“We are beginning to appreciate the scope of potential applications of AI”

Could you share an insight into the project and its aims?

Helen Kuht (HK): Optical coherence tomography (OCT) is considered an extension of the eye examination, allowing for detailed visualisation of the retina, with only a short acquisition time. OCT imaging has demonstrated impressive diagnostic accuracies in a range of ophthalmic conditions. However, analysis and interpretation of OCT scans requires significant training, expertise and clinical time.

Foveal hypoplasia describes the underdevelopment of the retina, one of the major causes of reduced visual acuity in paediatrics. Differentiating between grades of foveal hypoplasia poses a significant challenge, as subtle abnormalities are difficult to detect and there can exist inter-examiner variability with interpretation. Artificial Intelligence (AI) has previously been successful in image recognition for many areas of medicine, however to date there are no AI systems for paediatric OCT.

We therefore posed the question of whether we could develop an intelligent system to accurately and quickly differentiate between retinal developmental disorders. I have been involved in preparing the training dataset, reviewing clinical notes, labelling the OCT scans and collecting prospective data.

What were the biggest challenges faced in the project?

HK: OCT machines exist in different models, such as table mounted OCT and hand-held OCT. Subsequently, generated images from each machine may differ with regards to resolution and to size or scan window. For example, the optic nerve may be present in some images and not in others. It was therefore paramount to develop an AI system that is device agnostic, allowing the algorithm to recognise the fovea morphology, regardless of the image representation. We achieved this by training the algorithm on images from both types of OCT machine.

The curation of a high-yield dataset, demonstrating a representative sample of varying degrees of arrested retinal development, also presented a significant challenge. We therefore implemented data augmentation techniques, such as rotation, translation and altering the contrast or brightness of the images, as well as longitudinal data to balance out the groups, to train the algorithm on similarly represented grades of foveal hypoplasia.

How did you go about training the AI system?

HK: In Leicester, we are in a unique position to develop this system because we have access to the largest reference dataset for paediatric retinal scans worldwide. Using this dataset of approximately 20,000 volumetric B-scans, we extracted appropriate scans, segmented and annotated. We used the foveal hypoplasia grading system developed by my supervisor, Dr Mervyn Thomas, and the team in Leicester to annotate the images, characterising foveal scans with grade 1– 4 foveal hypoplasia, atypical foveal hypoplasia or normal morphology.

For a successful AI algorithm to be developed, a large amount of data is required, hence the term ‘big data’ that we frequently hear in relation to AI. Typically, thousands or even tens or hundreds-of thousands of images are required for deep learning systems – a subtype of AI. Our completed dataset included approximately 8000 foveal scans. This dataset was then split into a training, validation and test dataset.

Training the algorithm involves running a proportion of the dataset through a convolutional neural network (CNN) with annotations. The newly developed algorithm then receives completely new data with no annotations and generates predictions – thus establishing a validation accuracy.

I don’t think it will be long before AI has a part to play in most areas of our ophthalmology clinical pathway, complimenting but not eliminating, the work of our current clinicians

What stage are you now at with the research and what are the next steps?

HK: We have achieved a working algorithm to accurately classify retinal developmental disorders, arguably surpassing the performance of human interpreters. Our next step is to develop a working graphical user interface to enable non-coder clinicians to easily access and navigate through the system.

Due to the success of our current AI system, we anticipate adapting the algorithm and applying it to different datasets, such as different diseases, using a technique called transfer learning. This may allow us to introduce an intelligent automatic classification system for other conditions.

What other applications do you think AI could have for ophthalmology?

HK: I think this an exciting time when we are beginning to appreciate the scope of the potential application of AI in various areas of medicine. The impending impact of AI systems as part of our clinical practice is already augmenting the clinical pathway in many specialities, thus increasing time efficiency in busy clinics and reducing NHS cost burden.

AI has already been introduced to many areas of adult and paediatric ophthalmology including the likes of age-related macular degeneration, diabetic retinopathy, paediatric cataracts and strabismus detection. And the list just keeps on growing. I don’t think it will be long before AI has a part to play in most areas of our ophthalmology clinical pathway, complimenting but not eliminating, the work of our current clinicians.

Our AI system has the potential to be introduced to frontline clinical assessments, therefore offering diagnosis and prognosis much earlier on in the patients’ care pathway

What did your role within the project involve?

Dr Mervyn Thomas (MT): I am the principal investigator of this project and an NIHR clinical lecturer at the University of Leicester. I developed the concept and study design for this project around two years ago. This stems from my previous work developing the grading system for retinal developmental disorders.

To get this project to fruition required funding, bringing together a team to tackle this problem and developing industry links. My first aim was to put together a unique team with skills that were critical for this project. This included strong AI scientists in the UK and India; a highly skilled research orthoptist, Helen Kuht; an industry collaborator, Leica Microsystems; and importantly the support of the local ophthalmology clinical and research team led by Professor Gottlob.

I was subsequently successful in an Medical Research Council-Confidence in Concept award to perform this proof-of-concept work. I oversee and supervise the staff working on the project, ensuring we are meeting our milestones and deliverables as per our project timeline. We tend to have weekly meetings which involves a lot of troubleshooting and collectively coming up with solutions to overcome them.

Could you tell us about the grading system you developed and how this provided a basis for the project?

MT: One of our landmark papers is on the classification of retinal developmental disorders (foveal hypoplasia), which was published in 2011. This classification introduced a grading system for foveal hypoplasia, dividing the degrees of arrested development into grades 1-4, and an additional grade, termed ‘atypical’, for photoreceptor dystrophies. The grading system is universally accepted due to the diagnostic and prognostic implications. This work formed the basis of the AI project.

However, interpreting foveal OCT scans, especially in children, can be challenging. The idea to develop an automated system for OCT interpretation was therefore proposed to complement the role of the clinician by reducing the demand for experienced interpreters in addition to standardising the classification of retinal developmental disorders.

How did you bring together a network of researchers and how did they contribute to developing the system?

MT: I initially presented my concept and study design to Professor Yu-Dong Zhang, chair of machine learning, and Professor Gottlob, professor of ophthalmology at University of Leicester. Professor Gottlob has been working on a large normative dataset project for paediatric ophthalmology so was keen to support me to develop this system and utilise the wealth of data we have been accumulating at Leicester. Professor Zhang was particularly excited about the project since my previous work with the grading system formed a strong foundation and he could apply techniques he had developed for MRI datasets to OCT. Together we successfully received funding and commenced work in January 2019.

I developed the team and set up collaborations with an AI group based in India, collaborating with Dr Girish Varma, assistant professor, IIIT Hyderbad. Dr Varma works in the intersection of machine learning, applied computer vision and theoretical computer science with a strong track record.

Professor Zhang and Dr Varma and their associated teams developed the algorithms for our classification system, adapting to our specific requirements. Both teams used different CNNs so that we were able to compare and select the CNN with the highest validation accuracy.

What was important about this collaborative approach?

MT: From previous experience, the most impactful work often occurs at the interface between different scientific fields. Understandably, it would be extremely challenging to carry out this type of research within a defined time frame, without input from more than one discipline. Thus, it was crucial to involve multiple individuals with varying expertise and backgrounds in the fields of machine learning, clinical ophthalmology, paediatric retinal imaging and industry.

Our industry collaborators, Leica, were a vital part of the team, offering continuous support and technical assistance to enable the use of encrypted data from their equipment. This collaborative approach enabled the rapid development of an initial binary classification and subsequent successful developments with a six-point classification.

How will the outcomes of the project benefit ophthalmologists in diagnosis and prognosis?

MT: Childhood retinal diseases are a significant cause of reduced vision in children, reported to account for between 25–47% of severe paediatric visual impairment. The development of the foveal hypoplasia grading system already demonstrates significant diagnostic and prognostic value. Recently we have shown that it can predict future visual acuity in preverbal children. However, introducing our AI system to carry out automatic classification ultimately accelerates and standardises this process. Thus, much of the consultation time can then be focused with the patient and family rather than analysing and interpreting scans.

The system may also provide the benefit of reducing the requirement for additional tests such as electrical testing, genetic testing and MRI, thus reducing the staff and resource demands in such departments. We envisage that this will allow specialist orthoptists and optometrists to lead and setup virtual clinics allowing quicker triaging in paediatric ophthalmology. Our AI system has the potential to be introduced to frontline clinical assessments, therefore offering diagnosis and prognosis much earlier on in the patients’ care pathway.