Researchers at Washington University School of Medicine in St. Louis, working with colleagues at the University of Washington in Seattle and Genentech, developed an experimental artificial intelligence system designed to help physicians review retinal scans more quickly and identify signs of disease earlier.
The technology, called OCTCube-M, is built to analyze three-dimensional retinal images and other eye scans commonly used in clinical practice. The work addresses a practical challenge in ophthalmology. Optical coherence tomography scans generate hundreds of images per exam, and physicians typically must review them manually. That process can take time and may be vulnerable to human error.
In a recent study published in Nature Biomedical Engineering, the researchers reported that the system outperformed older models in identifying several retinal diseases and in predicting progression of geographic atrophy, a severe form of age-related macular degeneration.
Methodology
The researchers developed OCTCube-M as a family of three AI models designed to interpret 3D retinal images and additional eye imaging data.
To train the system, the team used more than 26,000 3D optical coherence tomography images, representing 1.62 million individual retinal slices. The researchers said they wanted to test whether training on 3D images, rather than only two-dimensional tomography images, would improve diagnosis and prognosis. They focused on 3D data because retinal disease often extends around the fovea in all three dimensions.
After building the 3D model, the researchers adapted it by adding data from two additional imaging techniques: infrared retinal imaging and fundus autofluorescence imaging. They then evaluated how the models performed when using optical coherence tomography alone and in combination with one or both of the other imaging types.
The study assessed the model’s ability to identify eight retinal diseases that affect the retina and optic nerve. The researchers also examined how well the system could predict the growth rate of geographic atrophy. In addition, the study evaluated whether the model could infer risks beyond eye disease, including outcomes such as heart attack, stroke and kidney failure, using retinal imaging alone.
Results
Compared with a model trained on two-dimensional images, OCTCube-M more accurately identified six of the eight retinal diseases by about four to six percentage points. According to the researchers, that would translate to 43 to 60 additional cases detected for every 1,000 people with eye disease. The study said this pattern held across scans from multiple clinical sites, imaging modalities and patient populations.
The system also improved prediction of disease progression in geographic atrophy. The version of the model trained on all three imaging types—optical coherence tomography, infrared retinal imaging and fundus autofluorescence imaging— outperformed the current state-of-the-art model that relies only on fundus autofluorescence images by an average of nearly 50%.
The researchers also reported that the model could predict outcomes beyond the eye, including heart attack, stroke and kidney failure, based only on retinal imaging. The study noted that retinal blood vessels share anatomical and developmental similarities with vessels in the kidney. Vascular changes associated with heart and brain disease may also leave detectable signs in the eye.
What this means for eyecare providers
For eyecare providers, the findings suggest that AI tools such as OCTCube-M may help manage the large volume of imaging data generated by routine retinal scans. A system that reviews scans more efficiently could support earlier identification of subtle disease features and provide another layer of analysis during image interpretation.
The study also suggests that combining multiple imaging types may improve assessment of diseases such as geographic atrophy. For clinicians who care for patients with retinal disease, more accurate estimates of how quickly a condition may progress could help inform monitoring and treatment planning.
The findings may also be relevant to research settings. The authors said better predictions of disease progression could support smaller and more efficient clinical trials by improving how patients are classified and monitored over time.
The technology remains in the research stage. The researchers said their next step is to train OCTCube-M on larger datasets that include more patients, more diseases and additional imaging types.
Read more news from AI in Eye Care here
Author
