Evaluation of artificial intelligence-based quantitative analysis to identify clinically significant severe retinopathy of prematurity.

PURPOSE To evaluate the screening potential of a deep learning algorithm derived severity score by determining its ability to detect clinically significant severe retinopathy of prematurity (ROP). METHODS Fundus photographs were collected, and standard panel diagnosis was generated for each examination by combining 3 independent image-based gradings. All images were analyzed using a deep learning algorithm and a quantitative assessment of retinal vascular abnormality (DeepROP score) were assigned on a 1-100 scale. The area under the receiver operating curve (AUROC) and distribution pattern of all diagnostic parameters and categories of ROP were analyzed. The correlation between the DeepROP score and expert rank ordering according to overall ROP severity of 50 examinations was calculated. RESULTS A total of 9882 individual examinations with 54626 images from 2801 infants were analyzed. 56 (0.6%) examinations demonstrated type 1 ROP, 54 (0.5%) examinations demonstrated type 2 ROP. The DeepROP score had an AUROC of 0.981 for detecting type 1 ROP and 0.986 for type 2 ROP. There was a statistically significant correlation between the expert rank ordering of overall disease severity and the DeepROP score (correlation coefficient 0.758, p<0.001). When hypothetical referral cut-off score of 35 was selected, all cases of severe ROP (type 1 and type 2 ROP) was captured, 8562 (87.6%) eyes with no or mild ROP were excluded. CONCLUSION The DeepROP score determined by deep learning algorithm was an objective and quantitative indicator for the severity of ROP, and it had potential in automated detecting clinically significant severe ROP.