Loosely coupled level sets for 3D retinal layer segmentation in optical coherence tomography scans of healthy and glaucoma subjects

Abstract

 

Purpose

 

To present and evaluate a novel level set (LS) approach for retinal layer segmentation in peripapillary OCT volumes. The accuracy and reproducibility were evaluated on normal and glaucomatous eyes

 

Methods

 

Interfaces between layers were detected simultaneously by using a novel LS coupling approach on attenuation coefficients (AC) derived from OCT volume scans. Coupling of the interfaces exploits the order of the layers in the retina and thus ensures anatomically correct segmentation results. This anatomical knowledge was incorporated in a probabilistic framework and combined with image data and prior knowledge on the layers’ AC and thicknesses. To assess the accuracy, one eye of 20 healthy subjects and 10 glaucoma patients was imaged with a Spectralis OCT system (Heidelberg Engineering, Germany). 9 healthy subjects were imaged again after a few minutes to evaluate the reproducibility. The scan protocol combined 193 B-scans composed of 512 A-lines of 496 pixels into one volume, with 5 times averaging. One randomly selected B-scan of every volume was manually segmented by a medical doctor. Six interfaces were considered: vitreous-RNFL, RNFL-GCL, IPL-INL, INL-OPL, OPL-ONL and the IS ellipsoid’s posterior boundary. Differences between segmentations were expressed as the root mean square error (RMSE) and mean absolute deviation (MAD).

 

Results

 

Examples of the segmentation on a B-scan of a normal and glaucoma eye are shown in Fig. 1. Evaluation results are listed in Tab. 1. The MAD for accuracy of the segmentation varied between 3.3-7.6 µm and 3.5-9.7 µm for healthy and glaucoma subjects, respectively. The reproducibility of manual annotations showed an MAD of 3.1-7.9 µm, while the automatic method (AM) showed an MAD of 2.5-4.6 µm.

 

Conclusions

 

A novel 3D segmentation method, which simultaneously detects multiple interfaces via a new coupling approach, was presented. The AM achieved a high accuracy in segmenting the volumes of healthy and glaucoma subjects. The reproducibility of the AM is better than manual annotations. The accuracy of the AM and the intra-observer reproducibility are similar suggesting that the accuracy of the AM is as good as the manual annotations.

 

 

Tab. 1. Accuracy and reproducibility evaluation results

 

 

Fig. 1. Segmentation results and resulting thickness maps for a normal and glaucoma eye. The A-lines passing through the vasculature were excluded

 

Keywords: 549 image processing • 551 imaging/image analysis: non-clinical • 552 imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)