AI system to predict "wet" age-related macular degeneration (AMD)

 AMD involves damage to the macula, the central part of the retina at the back of the eye. AMD causes loss of central vision, affecting the ability to read, drive, watch television, recognise faces, and many other activities of daily living. It is very common that patients develop wet AMD in one eye and start receiving treatment, before later developing it in their other eye.

The AI system developed by Moorfields, as well as researchers from DeepMind, and Google Health, may allow closer monitoring of the “good eye” in patients at high risk, or even guide use of preventative treatments in the future.

Pearse Keane, Associate Professor at IoO and Consultant Ophthalmologist at Moorfields Eye Hospital, said

 
Reena Chopra, a PhD student at IoO and research optometrist at Moorfields Eye Hospital, said: 

Links 

• Nature Medicine article
• Google Health blog
• DeepMind blog 

Image 

• Clinical setup and proposed system: a, After diagnosis of exAMD in one eye (the first eye), a patient commences intravitreal therapy in that first eye. Both the first eye and fellow eye are followed up regularly with further observation. b, Selected sequential scans from the fellow eye of a patient. This eye initially showed mild, early AMD and then converted to exAMD, following which it was treated with intravitreal therapy. The timing of each follow-up visit varies depending on the treatment regimen of the first eye, as well as on factors related to the individual patient and the clinic. At each visit, an OCT scan of the first eye is performed to assess the efficacy of treatment. An OCT scan of the fellow eye is also performed, as the presence of exAMD in one eye presents a high risk of fellow eye conversion. Here, the fellow eye converts to exAMD during follow-up at 11 months (red box). c, Illustration of the proposed AI system. The 3D OCT volume of the fellow eye (1) is used to provide a risk prediction of whether the eye will convert within a given time window. A deep learning (DL) segmentation model (2) outputs a 3D segmentation of anatomical and pathological tissue (3). A prediction model then takes this tissue segmentation as an input (4). A further prediction model takes the original 3D OCT volume as an input (5), and these two prediction models are ensembled (6) to assign a risk of conversion to exAMD within a clinically actionable time window of 6 months (7).