An ophthalmology conference in Sydney, hosted by Novartis, has looked at the knowledge, technologies and pharmaceuticals guiding treatment of retinal disease.
Over two jam packed half day sessions at Future Directions in Ophthalmology (FDO), ophthalmologists from around Australia were treated to insights on a broad range of topics, from the use of artificial intelligence (AI), retinal imaging and optical coherence tomography (OCT) in the diagnosis and management of age related macular degeneration, through to pregnancy and diabetic retinopathy, diet and supplements to minimise retinal (and other) disease progression, and gene therapy for glaucoma.
Prof Waldstein and his team have… demonstrated 70% accuracy in predicting low treatment requirements, 77% accuracy for high treatment requirements, and 75% accuracy for treat and extend protocol
The keynote speaker at FDO was Prof Sebastian Waldstein from the Medical University of Vienna, who is in Australia for 12 months under a vitreoretinal surgery fellowship at Westmead Hospital. At the University of Vienna, Prof Waldstein is a member of the Christian Doppler Laboratory for Ophthalmic Image Analysis (OPTIMA), a group pioneering the introduction of AI into ophthalmology.
AI AND RETINAL DISEASE MANAGEMENT
Prof Waldstein spoke extensively about the role of AI in helping medical specialists to analyse and understand rising volumes of increasingly high resolution retinal imaging and OCT scans.
“Imaging is getting better and we end up with a tremendous number of high resolution images without the time to analyse them… there are many bio markers but we need to know which ones to look for, which ones are relevant… (and AI can help us do that).”
Prof Waldstein told the audience that the areas in which AI will revolutionise retinal disease are diagnosis, treatment, and prognosis.
“The use of AI in diagnosing and managing retinal disease is already here, and some AI systems are already FDA approved and can be purchased. They have been shown to be as accurate as human experts,”1,2 he said.
Prof Waldstein stated his belief that in the future, AI will substantially enhance referral patterns, making diabetic retinopathy screening accessible in areas not currently available, such as in rural areas, in GP rooms and even supermarkets. However, its potential to screen for other retinal diseases could be limited.
“The key to accessibility will be low cost – it is easy to have low cost fundus cameras which can be used for screening for diabetic retinopathy, however in the case of macular degeneration we need OCTs and they are costly – it will be some time before OCTs are less expensive,” he observed.
Treatment paradigms are increasingly complex. Treatment for wet-AMD, for example, is now guided by the amount of retinal fluid present, but this needs to be measured.
“Specialists can’t spend hours manually assessing the amount of fluid that indicates a requirement to treat, but computers can do it very quickly – their automated measurements are as accurate as manual measurements by certified graders – and they measure more quickly and more reliably,” Prof Waldstein said.
Additionally, he highlighted the capacity of AI to predict how long we can leave individual patients between anti-VEGF treatments.
“This has been done by treating the patient monthly for the first three months, and taking OCT scans at each treatment then extracting the bio markers, such as subretinal fluid, age, gender and best corrected visual acuity at baseline, month one and month two. The biomarkers are collected over three months and used to predict the future need of treatments for the individual patient.”
Prof Waldstein and his team have used AI to do this for over 1,500 patients, and have demonstrated 70% accuracy in predicting low treatment requirements, 77% accuracy for high treatment requirements, and 75% accuracy for treat and extend protocol.
“The algorithm proved to be twice as good as the retina specialist,” he concluded. He added that being able to predict treatment intervals is very helpful for informing patients and clinics, and ensuring treatment availability.
Patients can be disappointed to hear that it is difficult to predict how they will do with their vision in response to a course of anti-VEGF therapy. Prof Waldstein and his team have shown that AI has a high level of accuracy when it comes to predicting visual acuity (VA) post intravitreal treatment. VA of 1,000 patients was measured monthly, at baseline, months one, two and three, then using this data, the system was able to predict VA with 70% accuracy after one year of treatment.
Room for Improvement
Professor Waldstein said target labels are the weakness of deep learning.
“Deep learning currently requires ‘supervised learning’, that is, you need to define what you are looking for. For instance, you need to take hundreds of thousands of images and manually mark up what you’re looking for on each, then give them to the computer. This is time and cost intensive.”
This has been overcome in one instance by training the AI system to recognise ‘normal’ as opposed to ‘abnormal’ retinal scans, a process which required far less imaging and manual processing.
“However, the area is moving into unsupervised learning – requiring no human input – and this will make it cheaper and less resource intensive for computers to become non-biased experts.”
CHANGING PARADIGM OF TREATING WET AMD
A presentation by Associate Professor Andrew Chang switched the focus to the changing paradigm of wet age-related macular degeneration treatments, and gave us a sneak peek at the future.
“We have moved from managing vision loss, and halting disease to now looking at how to optimise the current treatments we have,” he said noting that while the three current drugs used to treat wet AMD (Lucentis, Eylea and Avastin) all have slightly different characteristics, they all work.
Outcomes from clinical trials are not typically reflected in ‘the real world’ because of the challenges real world patients experience. Being a chronic disease, Professor Chang said, “the ability for the drugs to maintain vision in the real world, and over the long term, is the ultimate outcome we are looking for… we have shown that it is possible to maintain vision into five years with anti-VEGF therapy.”
Prof Chang reminded the audience that patients recruited for trials are highly motivated and offered financial support, which means they are more likely to stay with treatment compared to those who potentially are under-treated in the real world.
For those in the real world, “The main burden concern are the frequency of visits and travel time with patients on average visiting the clinic for treatment 10 times each year, which involves up to an average of four hours travel time to and from the clinic. Those with other health comorbidities are particularly challenged,” he said.
Long term challenges, which often lead to non-compliance with treatment, are the physical and financial burden of treatment, along with the patient’s expectations for visual improvement, which may differ with the clinician’s expectations.
Being a bilateral disease, patients need to be monitored for disease in their second eye. The second eye often responds to treatment better than the first, because it is diagnosed and treated earlier.
“The goal of management of wet AMD is to balance the treatment burden with its effectiveness. “With under-treatment, we don’t achieve the optimal outcomes. However over-treatment increases the burden of treatment, often over many years of treatment” said Prof Chang.
“The treat and extend protocol has been shown to help us achieve this balance with visual results that are similar to monthly injections. Good outcomes can be achieved with treat and extend protocols using Lucentis,” he said.
However, not all patients improve vision. “The trial data shows the average results achieved, however in some patients vision will initially improve then fall off, some will continue to lose vision despite treatment… Treatment resistance to anti-VEGF does occur in up to 25-30% of patients and there are a number of strategies to employ that include switching between the three agents, the results of which are variable.”
Professor Chang went on to describe future directions in treatment for wet AMD, including the much anticipated Brolicizumab, Faricimab, Abcipar and AXITINIB, a slow release implant currently in a Phase 1 trial to assess safety and tolerability.
- Clinically applicable deep learning for diagnosis and referral in retinal disease (Deep Mind),
- Pivotal trial of an autonomous AI-based system for detection of diabetic retinopathy in primary care offices (IDX)