Machine-Learning-Based Retinal Vessel Segmentation in Spectral-Domain Optical Coherence Tomography Volumes of Mice

Abstract

 

Purpose

 

With the increasing use of spectral-domain optical coherence tomography (SD-OCT) in longitudinal studies involving mice, automated approaches for the segmentation of the retinal vessels are critically needed. The purpose of this work is to present and evaluate two novel machine-learning approaches for the automated segmentation of retinal vessels within SD-OCT volumes of mice.

 

Methods

 

Twenty SD-OCT volumes (400 × 400 × 1024 voxels) from the right eyes of 20 C57BL/6J (Jackson Laboratory, Bar Harbor, ME) mice were obtained using a Bioptigen SD-OCT machine (Morrisville, NC). Seven retinal layers in each volume were first segmented using our 3D graph-based algorithm (Antony et al., BOE 2013). Two different approaches were then used to segment the projected locations of the retinal vessels: a baseline approach (similar to the human-based approach presented by Niemeijer et al., SPIE 2008) and an all-layers approach. In the baseline approach, a single projection image was generated by averaging the projected intensity within the retinal-pigment-eptithelium complex and Gaussian-based features were used for k-nearest-neighbor (k-NN) classification. In the all-layers approach, the projection images from all seven layers were generated and Gaussian- and Hessian-based features were used for k-NN classification. Both approaches were evaluated using a leave-four-out cross-validation strategy and a CUDA-based implementation of k-NN was used.

 

Results

 

The receiver operating characteristic (ROC) curves are provided in Figure 1. The area under the curve (AUC) for the all-layers approach, 0.93, was significantly larger than the AUC for the baseline approach, 0.88 (p < 0.05). Example classification results are provided in Figure 2.

 

Conclusions

 

Use of a machine-learned combination of imaging SD-OCT information from multiple 3D retinal layers is more effective for retinal vessel segmentation in mice than use of information from a single layer. Having such an automated retinal vessel segmentation approach will greatly enhance the image-analysis possibilities of longitudinal SD-OCT volumes of mice.

 

 

Figure 1. ROC curves for baseline approach and all-layers approach.

 

 

Figure 2. Results for SD-OCT volume of one mouse. (A) SD-OCT projection image for baseline approach. (B) Set of SD-OCT projection images for all-layers approach. (C) Baseline classification result. (D) All-layers classification result.

 

Keywords: 549 image processing • 688 retina