Development of AI Deep Learning Algorithms for the Quantification of Retinopathy of Prematurity

Abstract

Purpose : Retinopathy of prematurity (ROP) remains a leading cause of childhood blindness worldwide. Recently, handheld spectral-domain optical coherence tomography (SD-OCT) has become more widely used in neonates because it allows for non-invasive, quick in vivo imaging of the retinal structures offering more detail than clinical examination. However, objective quantification of SD-OCT images are difficult because they require time-consuming human expert interpretation. We sought to create a fully automated algorithm to identify and quantify the ROP clinical features in SD-OCT images from neonates.

Methods : 4,687 SD-OCT images from 54 neonates were collected. A fully automated pipeline was created 1) to segment the border between the vitreous and internal limiting membrane and 2) detection of vitreous punctate hyperreflective opacities (VHO) using deep learning and Gaussian convolutions, respectively. The data was partitioned by neonate into training, validation, and test sets. A total of 4 independent expert graders created the ground truth for each image in the test set, which were used to validate the models. We tested the performance of the automated pipeline using Dice coefficient and F1 score.

Results : The mean Dice coefficients for the human interrater reliability and the models for the vitreous border segmentation were 0.95 (SD=0.12) and 0.90 (SD=0.21), respectively (Wilcoxon p<0.05). The mean F1 scores for the human interrater reliability and the models for VHO detection were 0.75 (SD=0.25) and 0.68 (SD=0.33), respectively (Wilcoxon p=0.60).

Conclusions : Deep learning models can achieve high accuracy in identifying and detecting certain clinical features of ROP. Features with subjective manual quantification such as VHO may eventually be screened using fully automated algorithms. Further studies are needed to extend the pipeline to include other clinical features such as cystoid macular edema and vitreous bands.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Figure 1. Deep learning model schematic. Modified U-Net architecture for vitreous and internal limiting membrane segmentation.

 

Figure 2. Example segmentations of the vitreous border (top) and punctate hyperreflective opacities (bottom). The left column shows the original image, the middle column shows expert segmentation, and the far right column shows segmentation by the models. Insets are zoomed images of opacities.