Assessment of AI-enabled screening for ROP in low-resource settings

(Image credit: AdobeStock/ARAMYAN)

Anthony Ortiz, PhD, and colleagues recently reported that the access to screening for retinopathy of prematurity (ROP) can be expanded by using machine learning algorithms developed by using high-quality retinal images collected on a smartphone.¹ Using this advanced technology would lift some of the burden off the few pediatric ophthalmologists that are available in that it could be used by lightly trained personnel who were not pediatric ophthalmologists to screen premature neonates for ROP using a smartphone camera.

The findings were reported in JAMA Network Open.

If successful, this technology could be used to provide screening service to areas with little access to screening, according to Dr. Ortiz, who is from Microsoft AI for Good Lab, Redmond, WA. He was joined in this study by investigators from Business Data Evolution, Mexico City; the Ophthalmology Department, Hospital Italiano de Buenos Aires, Buenos Aires; and the Asociación Para Evitar la Ceguera en México, Mexico City.

Expanding screening capabilities is importance because, with improvements in neonatal intensive care, “more preterm neonates are living and ROP cases have increased worldwide,² particularly in low- and middle-income countries in Africa³ and Latin America,⁴ where no routine screening for ROP occurs,” the investigators explained.

Ortiz and colleagues cited previous studies that have applied artificial intelligence (AI) techniques to identify ROP and classify its severity; however, they pointed out that that work was limited because of its reliance on pediatric ocular camera images, which are not generally available in low-resource settings. While AI algorithms that have been trained on those images were used to identify ROP and classify its severity⁵ even in low- and middle-income populations,⁶ a systematic review of 27 published algorithms found that all but 3 used pediatric ocular camera images⁷; of these three, one used poor-quality images from an unspecified source,⁸ another did not identify how retinal images were obtained,⁹ and the third evaluated the quality of images obtained through the i-ROP study,¹⁰ which used digital wide-field retinal images.¹¹

Study methodology

The authors conducted this diagnostic study in Mexico and Argentina using fundus videos obtained on a smartphone from premature neonates who had and did not have ROP.

The machine learning–driven algorithms were developed to process videos, identify the best frames in the videos, and use those frames to determine whether ROP was or was not likely. Neonates were included who were born at a gestational age less than 36 weeks or had a birth weight below 1,500 grams.

The study goals were to develop and assess the performance of the algorithm to perform retinal screening for ROP in low-resource settings.

The algorithm’s classifications were compared with those of three pediatric ophthalmologists.

ROP screening assessment

The investigators collected 524 videos from 512 neonates. The babies had a median gestational age of 32 weeks (range, 25-36 weeks) and median birth weight of 1,610 grams (range, 580-2,800 grams).

“The frame selection model identified high-quality retinal images from 397 of 456 videos (87.1%; 95% confidence interval [CI], 84.0%-90.1%) reserved for testing model performance. Across all test videos, 97.4% (95% CI, 96.7%-98.1%) of high-quality retinal images selected by the model contained fundus images,” the authors reported.

When evaluated at the frame level, the ROP classifier model had a sensitivity of 76.7% (95% CI, 69.9%-83.5%) and at the patient level, the classifier model had a sensitivity of 93.3% (95% CI, 86.4%-100%), the investigators found.

The combined results surpassed the sensitivities of the pediatric ophthalmologists at both the frame and patient levels, which were 71.4% (95% CI, 64.1%-78.7%) at the frame level and 73.3% (95% CI, 61.0%-85.6%) at the patient level.

The ophthalmologists’ classifications had higher specificity and accuracy compared with the machine-learning model.

In this diagnostic study, the authors concluded, the process had high sensitivity to classify such images as indicating or not indicating ROP but had lower specificity and accuracy than the classification of the pediatric ophthalmologists.

“This usable process,” they said, “may substantially increase screening for ROP in low-resource settings. Because screening is one of the major limitations of receiving ROP treatment, by offloading the process to a less scarce physician type with less training, our process could liberate pediatric ophthalmologists to focus more on activities that only they are qualified to do, such as diagnosis and treatment, and less on conducting screening.”

References

• Ortiz A, Patino S, Torres J, et al. AI-enabled screening for retinopathy of prematurity in low-resource settings. JAMA Netw Open. 2025;8(4):e257831. doi:10.1001/jamanetworkopen.2025.7831

• Nair A, El Ballushi R, Anklesaria BZ, Kamali M, Talat M, Watts T. A review on the incidence and related risk factors of retinopathy of prematurity across various countries. Cureus. 2022;14(11):e32007. doi:10.7759/cureus.32007

• Gilbert C, Malik ANJ, Nahar N, et al. Epidemiology of ROP update—Africa is the new frontier. Semin Perinatol. 2019;43:317-322. doi:10.1053/j.semperi.2019.05.002

• Carrion JZ, Fortes Filho JB, Tartarella MB, Zin A, Jornada ID Jr. Prevalence of retinopathy of prematurity in Latin America. Clin Ophthalmol. 2011;5:1687-1695.

• CampbellJP, Chiang MF, Chen JS, et al; Collaborative Community in Ophthalmic Imaging Executive Committee and the Collaborative Community in Ophthalmic Imaging Retinopathy of Prematurity Workgroup. Artificial intelligence for retinopathy of prematurity: validation of a vascular severity scale against international expert diagnosis. Ophthalmology. 2022;129(7):e69-e76. doi:10.1016/j.ophtha.2022.02.008

• Coyner AS, Oh MA, Shah PK, et al. External validation of a retinopathy of prematurity screening model using artificial intelligence in 3 low- and middle-income populations. JAMA Ophthalmol. 2022;140:791-798. doi:10.1001/jamaophthalmol.2022.2135

• Ramanathan A, Athikarisamy SE, Lam GC. Artificial intelligence for the diagnosis of retinopathy of prematurity: a systematic review of current algorithms. Eye (Lond). 2023;37:2518-2526. doi:10.1038/s41433-022-02366-y

• Worrall DE, Wilson CM, Brostow GJ. Automated retinopathy of prematurity case detection with convolutional neural networks. In: Carneiro G, Mateus D, Peter L, et al. Deep Learning and Data Labeling for Medical Applications. Springer; 2016:68-76. doi:10.1007/978-3-319-46976-8_8

• Wang J, Ju R, Chen Y, et al. Automated retinopathy of prematurity screening using deep neural networks. EBioMedicine. 2018;35:361-368. doi:10.1016/j.ebiom.2018.08.033

• Coyner AS, Swan R, Brown JM, et al. Deep learning for image quality assessment of fundus images in retinopathy of prematurity. AMIA Annu Symp Proc. 2018;2018:1224-1232.

• Open-label, randomized, 2-arm, controlled study to assess the efficacy, safety, and tolerability of intravitreal (IVT) aflibercept compared to laser photocoagulation in patients with retinopathy of prematurity (ROP). ClinicalTrials.gov identifier: NCT04004208. June 23, 2020. Accessed March 29, 2024.