Automated layer segmentation of macular OCT images using dual-scale gradient information

Abstract

A novel automated boundary segmentation algorithm is proposed for fast and reliable quantification of nine intra-retinal boundaries in optical coherence tomography (OCT) images. The algorithm employs a two-step segmentation schema based on gradient information in dual scales, utilizing local and complementary global gradient information simultaneously. A shortest path search is applied to optimize the edge selection. The segmentation algorithm was validated with independent manual segmentation and a reproducibility study. It demonstrates high accuracy and reproducibility in segmenting normal 3D OCT volumes. The execution time is about 16 seconds per volume (480x512x128 voxels). The algorithm shows potential for quantifying images from diseased retinas as well.

Fig. 1

Illustration of nine intra-retinal boundaries from top to bottom: boundary 1 ILM, boundary 2 NFL/GCL, boundary 3 GCL/IPL, boundary 4 IPL/INL, boundary 5 INL/OPL, boundary 6 ELM, boundary 7 IS/OS, boundary 8 OS/RPE and boundary 9 BM/Choroid.

Fig. 2

Illustration of the complementary gradient information used for segmentation of the IPL/INL boundary. The sigma of the Gaussian kernel used in the Canny edge detector was 2, while the sigma of Gaussian kernel used in the axial gradient calculation was 4. The thresholds of the Canny edge detector were [0.1 0.55 0.8].

Fig. 3

Illustration of nine boundary segmentation flow. Panel a is the original OCT image acquired using Topcon 3D OCT-1000 equipment. The image is first aligned as shown in b and the ILM and IS/OS are detected as in c; d, e and f illustrate the BM/Choroid, OS/RPE, IPL/INL, NFL/GCL, GCL/IPL, INL/OPL and ELM are detected in order; in the end, all nine boundaries are converted back to the original OCT image coordinates as shown in g.

Fig. 4

Results of a normal 3D vertical scan volume. NFL, GCC (from ILM to IPL/INL) and total retinal thickness maps (from ILM to OS/RPE) covered 5×5mm² area, in which three B-scans (the 32nd, 64th and 96th images) of this volume were shown. T: temporal; I: inferior; N: nasal; S: superior.

Fig. 5

Comparison between manual segmentation (average of four segmenters’ results, yellow lines) and algorithm segmentation (blue lines). The blue arrows illustrate the commonly seen NFL/GCL discrepancy between human segmenters and the automated algorithm. The red arrows indicate the commonly seen IPL/INL difference between human segmenters and the automated algorithm around the fovea.

Fig. 6

Illustration of segmentations on images with poor quality (panel a) and blood vessel artifacts (panel b).

Fig. 7

Results of a 3D horizontal scan volume with drusen. The total thickness map (from ILM to OS/RPE) and the RPE complex thickness map (from OS/RPE to BM/Choroid) cover 5×5mm² area, in which two B-scans (the 20th and 40th images) of this volume are shown. T: temporal; I: inferior; N: nasal.

Fig. 8

Results of a 3D vertical scan volume from a glaucoma patient. NFL and total retinal thickness maps (from ILM to OS/RPE) cover 5×5mm² area, in which two sample B-scans (the 15th and 47th images) from the volume are shown. T: temporal; I: inferior; N: nasal; S: superior.