Machine Learning for Prediction of Retinopathy of Prematurity Fundus Image Quality from Clinical Data

Abstract

Purpose : Telemedicine and artificial intelligence have shown great promise for improving management of ophthalmic diseases such as retinopathy of prematurity (ROP). A significant challenge in their implementation is the need for high-quality images. We have previously trained a convolutional neural network to detect Acceptable Quality (AQ) images from those that are not. However, it remains unclear as to what degree image quality is related to clinical factors versus imaging technique. The aim of this study is to identify what clinical factors contribute to image quality and to predict image quality with them.

Methods : We reviewed 4,030 examinations from the Imaging and Informatics in ROP (i-ROP) cohort study, in which images were obtained (Retcam, Natus Medical Systems, Pleasanton, CA) during routine ROP screenings. Six experts assessed image quality, voting AQ, Possibly AQ (PAQ) or Not AQ (NAQ). Each expert voted on each image, and a majority vote was used for the image quality label. The data set was split into separate training (75%) and test sets (25%). A random forest classifier (RF) was trained, via 5-fold cross-validation, to classify AQ images from non-AQ images. Variables with a mean decrease in Gini greater than 0.5 were considered significant. To confirm variable importance, logistic regression was performed, and variables with a p-value less than 0.05 were considered significant. A new RF was trained using only the significant predictors identified by both models and its performance was evaluated on the test set.

Results : The variables post-menstrual age, gestational age, race, site (different imaging technician at each site), and session (number of times a baby was imaged) were identified as significant predictors of image quality by both models (Figure 1). Using only these predictors, a RF classified AQ images from non-AQ images with an area under the receiver operating characteristics curve (AUC) of 0.875 on a test set, with a sensitivity and specificity of 86% and 76%, respectively (Figure 2).

Conclusions : Clinical and demographic factors can be used to predict retinal fundus image quality in ROP better than chance. This suggests that poor image quality is partially dependent upon the clinical demographics of the subject being imaged.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Descriptive plots of clinical factors found to be important for image quality.

 

Evaluation of final RF model on the test set via ROC curve. AUC = 0.875.