Using Novel Fundus Image Preprocessing to Improve the Classification of Retinopathy of Prematurity (ROP) Using Deep Learning

Abstract

Retinopathy of Prematurity (ROP) can affect babies born prematurely. It is a potentially blinding eye disorder which can arise from the complications of undeveloped retina. Thus, screening of ROP is essential for early detection and treatment in these infants. An Retcam digital camera is used to capture patient’s retinal image. The captured image quality is constrained due to many factors. Effective and accurate image pre-processing methods for ROP Retcam images are required for improving ROP clinical features prior to being used in any CNN based classification system.

We reviewed present literature on image pre-processing pertaining to digital retina images. This included image domain, restoration based methods and the latest machine learning methods. Our first contributions were two improved novel restoration image pre-processing methods. An image domain method was then applied to the output of these improved methods to create new hybrid methods. These new pre-processing methods improved ROP clinical features in ROP Retcam images. The third contribution used a novel approach using deep learning-based segmentation classifier to generate vessel map from an ROP Retcam image. The purpose was to erode the blood vessels from the original image thereby reducing the blood vessel noise. For our fourth contribution, we used transfer learning based CNNs, namely, InceptionResv2, and ResNet50 to create 3 sets of classifiers representing each ROP condition, namely Plus Disease, Stages and Zones. These CNNs were trained and validated using the improved pre-processing methods and traditional methods independently. The comparative evaluations of all identified pre-processing methods showed that these new pre-processing methods contributed to higher accuracy when classifying ROP using limited training images. With these methods, our results were as equal or better than comparative peer results using limited data. In this research, using the above components, we created a framework, McROP, that deals with key three ROP conditions. This framework can be extended easily to other pediatric ophthalmology conditions. To our knowledge, this is the first known use of restoration-based image pre-processing for ROP Retcam for improving ROP clinical features. These methods demonstrated effectiveness in CNNs based classification for ROP when compared against traditional pre-processing methods.