An international team of researchers recently expanded our understanding of deep learning and its ability to not only detect but also grade age-related macular degeneration (AMD) fundus photographs. While research has already proven the technology useful in identifying whether patients have intermediate or advanced AMD, this team sought to better understand deep learning’s ability to grade AMD fundus photographs based on the classic AREDS nine-step detailed scale, as well as a simplified four-step scale designed for the study.

Using  67,401 color fundus images from the 4,613 AREDS participants, the researchers tested the deep learning tool’s accuracy with two AMD severity classification criteria, one with four steps and the classic AREDS nine-step AMD severity scales.

When comparing artificial and human intelligence in grading of fundus photos, they found good strength of agreement between the two. Inter-rater agreement was 0.77 for the four-step AMD scale and 0.74 for the nine-step AMD severity scale, just shy of the 0.81 threshold for “very good” strength of agreement on this scale. They also found the overall mean estimation error—i.e., the likely difference between the sample mean and that of the population studied—for the five-year risk ranged from 3.5% to 5.3%.

The researchers note this performance is comparable with that of the human graders for both the four-step scale and is promising for providing severity grading with the nine-step AREDS classification and for estimating the five-year risk of progression to advanced AMD.

Using deep learning “has the potential to assist physicians in longitudinal care for individualized, detailed risk assessment as well as clinical studies of disease progression during treatment or as public screening or monitoring worldwide,” the study concludes.