Skip to main content
  • Pediatric Ophth/Strabismus, Retina/Vitreous

    Review of: Evaluation of a deep learning-derived quantitative retinopathy of prematurity severity scale

    Campbell J, Kim S, Brown J, et al. Ophthalmology, July 2021

    Using clinical examination images, investigators reviewed and evaluated the feasibility and reproducibility of a quantitative vascular severity scale for assessing retinopathy of prematurity (ROP) cases.

    Study design

    The investigators evaluated the relationship between a deep-learning quantitative (1–9) vascular severity scale for ROP, developed by the Imaging and Informatics in ROP Consortium, with stage of disease (stages 1-3), extent of stage 3, and zone. The investigators also evaluated whether this algorithm could easily be used by trained clinical observers. De-identified clinical examination images (N = 6344) with 5 standard fields were used.


    A higher vascular severity score, as determined by both clinical observers and the deep-learning algorithm, was associated with increasing stage of disease in zones I–III and a greater number of clock hours in stage 3. Multivariate regression analysis found that stage, extent, and zone were all independently associated with the 1–9 score, though there was significant overlap in the distributions for individual eyes. There was moderate-to-high agreement among trained clinical graders for both the absolute and relative 1–9 scores.


    To fully validate the proposed scale, further testing by clinicians outside the current study group and outside of North America is needed. Additional validation with other datasets using different fundus photography platforms is required for the deep-learning algorithm.

    Clinical significance

    The proposed quantitative scale for ROP is feasible for clinical adoption and may reduce interobserver variability between clinicians, allow for better tracking of disease progression, and improve consistency between clinical trials. Furthermore, the proposed quantitative scale lends itself to using technology for more consistent and improved ROP diagnosis.