Deep learning for the identification of plus disease in retinopathy of prematurity

Abstract

Purpose : Retinopathy of Prematurity (ROP) is leading cause of childhood blindness worldwide. In the United States and globally, there is a shortage of adequately trained clinicians who can manage ROP at point of care. Identification of plus disease is key component of the diagnostic process. However, plus disease classification has been shown to be subjective and qualitative. This study explores the use of deep learning to identify infants with clinically significant ROP and to calculate a severity score.

Methods : We compared three deep learning approaches to ROP disease classification. We used two datasets for the experiments. Our smaller, highly characterized dataset of 195 images had manual segmentations of the vasculature, plus disease classification, and disease severity rankings based on pairwise comparisons. Our larger dataset consisted of over 6000 images (multiple views) from 650 patients, including 2400 images of the posterior pole. The larger set has a 3 level classification provided by 3-5 experts.

We utilized a deep learning approach with a U-network architecture to develop a segmentation algorithm. We then extracted quantitative measure of the vascularity from the segmented images and created a random forest classifier for 3 level disease classification (plus, pre-plus, normal). Neighborhood approximate forests were used to create a severity score. Finally, neural networks were developed to perform image classification without segmentation. Data pre-processing and augmentation were employed. We calculated the area under the receiver operating characteristic curve (AUC) for the deep learning algorithm, as well as correlation coefficients (CC) for the algorithm rankings versus pairwise comparison rankings.

Results : A deep learning system for vascular segmentation based on a U-network architecture had an AUC of 0.97. Features extracted from the segmented images are able to classify images into 3 classes and correlated well with the severity ranking (CC 0.78, p<0.05).

Conclusions : A deep learning system for vascular segmentation based on a U-network architecture is extremely effective for classifying plus disease in ROP diagnosis. Deep learning has great potential in automated analysis of ROP.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.