Purchase this article with an account.
Miguel Angel Zapata, Eduardo Ulises Moya-Sanchez, Jonatan Moreno, Dario Garcia-Gasulla, Ferran Parés, Didac Royo, Armand Vilalta, Ulises Cortés, Eduard Ayguadé, Jesus Labarta; Artificial intelligence for evaluating retinographies: beyond the diabetic retinopathy. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1708.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Automatic image classification systems have recently made a significant leap in performance thanks to the application of Convolutional Neural Networks (CNNs), they are able to achieve human-level on image recognition tasks. Researchers have focused on diabetic retinopathy with significant success. We follow and extend this research line, widening the scope to assess normality/abnormality, and detection of five different entities: Increase cup/disc ratio, nevi, age related macular degeneration (AMD), epiretinal membrane and pigmented abnormalities in macula. Our main goal is to explore the limits of CNNs for retinographies, and to generate relevant tools in the process
Originates from a database of 300K photographs labeled by ophthalmologists, collaboration between medical experts from Optretina and researchers from High Performance Artificial Intelligence group at Barcelona Supercomputing Center. Details of database are shown in Table 1. We have designed and trained six different CNNs, considering the particularities of each task separately. For evaluating normality/abnormality, the available set of images was larger, which allows us to use a deep model which perceives the image at different levels of resolution (128x128, 256x256 and 512x512). For the pathology detectors the available datasets were smaller, we used a different approach. We consider only images at 512x512 resolution, but we use of the rest of available images to pre-train the network. This will allow us to start the detector training from a network status where basic retinography characteristics are already known
Table 2 shows results separated by task. Accuracy of the CNN indicating normality or not normality of any retinography was 70 %. Other accuracies were: cup/disc ratio 85.5%, AMD 91.07%. Nevi 64.64%, epiretinal membrane 76.17%, macular pigmented abnormalities 63.03%.
Results indicate that CNNs could determinate if we are in front of a normal retinography or not with a medium level of certainty. Beyond diabetic retinopathy, CNN’s can be helpful for screenings of other entities as suspicion of glaucoma or AMD, also interesting for epiretinal membrane or nevi. We expect the models to reach an even higher level of reliability in the short-term, which will soon allow for their deployment as decision support tools for experts
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Images for training, test and validation
This PDF is available to Subscribers Only