Automated classification of Retinopathy of Prematurity RetCam images using the Apple CreateML platform: gradeable versus ungradeable image classification

Abstract

Purpose : Moorfields Eye Hospital has accumulated over 40,000 retinal images using the RetCam from infants screened for Retinopathy of Prematurity (ROP) over the course of ten years, one of the largest such datasets globally. A programme of research aims to devleop robust, sensitive deep-learning prediction algorithms for interpreting retinal images for the detection of retinopathy of prematurity (ROP) that is non-inferior to current screening methods. We are reporting here the performance characteristics of an automated Machine Learning platform (Apple CreateML) for distinguishing between gradebale and ungradeable RetCam images in ROP.

Methods : 1732 posterior pole RetCam images (952 gradeable and 780 ungradeable) were used to train an Apple CreateML classifier. The Apple CreateML platform enables training and testing Machine Learning models with minimal or no Machine Learning skills required. The validation set consisted of 65 gradeable and 40 ungradeable images. The testing set consisted of 1430 gradeable and 46 ungradeable images.

Results : In the validation set, the precision for detection of gradeable images was 98% and the sensitiovity 100%. In the testing set, where the ratio of gradeable to ungradeable images was representative of real-life condtions, the sensitivity for distinguishing between gradeable and ungradeable RetCam images was 86% and the specificity 83%.

Conclusions : Our model achieved good perfromance in distinguishing between gradeable and ungradeable RetCam images from infants undergoing ROP screening. This automated classification is of clinical relevance as the identification of images of inadequate quality to be graded remotely, either by a human expert or by an automated Decision Support System, could prompt referral of these cases for clinical examination by expert paediatric ophthalmolgists. In this report, we are also demonstrating the potential of an automated ML platform (Apple CreateML) to enable the development of automated screening tools for medical use even when expert Maching Learning skills are not available. Further work will utilise alternative ML approaches and address additional meaningful classifications, such as between referrable and non-refarrable disease.

This is a 2020 ARVO Annual Meeting abstract.