Artificial intelligence in retinopathy of prematurity: identification of clinically significant retinal vascular findings using computer-based image analysis

Abstract

Purpose : Plus disease is the key parameter for identifying infants at risk for blindness from ROP, and is defined as arterial tortuosity and venous dilation in the posterior pole. Studies have shown that: (1) diagnosis is subjective and variable, (2) experts often consider vascular features beyond those in the definition and are often unable to explain their diagnostic process, and (3) computer-based analysis can diagnose plus disease with comparable or better accuracy than experts. This study uses outputs from image analysis systems to identify vascular features considered most significant by experts for plus disease diagnosis.

Methods : 31 wide-angle retinal images with plus disease, based on a consensus reference standard, were analyzed using (a) a deep learning system (i-ROP DL) performing occlusion analysis with a convolutional neural network to compute the significance of each 12x12 pixel region to the network’s diagnostic ability, visualized as a heatmap (Figure) from which vascular features were extracted that most affected the probability of diagnosis; (b) a supervised machine learning (SML) system (i-ROP ASSIST) that quantified 50 different retinal vascular features, from which the ones most closely associated with the reference standard diagnosis were identified.

Results : Retinal features identified as important for plus disease diagnosis based on deep learning occlusion analysis were: (1) central retinal vessel location (31/31 images), (2) mid-peripheral vessel location (25/31 images), (3) arterial tortuosity (31/31 images), (4) venous dilation (31/31 images), (5) arterial dilation (31/31), (6) venous tortuosity (31/31). SML feature output identified the following features as most important: (1) cumulative tortuosity index (30/31), (2) acceleration or rate of change of direction of vessel (28/31), and (3) average point diameter (23/31).

Conclusions : Backward reasoning from outputs of computer-based image analysis systems shows that experts consider numerous vascular features during plus disease diagnosis, including dilation and tortuosity of both arteries and veins in all fields of view. This has important implications for clinical care and education in ROP, and this methodology may be applied to other ophthalmic diseases.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Heatmap identifying key regions for plus disease diagnosis using deep learning occlusion analysis.