Skip to main content

  • AI Used to Detect Early Keratoconus

    By Lynda Seminara
    Selected by Prem S. Subramanian, MD, PhD
    Add to My Bookmarks

    Journal Highlights

    British Journal of Ophthalmology
    Published online Dec. 16, 2021

    Download PDF

    Previous work by Kundu et al. showed that segmental tomography of the corneal epithelium and stroma may identify early keratoconus (KC) chang­es and that 3D mapping is sensitive for specific changes after different types of refractive surgery. In a newer study, the same authors devised a universal archi­tecture of segmental layered tomogra­phy variables and tested it on eyes with KC and healthy fellow eyes. They found that the artificial intelligence (AI) model could distinguish healthy eyes from those with KC and could be used to further classify those eyes with very asymmetric ectasia (VAE) as subclinical or forme fruste.

    For this work, the authors used data from preoperative images of patients with unilateral or bilateral KC who had refractive surgery and stable outcomes. Layer thickness (epithelium and stro­ma) and posterior surface curvature were measured with high-resolution OCT. Twelve radial B-scans were ob­tained (10-mm maximum scan diam­eter) and were checked for the miss­ing-edge artifact. Curvature, wavefront aberrations, and thickness distributions were analyzed by Zernike polynomials and a random-forest AI model. For training and validation purposes, three groups were defined: healthy (n = 527), KC (n = 454), and VAE (n = 144).

    For healthy eyes, the AI system had area under the curve of 0.994, accuracy of 95.6%, recall of 98.5%, and precision of 92.7%. The respective values for eyes with KC were .997, 99.1%, 98.7%, and 99.1%. For VAE eyes, the results were .976, 95.5%, 71.5%, and 91.2%, respectively. The model reclassified 36 VAE eyes initially deemed subclinical as healthy, even though these eyes were distinctly different from healthy eyes. Most other VAE eyes (forme fruste, n = 104) retained their original classifica­tion. The variables ranked highest for diagnostic value related to the anterior surface; posterior surface features were less useful.

    The original article can be found here.