Comparison of coupled level sets and graph cuts for retinal layer segmentation in optical coherence tomography

Abstract

 

Purpose

 

The accuracy and reproducibility of two retinal layer segmentation methods were evaluated. Graph cuts (GC) and a novel level set (LS) approach were applied to peripapillary OCT images, which were first converted to attenuation coefficients. As an optical tissue property, the attenuation coefficient is not affected by common imaging artefacts such as shading.

 

Methods

 

Segmentation methods detect interfaces between layers principally based on intensity information. A novel LS method exploits anatomical knowledge about the retina and incorporates it via LS coupling, thereby simultaneously detecting interfaces. A GC approach was adapted to favour layered structures. Starting from the outer retina, it iteratively detects retinal layers. One eye for each of 24 normal subjects was imaged with a Spectralis OCT system. 10 eyes were used for training of the algorithms, the other 14 eyes were used to assess the accuracy. 6 eyes were imaged again on the next day to evaluate the reproducibility. One B-scan of every scan was manually segmented. Three interfaces were considered: the vitreous-RNFL interface, the RNFL-GCL interface and the IPL-INL interface and thicknesses of two layers: the RNFL and the GCC (the RNFL, GCL and IPL). Errors used were the root mean square error (RMS), mean absolute deviation (MAD) and Dice coefficient; blood vessels were excluded from evaluation.

 

Results

 

An example of the segmentation results on the attenuation coefficient image of a normal and glaucoma eye is shown in figure 1. All evaluation results are listed in figure 2. The accuracy of LS expressed by MAD is on average 1.5 µm better than the GC. The reproducibility of automated methods has on average 1 µm smaller MAD than the reproducibility of manual annotations. The reproducibility of LS is on average 0.3 µm (MAD) better than the reproducibility of GC. The accuracy of both methods is within 2 µm of the reproducibility of manual annotations.

 

Conclusions

 

Both methods are capable of segmenting OCT data, but our LS approach shows better accuracy and reproducibility. Automated methods are more consistent than manual annotations. Accuracy of both methods and intra-observer reproducibility of manual annotations are similar suggesting that the accuracy of the automatic segmentation is at least as good as the manual annotations.

 

 

Segmentation results on a normal and glaucoma eye

 

 

Accuracy and reproducibility evaluation

 

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