Approved research

Developing Novel, Integrated, Interpretable Deep Learning Systems for Ophthalmology

Lay summary

There are many unmet needs in ophthalmology that lead to preventable blindness. Across the world, there is an increase in individuals needing proactive access to eye care to prevent blindness or to manage existing conditions. This increase has outpaced the number of trained doctors who can help screen, diagnose, and monitor eye disease, resulting in an access issue. AI can help improve access to support the early identification of disease. Where humans are unable to make time to look at all the images for patients, AI can help examine or diagnose a larger volume and ask for human help only in unclear cases. Furthermore, where ophthalmologists are able to see patients, AI can assist them by identifying and surfacing relevant information quickly. These approaches have not yet been widely implemented because current AI systems do not integrate the broad information used by humans to make decisions. Instead, most current systems use a single imaging modality as the input. Additionally, these systems do not yet do a good enough job of explaining how they reached their conclusion for each patient. Our goal is to engineer better AI systems that use all of the available ophthalmological imaging modalities to make diagnoses and provide high-quality explanations for their decisions. As we make progress towards our goal, we believe that more people will be able to be screened and diagnosed for their eye conditions sooner, which will ensure that they receive the treatment they need before their health suffers. Specifically, conditions like AMD, glaucoma, and diabetic retinopathy will see major reductions in the number of patients adversely impacted. Our project is forward-looking, so we anticipate a three year timeline to fully take advantage of rapid innovations in computer vision.