Automatic segmentation of retinal and choroidal thickness in OCT images using convolutional neural networks

Abstract

Purpose : To evaluate the performance of a fully automatic method based on a deep learning approach to segment retinal and choroidal boundaries in OCT images, and derive retinal thickness (RT) and choroidal thickness (ChT) using data obtained from a healthy pediatric cohort.

Methods : Custom designed convolutional neural networks (CNN) were trained to classify three boundaries; the inner limiting membrane (ILM), the retinal pigment epithelium (RPE) and the chorio-scleral interface (CSI). The CNN uses a rectangular path size of 61x31 (VxH) pixels to train the network and three network-input options were tested during training; (i) standard intensity, (ii) attenuation coefficient equivalent and (iii) a combination of both (dual). For each option, the network was trained on the same 137 randomly selected images (70 subjects) and validated on 28 images to ensure adequate training. The network was then tested on 30 different images from 30 different subjects. To test repeatability, consecutive images from the same subject/location were used. The CNN outputs a probability map for each boundary position that was traced with a graph-search technique. The results from the automatic method were compared to data from manual segmentation by an experienced observer.

Results : The well-defined ILM and RPE boundaries showed small errors (1 pixel) in comparison to the CSI which exhibited slightly larger errors (4 pixels) across all tested options (Table 1). The different CNN inputs had a small effect on the boundary error, with the dual input yielding a slightly smaller mean error and SD. Analysis of the ChT and RT, revealed errors of -0.1 and -2 pixels respectively. The mean repeatability difference results (in pixels) for the RT [1.10;1.09;1.08] and ChT [3.40;3.43;3.38] across all options were comparable with the repeatability from manual segmentation [RT 1.24,ChT 2.51].

Conclusions : The automatic and manual segmentation methods showed very close agreement, which suggests that the proposed method provides robust detection of the retinal and choroidal boundaries of interest.

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

 

Table 1. Difference in boundary position and thickness between the proposed algorithm and the manual observer for the data set (n=30) and the different input options.

 

Fig 1. B-scans representing different examples, with typical variations in thickness and contrast. Manual segmentation (red-solid) and automatic using dual input (green-dotted).