Risk assessment in retinopathy of prematurity: improvement of clinical models using automated image analysis

Abstract

Purpose : ROP risk is traditionally predicted using clinical parameters such as birth weight, gestational age, and systemic illness. We propose that risk models for predicting treatment-requiring (Type 1) ROP may be improved with the inclusion of image severity scores derived from a deep learning algorithm.

Methods : A multi-institutional dataset of posterior retinal photographs was collected as part of the ongoing “Imaging and Informatics in ROP” (i-ROP) study. A reference standard diagnosis (RSD) for treatment-requiring ROP was assigned to each image using previously published methods. An i-ROP image severity score for plus disease (i-ROP-SS) was automatically computed using a deep convolutional neural network, which grades images on a scale from 1.0 - 9.0. The change in severity between first and second sessions (Δi-ROP-SS) was also calculated, per subject. Clinical parameters included birth weight, gestational age, target oxygen saturation, chronic lung disease, intraventricular hemorrhage grade, and sepsis. Five logistic regression models were constructed and compared to predict whether an infant would require treatment: (1) clinical parameters only, (2) clinical parameters, i-ROP-SS at baseline, (3) clinical parameters, i-ROP-SS at baseline, Δi-ROP-SS at follow-up, (4) i-ROP-SS at baseline, Δi-ROP-SS at follow-up, (5) birth weight, gestational age, i-ROP-SS at baseline, Δi-ROP-SS at follow-up. All models were evaluated using five-fold stratified cross-validation.

Results : Mean ± standard deviation areas under the receiver operating curve (AUCs) were (1) 0.85±0.05, (2) 0.89±0.06, (3) 0.93±0.05, (4) 0.84±0.15 and (5) 0.91±0.06. Both the mean sensitivity and specificity of risk model (3) were 89%. The high specificity is favorable compared with previous models such WINROP (95.7% specificity, 24% sensitivity), CO-ROP (96.4% sensitivity, 33.7% specificity), CHOP-ROP (98% sensitivity, 53% specificity).

Conclusions : Inclusion of a deep learning-based image severity score in risk models improves their power in predicting treatment-requiring ROP. This may have implications for clinical ROP management.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Table summarizing data and AUCs in each of the logistic regression risk models.

 

Areas under the ROC curve (mean + standard deviation) for risk models (1) clinical only, and (3) clinical parameters, i-ROP-SS at baseline, Δi-ROP-SS at follow-up.