Skip to main content

  • Smartphone-Based ROP Screening

    By Lynda Seminara
    Selected and reviewed by Neil M. Bressler, MD, and Deputy Editors
    Pediatric Ophth/Strabismus, Retina/Vitreous
    Add to My Bookmarks

    Journal Highlights

    JAMA Ophthalmology, June 2023

    Download PDF

    Telemedicine screening programs for retinopathy of prematurity (ROP) are effective, but they require expensive widefield digital fundus imaging (WDFI) and thus may not be accessi­ble in low-income countries. Smart­phone-based fundus imaging (SBFI) systems are inexpensive but have a narrower field of view and lack testing in real-world telemedicine settings. In a study conducted in India, Young et al. looked at the efficacy of ROP screening with SBFI and the accuracy of grading by artificial intelligence (AI) versus humans. They found SBFI systems to be highly sensitive for detecting ROP re­quiring treatment and noted that both types of grading were effective.

    This prospective cross-sectional study was part of a teleophthalmology program in India. From January 2021 to April 2022, premature infants who met normal ROP screening criteria were enrolled. Both eyes of each infant underwent WDFI as well as imaging via one of two smartphone-based oph­thalmoscopy devices (Make-In-India RetCam or Keeler Monocular Indirect Ophthalmoscope). The images were obtained by trained technicians, usually general nurses. For every image, two masked readers evaluated the zone, stage, plus, and vascular severity. The smartphone-obtained images were stratified into three datasets: training (70%), validation (10%), and test (20%). These datasets were used to train a ResNet-18 deep-learning architecture for classification of disease (normal, preplus, or plus), which was applied to derive patient-level predictions of referral-warranted ROP (RW-ROP) and treatment-requiring ROP (TR-ROP). Main outcome measures were 1) the sensitivity and specificity of RW-ROP and TR-ROP detection by human and AI graders and 2) area under the receiver–operating characteristic curve (AUC) for human-graded vascular severity scores, which ranged from one to nine. The sensitivity and specificity of the two SBFI systems were compared using Pearson χ2 testing.

    The study population comprised 156 infants (312 eyes; mean gestational age, 33.0 weeks). There was no significant difference in sensitivity or specificity between the two smartphone systems, so their findings were combined. SBFI was only moderately sensitive for detecting more-than-mild ROP (59%). However, as a screening tool, using the cutoff of at least RW-ROP (i.e., type II ROP and/or preplus or worse), its sensitivity was 100%. SBFI plus human grading detected TR-ROP with sensi­tivity of 100% and specificity of 83.49%. The AUCs for human-graded severity were .95 for RW-ROP and .96 for TR-ROP. With SBFI plus AI grading, sensitivity and specificity were 100% and 58.6% (respectively) for detecting TR-ROP and were 80.0% and 59.3% (respectively) for detecting RW-ROP.

    These findings suggest that the combination of SBFI and AI may be effective for detecting TR-ROP. Use of these low-cost methods would increase the global reach of ROP screening.

    The original article can be found here.