Complications from diabetes (diabetic retinopathy) is one of the leading causes of VI. The total number of individuals with diabetes is expected to double between 2000 and 2030, greatly increasing the number of VI cases worldwide.
When diagnosed early, severe vision loss can be reduced by up to 57%. Unfortunately, there has been little improvement in the screening process to discover retinal diseases in the early stages. In the countries that are affected the most, patients are not receiving regular health screenings. Ophthalmologists have a fairly low accuracy rate in properly detecting retinal disease through in-person dilated eye examinations. With other life-threatening diseases at the forefront of our mind, retinal disease and visual impairment have not been as prominent in the public eye.
An Eye on Artificial Intelligence
AI may hold the key to significantly reducing the impact of retinal diseases by helping ophthalmologists detect disease more effectively, augmenting the human experience. Through a collaboration with Lenovo, the Barcelona Supercomputing Center (BSC) has set out to explore how AI can improve the accuracy of the screening process and potentially detect a retinal disease sooner. AI technology further increases the likelihood of early detection by putting the power of screening in the hands of patients in underserved populations, allowing them to self-administer an initial screening in a matter of minutes, using a smartphone.
In addition to diabetic retinopathy, many other pathologies cause VI such as glaucoma, macular degeneration, nevus, and epiretinal membrane. Machine learning models make identifying these various pathologies much easier than the current screening methods. Dario Garcia-Gasulla, postdoctoral researcher at the Barcelona Supercomputing Center, is optimistic about how this technology can be used. Garcia-Gasulla explains, “Scaling the design, training and validation of machine learning models to study these vision issues may be challenging. But the impact of potential solutions for this is huge, as the same challenge is found within other medical domains and many other industrial applications.”
Transfer Learning Overcomes Data Challenges
The issue with training an AI model to detect certain retinal diseases lies in the lack of clean data available to train an AI neural network. For pathologies with limited dataset availability (e.g., < 5,000 images), training a reliable deep neural network from scratch may not be feasible. Transfer learning is based on models trained for problems with larger datasets, which are then reused to solve other problems with little data availability. When it is used as a feature extractor, transfer learning can also reduce the training time (to minutes), research hours, and ultimately, the cost associated with the development of a solution.
Lenovo Announces New AI Technology
At the International Supercomputing Conference (ISC) in Frankfurt, Germany, Lenovo and BSC will demo a Transfer Learning Application from the Lenovo AI Innovation Center in Morrisville, North Carolina (USA). The application allows visitors at Lenovo’s booth to compete through an intuitive gamified interface, allowing anyone to take an active role in improving retinal disease screening.
Garcia-Gasulla explains, “The goal of the demo is to show how easy it is to use pre-trained deep neural networks as feature extractors which feed other simpler and faster models (in this case, SVMs). Within ten minutes, each participant will be able to design, train and validate the performance of a machine learning model to detect a retina pathology. Participants working on the same pathology will be compared, to find and award the best model designed during the whole conference.”