Deep Learning for Image Quality Assessment of Fundus Images in Retinopathy of Prematurity

Abstract

Accurate image-based medical diagnosis relies upon adequate image quality and clarity. This has important implications for clinical diagnosis, and for emerging methods such as telemedicine and computer-based image analysis. In this study, we trained a convolutional neural network (CNN) to automatically assess the quality of retinal fundus images in a representative ophthalmic disease, retinopathy of prematurity (ROP). 6,043 wide-angle fundus images were collected from preterm infants during routine ROP screening examinations. Images were assessed by clinical experts for quality regarding ability to diagnose ROP accurately, and were labeled "acceptable" or "not acceptable." The CNN training, validation and test sets consisted of 2,770 images, 200 images, and 3,073 images, respectively. Test set accuracy was 89.1%, with area under the receiver operating curve equal to 0.964, and area under the precision-recall curve equal to 0.966. Taken together, our CNN shows promise as a useful prescreening method for telemedicine and computer-based image analysis applications. We feel this methodology is generalizable to all clinical domains involving image-based diagnosis.

Figure 1.

Representative images from the (A) Not acceptable for diagnosis of ROP, (B) Possibly acceptable for diagnosis of ROP, and (C) Acceptable for diagnosis of ROP image sets.

Figure 2.

Receiver operating characteristics (ROC) curve for model evaluation on the test set data displayed in the left panel. Area under the ROC curve (AUROC) is ~0.96. Precision-recall curve for model evaluation on the test set data is shown in the right panel. Area under the precision-recall curve (AUPR) is ~0.97. Red dashed lines indicate expected results of a naïve algorithm.

Figure 4.

Scatterplot of experts’ consensus rank versus CNN rank. Each point represents an image and its corresponding expert consensus rank (y-axis) and CNN rank (x-axis). Red line indicates best-fit line. Confidence interval is represented by dark gray band. Summary statistics are displayed within the plot.