Principal Investigator: Professor Mirza Faisal Beg
Simon Fraser University, Burnaby, CanadaTags: 43402, fully convolutional network, multiclass segmentation and detection, optical-coherence-tomography, retinal fluid
Age-related macular degeneration (AMD) is the third leading cause of blindness in the world and the first in industrialized countries. Patients with AMD can suffer from blurriness and dark spots in their central vision. Diabetic retinopathy (DR) affects approximately one third of the people with diabetes, and in advanced stages it also leads to blurry vision and floating spots.
In these and other eye diseases, a major problem is fluid build-up and swelling inside the retina created by leaky blood vessels. Early detection of these fluids is important for successful treatment and management of the diseases.
Optical coherence tomography (OCT) provides 3D cross-sectional images of the retina and visualization of the retinal fluids. However, the areas of fluid build-up can be small and difficult to see without careful examination of each image, and this can be a costly and time-consuming process for clinicians with sometimes several hundreds of cross-sections to examine in each eye.
We have developed a novel framework using deep neural network (DNN) to automatically detect multiple types of fluid regions in OCT images, and identify eyes with specific fluid types. Our framework has achieved the highest level of accuracy for these tasks and won the first place in the 2017 MICCAI RETOUCH competition.
With our promising results, we hope to improve our framework in the next few years by training the DNN with a much larger number of data, including from UK Biobank, and develop a clinical software for ophthalmologists for accurate detection and detailed information of the retinal fluids, such as the shape and volume of the fluid regions. This tool has potential for helping ophthalmologists with early detection of harmful fluid-build ups in the retina, and tracking the disease progress and treatment effectiveness with accurate, quantitative information.