Skip to main content

  • Artificial Intelligence and Glaucoma Detection

    By Jean Shaw
    Selected By: Henry D. Jampel, MD
    Add to My Bookmarks

    Journal Highlights

    Ophthalmology Glaucoma, July/August 2018

    Download PDF

    Using monoscopic fundus photos, Liu et al. developed a deep learning–based algorithm to detect glaucoma­tous optic discs. They found that their artificial intelligence (AI) algorithm was highly accurate in identifying glaucomatous discs. In addition, they concluded that, as it is relatively easy to obtain monoscopic images, the algo­rithm has potential for use in screening large populations and in telemedicine.

    For this database study, the research­ers obtained fundus photos (n = 3,768) from several previous clinical studies and images from publically available online databases (n = 626), including the High-Resolution Fundus (HRF) database. They then merged the data­bases, with the exception of the HRF database, and divided the images into a training set that comprised 80% of all cases and a testing set that comprised 20% of all cases. The HRF images were used as an additional testing set. Both healthy and glaucomatous eyes were represented in all datasets.

    The researchers tested their AI model and found that its accuracy was 92.7% and that it achieved 89.3% sen­sitivity and 97.1% specificity. When the HRF dataset was used for additional testing, the AI model again was highly accurate and achieved 86.7% in both sensitivity and specificity.

    In order to compare the AI model’s accuracy with the diagnostic skill of experienced clinicians, the researchers randomly selected a series of mono­scopic images and submitted them to a panel of 18 ophthalmologists, which included 11 glaucoma specialists from several countries. They also submitted the HRF images to 3 of the 18 ophthal­mologists for evaluation. The clinicians’ overall accuracy rate was 65%; those who evaluated the HRF images achieved a higher level of accuracy (77%).

    In previous studies, clinician accuracy has been found to be higher when stereoscopic fundus images are used, and the authors noted that stereoscopic images tend to provide better inter- and intraobserver reproducibility. The monoscopic images used in this study varied in terms of quality and resolution, and the testing set included a considerable number of images of anomalous optic discs and photos rep­resenting different disease stages.

    The original article can be found here.