Artificial Intelligence and ophthalmology
AI set to be used in more eyecare applications
Béatrice Cochener-Lamard MD
The future of artificial intelligence (AI) looks exciting to assist but not replace ophthalmologists, concluded speakers during a dedicated Clinical Research Symposium on AI at the 38th Congress of the ESCRS.
Béatrice Cochener-Lamard MD, PhD, France, outlined the development of AI in ophthalmology and the role of deep learning to assist in diagnosis, citing its successful use in diabetic retinopathy.
Looking at the latest developments, she said that AI technology is now well on the road to automatic image classification, with AI to soon become a widespread tool in all imaging modalities (2D and 3D and beyond).
Thanks to the creation of more refined algorithms, the use of ‘big data’ is not always necessary now in AI and there are “multiple additional applications” on the way, she said. These include using AI as an integrated part of “screening, diagnosis, decision support and maybe even surgical help”.
“So the future looks very exciting [for AI] to help ophthalmologists for sure, but never to replace us,” Prof Cochener-Lamard concluded.
Although progress in deep learning no longer requires big data, it is crucial to emphasise that the development and validation of image recognition software relies on the number and quality of images and their interpretation, which condition the training of the algorithms.
Also speaking during this session, Bruce Allan MD, UK, gave a practical presentation on the development of a machine learning accessible electronic healthcare record suitable for a patient registry.
“All machine learning studies require the same thing; high-quality labelled data, usually collected for routine clinical practice,” he explained.
Data for machine learning and registry studies has four essential attributes, Dr Allan said.
“First of all it has to be legal, in line with GDPR legislation. It has to be high quality, it has to be accessible and searchable, and it has to be secure.”
Healthcare data does qualify for certain GDPR exemptions to the strict requirement for prior consent to research use. Data does not always have to be anonymised, rather “pseudo-anonymisation”, in which personally identifiable elements are held at arms’ length from other healthcare data using an encryption key, can be sufficient, “but it is useful to have clear advice about this”, Dr Allan said.
ESCRS has now commissioned legal expertise on this topic, “which should help us in each member state to know where we stand, and remove some of the obstacles to research progress”.
In terms of collecting good-quality data, there are some generally applicable criteria that can be applied, including averaging key measurements across a sequence of three scans to improve precision, and labelling poor-quality scans that are still deemed to be clinically useful.
“ESCRS can help by setting standards for doing this and standards for simple aspects of data acquisition, like, for example, measuring intermediate visual acuity of 63cm. These kind of standards are not well defined at the moment,” Dr Allan said.
Also speaking during this session, Robert Wisse MD, PhD, the Netherlands, who discussed the use of digital eye testing, including on smartphones, in cataract and refractive care and the use of telemonitoring.
He explained that by 2040, older people will make up half of the population in Europe, and an estimated 40% of these will have three or more chronic conditions. This will create a significant extra demand on healthcare services and strengthens the need to increasingly utilise AI and telemedicine. “A paradigm shift in healthcare delivery is needed,” Dr Wisse maintained, adding that digital transition is not a goal per se but as a means to an end.
A prospective international RCT is one of eight research programmes to assess the true clinical validity and safety of this novel method for testing visual function. Easee BV (http://easeee.online) is a medtech start-up and the private partner in a public-private Digital Eye Health consortium that develops the web-based test. The tool will be integrated in several hospitals and Electronic Health Records.
IOL power calculation
Warren Hill MD, USA, gave an update on IOL power calculation driven by AI.
He explained how the use of pattern recognition based on AI data calculations, rather than the more common theoretical formula approach, could help achieve more accurate IOL power calculation results, and is particularly helpful for eyes with short axial length.
“A ±0.50D accuracy of around 78% [with traditional formulas], which is where most surgeons are now, but with AI we can routinely get to 90% provided the measurements are good. A critical aspect of the best possible refractive outcomes is validating the measurements and making sure the ocular surface has been optimised.”
Dr Hill’s quoted positive comparison data from a number of international centres on the AI-based Hill-radial basis function (RBF) calculator, of which he is the author. The latest validated data on the updated RBF calculator show results of over 91% accuracy within ±0.5D of the intended target.
Renato Ambrósio MD, PhD, Brazil, discussed the detection of corneal ectasia using AI, noting that while laser vision correction and eye rubbing are the primary environmental culprits in the development of ectasia, detection of other risk factors is key to reducing cases and ensuring early diagnosis which AI is proving useful in.
The final speaker in this session was Jodhbir S Mehta MD, Singapore, who spoke about the use of AI for classification of corneal dystrophies. He said that while retinal disease is currently leading the way in AI-related ophthalmic applications, there are plenty of anterior segment classification uses; “I see this as a platform technology we can use for multiple things.”
For stromal dystrophies he has helped create validated AI software that can monitor disease progression, recurrence after transplantation, and monitor the effect of various treatments.
Concluding, Dr Mehta said AI offers much promise in any field where there is large data and good imaging, thus it is here to stay in ophthalmology.