Improving the pathway for Keratoconus diagnosis and monitoring through AI (the KerAItoconus study)

Background: Keratoconus is a common eye condition affecting the cornea which predominantly impacts young people, and if left untreated can lead to severe visual disability. Its prevalence has been estimated to be as high as 1 in 375. If diagnosed early it can be managed with cross-linking, circumventing the need for expensive corneal transplant surgery and its lifelong burden. Current hospital waiting times are up to 44 weeks for a specialist. Deep learning (DL) is a form of artificial intelligence (AI) that can achieve human expert-level performance in some medical classification tasks. Our group's early work has shown promising results in using DL to automatically detect normal from keratoconus anterior segment optical coherence tomography (AS-OCT) scans. However, high quality, precise scans are required to accurately diagnose, monitor and train the AI algorithm. Research Question: How can deep learning algorithms be developed and utilised in keratoconus to improve early diagnosis and recognition of progression? Aims: 1. Explore AS-OCT measurement precision to improve the quality of image acquisition (WP1) 2. Sort and create a large, high-quality, open-access dataset of labelled AS-OCT images for normal and pathological scans (WP1) 3. Develop and validate DL algorithms to automatically identify early keratoconus and predict progression (WP2) 4. To clinically implement the AI tool and investigate factors influencing adoption in keratoconus pathways (WP3) Design and methods The project plan comprises three work packages (WPs): WP 1 - Improving precision and dataset curation (Year 1): Measurement precision is known to deteriorate with increasing disease severity in keratoconus imaging. I will examine key keratoconus progression indices to help develop automated discard strategies for poor quality scans. I will then create and integrate a database of normal and pathological scans to train the AI algorithm. In collaboration with the Health Data Research INSIGHT hub for ophthalmology, the dataset will be fully anonymised utilising established INSIGHT data governance structures. WP2 - Development and validation of AI algorithm (Year 2 - 3): to classify scan images, I will use a custom convolutional neural network (CNN) and code-free models. Ground truth labelling will include current advanced automated labelling and human experts. I will combine with this other data such as demographics and externally validate the algorithm. WP 3 - Clinical implementation and qualitative study (Year 3): I will conduct a clinical implementation study with optometrists testing the AI tool. I will embed a qualitative study by conducting semi-structured interviews on their experience and perceived barriers to adoption. Two PPI members will help with topic guide and theme coding for WP3. Impact AI algorithms could be leveraged to detect early subclinical cases without specialist input, giving new opportunities for early treatment. Newly referred patients are waiting several months, often leading to disease progression and preclusion of less invasive/expensive cross-linking. Automated diagnosis and disease progression detection in community hubs could relieve the burden on hospital services for both new referrals and those requiring regular specialist hospital review. Furthermore, insights from the qualitative study could be extrapolated to aid adoption of AI beyond ophthalmology.