Skip to main content

  • Using AI to Grade Cataract Severity

    By Jean Shaw
    Selected by Emily Y. Chew, MD
    Add to My Bookmarks

    Journal Highlights

    Ophthalmology Science, June 2022

    Download PDF

    Young Son et al. aimed to develop and validate an automated platform for diagnosing and grading cataracts using slit-lamp and retroillumination lens photographs. They found that their deep learning (DL)–based artificial in­telligence (AI) platform was able to ac­curately and precisely detect and grade nuclear opalescence (NO), nuclear color (NC), cortical opacity (CO), and posterior subcapsular opacity (PSC).

    For this cross-sectional study, the researchers evaluated 596 patients (887 eyes) who were 14 to 94 years old and had nuclear sclerotic, cortical, or poste­rior subcapsular cataracts. The images were graded by two trained graders, using the Lens Opacities Classification System (LOCS) III, and were divided into training, validation, and test datasets. Main outcome measures were diagnostic and LOCS III–based grading prediction performance.

    A total of 1,335 slit-lamp and 637 retroillumination lens images were obtained. Of the slit-lamp images, 918 were in the training dataset, 152 were in the validation dataset, and 265 were in the test dataset. For the retroillumi­nation lens images, these numbers were 435, 71, and 131, respectively.

    For nuclear sclerotic cataracts, the AI system’s diagnostic performance of NO and NC showed excellent results—area under the curve (AUC) of .9992 and .9994 and accuracy of 98.82% and 98.51%, respectively. Excellent results also were achieved for cortical and posterior subcapsular cataracts: the AUC for CO and PSC was .9680 and .9465, while accuracy was 96.21% and 92.17%, respectively.

    All told, the results show the plat­form’s potential for accurate detection and grading in LOCS III 6- and 7-level classifications for all types of cataracts, the authors said.

    The original article can be found here.