How Accurate is Automated Segmentation of Optic Nerve Head Structures in Spectral Domain Optical Coherence Tomography?

Abstract

 

Purpose:

 

To quantify and characterise the accuracy of automated segmentation of optic nerve head (ONH) structures from images obtained with spectral domain optical coherence tomography (SD-OCT).

 

Methods:

 

SD-OCT imaging (Spectralis, 24 high resolution, 15º B-scans centred on the ONH) was performed on 105 glaucoma patients (median age 71 y) and 46 healthy controls (median age 65 y). The positions of the internal limiting membrane (ILM) and Bruch’s membrane opening (BMO) were manually segmented in each B-scan using custom visualisation software. These positions were used to measure the accuracy of a Beta version automated segmentation algorithm. Error was quantified by the distance between automated and reference ILM curves and also between BMO margins in each B-scan. Image quality score and ILM height variation were investigated as predictors for segmentation accuracy. The proportions of B-scans with no available position for a structure (a "failure") for each subject were also quantified.

 

Results:

 

The median (IQR) mean distance for the ILM was 10.0 (6.3, 26.7) µm for patients and 8.2 (5.6, 17.9) µm for controls (Mann-Whitney p=0.01). ILM segmentation error was higher inside the BMO than outside (Mann-Whitney p=0.01). ILM error outside the BMO represented, on average, 12% of retinal nerve fibre layer thickness in these subjects. The median (IQR) BMO error was 36.0 (29.1, 49.6) µm in patients and 43.3 (31.1, 47.6) µm in controls (p=0.16). Image quality score was weakly and non-significantly correlated to segmentation accuracy of both ILM and BMO (Figure). A moderate (Spearman’s ρ=0.41) but statistically significant (p<0.001) correlation was found between the ILM height variation and the ILM segmentation error. Failure to segment BMO in 1-2, 3-4 and >4 out of 24 B-scans occurred in 19%, 10% and 0% of patients and 11%, 4% and 2% of controls respectively and did not occur with ILM segmentation.

 

Conclusions:

 

Advances with SD-OCT and its clinical application require accurate, automated segmentation tools. Image quality score was not related to ILM and BMO segmentation accuracy. ILM segmentation errors were more likely to occur inside the BMO and in ILM morphologies with more height variation.  

 

Keywords: optic disc • imaging/image analysis: clinical • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)