, I developed a tool initially for segmenting the choroid region in OCT, called DeepGPET. Next, we developed Choroidalyzer, which segments the choroid and the choroidal vasculature while also identifying the location of the fovea. This allows for fully-automated computation of choroidal thickness, area, vascular index in a fovea-centred region of interest. Third, I developed QuickQual an efficient and easy-to-use method for CFP quality assessment that obtains state-of-the-art performance on a commonly used quality assessment dataset.

Finally, for the third theme, I applied DART to real-world, primary care data and found a significant association between lower FD and prevalent systemic health conditions. Furthermore, I compared the repeatability and robustness of DART to AutoMorph, a method that follows the traditional paradigm for computing FD, finding that DART was not only more robust to image quality issues but also more repeatable even for high quality images.

In my opinion, this thesis exemplifies the potential of machine learning for retinal image analysis. I hope that my work will – eventually and incrementally – advance the field of retinal image analysis and one day make a positive difference for clinical practice.