Australian researchers recently tested an artificial intelligence (AI) system built to detect diabetic retinopathy. The researchers concluded that the system did not perform well in real-world clinical settings. Of the nearly 200 patients evaluated, the system marked 17 patients as having severe diabetic retinopathy that required a referral. While the system correctly judged two patients with the disease, it also generated 15 false positives. These false positives were driven by dirty lens reflections, uneven light exposure, or drusen that appeared similar to exudates.
Despite these limitations, researchers remain hopeful in the future use of the AI system for improving screening efficiency for diabetic retinopathy. Further training of the system to differentiate exudates, sheen reflections, and drusen may improve results.
These findings were published in JAMA Network Open on September 28, 2018.