Automated 3D Lamina Cribrosa Segmentation in Optical Coherence Tomography Volumetric Scans

Abstract

Purpose : The lamina cribrosa (LC) has been implicated as the primary site of glaucomatous optic nerve damage. Previous histologic studies and imaging studies using optical coherence tomography (OCT) have reported the deformations and displacement of the LC in eyes with glaucoma. Automatic segmentation of the anterior border of the LC would facilitate three-dimensional (3D) morphologic analyses of the LC in a large dataset. However, visibility of the LC in original OCT images are often insufficient for automatic segmentation because of signal attenuation or vascular shadowing. We present an automated 3D segmentation method of the anterior border of the LC using a deep-learning (DL) based image enhancement approach.

Methods : The anterior LC segmentation consists of two main procedures. Firstly, 3D OCT images are processed to enhance the visualization of LC. This consists of a DL-based noise reduction process, a 3D projection artifact reduction process, and a contrast enhancement process. Secondly, the enhanced 3D volumetric data is segmented within the optic disc with a method that is similar to that used in the commercial Topcon OCT. The auto-segmentation results of twenty subjects are validated by experts. B-scan images at the center of the disc are reviewed to check the segmentation error, defined as inappropriate delineation of more than 1/3 of the anterior LC border.

Results : Shown in Fig 1, our noise reduction algorithm improved peak SNR (PSNR) by more than 5 dB. Example segmentation results are shown in Fig 2. Automatic segmentation delineated the anterior LC border in 85% in denoised images, compared to 55% in raw images.

Conclusions : A novel method to enhance LC visualization and automatically segment LC from 3D volumetric scan is presented. For the first time, to our knowledge, LC anterior segmentation was performed on a whole volume in high resolution. The present method has the potential to enable large-scale quantitative research of the LC morphology and the developments of new biomarkers.

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

 

Figure 1. Results of image noise reduction. Left: single B-scan with PSNR = 28.7 dB. Middle: Noise reduced image of the image in the left with PSNR = 34.0 dB. Right: Averaged image of 128 repeated scans at the same location.

 

Figure 2. LC segmentation results (green line) from 6x6 mm 3D scan at ONH. Left: B-scan from superior region. Middle: B-scan crossing ONH center. Right: B-scan from inferior region.