Expanding the application of AI

Researchers describe first machine learning OCT-A-based AMD classification

Cheryl Guttman Krader

Posted: Sunday, September 1, 2019

Konstantinos Balaskas MD
Initial evaluation of a machine learning algorithm for the automated interpretation of optical coherence tomography angiography scans shows that it has promising diagnostic potential for age-related macular degeneration (AMD).
The project, which was a collaboration between researchers from Moorfields Eye Hospital, London, UK, and the School of Computer Science, University of Manchester, Manchester, UK, included approximately 200 scans from patients with AMD and 100 scans from healthy controls.
As reported at the 2019 annual meeting of the Association for Research in Vision and Ophthalmology (ARVO) in Vancouver, Canada, the classifier demonstrated good performance and resilience in distinguishing eyes with wet AMD both from normal controls and from eyes with dry AMD.
“This work is a promising step towards the development of machine learning support systems for the interpretation of OCT-A scans in the diagnosis and classification of AMD,” said Konstantinos Balaskas MD, Director of the Ophthalmic Imaging Reading Centre, Moorfields Eye Hospital.
“Our future plans include enriching the dataset with more scans, conducting additional validation exercises with other datasets and developing OCT-A-based classifiers for other pathologies.”
The algorithm was based on a hybrid approach that decreases the chance of having multiple correlated features and redundant features. Two classifiers were tested, and for both the algorithm for AMD demonstrated high sensitivity and specificity when combining information from all of the retinal vascular layers or when focusing on individual layers, including the inner retinal layers.
“The inner retinal layers are not where we would expect to see changes indicative of wet AMD; although we do not yet know what pattern the classifier was identifying, this is quite an intriguing finding. There is the potential, to be confirmed by further work, that the algorithm is able to pick up differences that are not identified by clinicians,” Dr Balaskas told EuroTimes.
Looking ahead, Dr Balaskas also noted the potential of combining OCT-A diagnostic algorithms for AMD with other artificial intelligence decision support systems to develop tools for use in clinical practice.
“Artificial intelligence support systems have been developed for interpreting colour fundus images from eyes with diabetic retinopathy and OCT scans from eyes with AMD. Combining output from machine learning algorithms that interpret different but complementary imaging modalities (such as OCT and OCT-A) could provide an additional layer of certainty for diagnostic systems and support their adoption,” he said.
Konstantinos Balaskas, MD