Learning to Segment Corneal Tissue Interfaces in OCT Images

TL;DR

A Convolutional Neural Network (CNN) based framework called CorNet is presented that can accurately segment three corneal interfaces across datasets obtained with different scan settings from different OCT scanners.

Abstract

Accurate and repeatable delineation of corneal tissue interfaces is necessary for surgical planning during anterior segment interventions, such as Keratoplasty. Designing an approach to identify interfaces, which generalizes to datasets acquired from different Optical Coherence Tomographic (OCT) scanners, is paramount. In this paper, we present a Convolutional Neural Network (CNN) based framework called CorNet that can accurately segment three corneal interfaces across datasets obtained with different scan settings from different OCT scanners. Extensive validation of the approach was conducted across all imaged datasets. To the best of our knowledge, this is the first deep learning based approach to segment both anterior and posterior corneal tissue interfaces. Our errors are 2x lower than non-proprietary state-of-the-art corneal tissue interface segmentation algorithms, which include image analysis-based and deep learning approaches.

Figures (5)

• Figure 1: (a)-(b) Original B-scans from a 3$\times$3mm UHR-OCT and 6$\times$6mm SD-OCT volume; (c) Failed Epithelium segmentation result (cyan) from algorithms in LaRocca2011Ge2012Zhang2017; (d)-(e) Our segmentation results for Epithelium (red), Bowman's layer (green), and Endothelium (orange) for images in (a) and (b).
• Figure 2: Our framework takes as input an OCT image, predicts the location of corneal interfaces using the CorNet architecture, and fits curves to the detected interfaces.
• Figure 3: Our network architecture comprises of contracting and expanding branches. The dark green and blue blocks represent downsampling and upsampling computations respectively. Our network makes efficient use of residual and dense connections to generate the corneal interface segmentation in the final image, where each pixel is assigned the label of the tissue it belongs to. The input image is split width-wise into a set of slices of dimensions 256$\times$1024 pixels, the network predicts an output for each slice, and the slices are aligned to recreate the original input dimension. Dense connections concatenate feature maps from previous layers. The light blue block at the bottom of the "U" does not perform upsampling, but it functions as a bottleneck and generates feature maps of the same dimensions as the output feature maps from the previous layer.
• Figure 4: Original B-scans and segmented interfaces from different datasets: (a)-(b) 3$\times$3mm UHR-OCT, (c)-(d) 6$\times$6mm UHR-OCT, and (e)-(h) 6$\times$6mm SD-OCT.
• Figure 5: Error comparison between expert annotation and automated segmentation (fitted with curves) obtained from different deep learning based methods across all 30 testing datasets.