Glaucoma is the second leading cause of blindness worldwide. The disease progresses very slowly and destroys vision gradually, starting with the side vision and narrowing over time. It often remains undetected until irreversible eye sight is lost at later stages. It’s no surprise then, that glaucoma has earned a reputation as the silent thief of sight, with an estimated 50 percent of cases going undetected, leaving people unaware that they’re slowly going blind.
It can be treated but early detection is critical in ensuring effective treatment. The first challenge we face with eye diseases like glaucoma as well as diabetic retinopathy and age-related macular degeneration is that in many cases, blindness is preventable (or at least slowed). If detected early enough in the majority of patients, it could have a profound impact on not only their quality of life but also the economic strain on health care systems.
Today, IBM Research is using the cognitive computing power of Watson, to progress the science of medical imaging analysis of eye images, which could in the future make the early detection process significantly faster and more accessible for all patients. Another challenge we face is that there may be a limited supply of specialised clinicians experienced in identifying subtle changes in retinal images. Often this means it is costly to visit or difficult to access for those patients in remote areas. Convenient and affordable access to regular screening is critical in the identification of not only glaucoma, but also all preventable eye diseases. We need to innovate to improve access to regular eye disease screening, for everyone.
Since 2015, scientists from the IBM Research Lab in Australia have been applying deep learning and image analytics capabilities to 88,000 retina images accessed from EyePACS, a global web-based platform that enables the exchange of eye-related images and clinical information. By understanding what constitutes the regular anatomy of an eye, the technology is being trained to identify possible abnormalities which may indicate the early onset of eye diseases like glaucoma.
The research results that we’ve announced today have indicated a statistical performance of 95 percent in the technology’s ability to measure the ratio between two parts of the eye, the optic cup and the disc. Identifying an increased optic cup to disc ratio could be a sign of Glaucoma and inform a need for further tests. Another key factor for eye disease analysis is the ability to automatically identify left from right in a retina image. With 94 percent confidence in determining left from right images, this technology could help streamline some of the manual processes that support optometrists and ophthalmologists today.
A clear vision for eye health
IBM Research’s early successes could help make eye examinations accessible to far more people worldwide than ever before. Our researchers will continue to progress the science of medical imaging analysis for retinal images, including the ability to understand a broader range of eye diseases such as diabetic retinopathy, cataracts and age-related macular degeneration. Giving Watson eyes is an exciting area of research. It could one day be key to helping free up our expert clinicians to focus more of their efforts on targeted treatment and management of these diseases.
P. Roy, et al. “Automatic Eye Type Detection in Retinal Fundus Image Using Fusion of Transfer Learning and Anatomical Features.” International Conference on Digital Image Computing: Techniques and Applications (DICTA), 2016.. http://ieeexplore.ieee.org/abstract/document/7797012/
D. Mahapatra et al. “Retinal Image Quality Classification Using Saliency Maps and CNNs.“ Machine Learning in Medical Imaging, Volume 10019 of the series Lecture Notes in Computer Science, pp 172-179. http://link.springer.com/chapter/10.1007/978-3-319-47157-0_21.
S. Sedai, et al. “Segmentation of Optic Disc and Optic Cup in Retinal Fundus Images Using Coupled Shape Regression.” , Proceedings of the Ophthalmic Medical Image Analysis Third International Workshop (OMIA 2016) Held in Conjunction with MICCAI 2016, pp 1-8. http://ir.uiowa.edu/omia/2016_Proceedings/2016/1/