Segmentation error in spectral domain optical coherence tomography measures of the retinal nerve fibre layer thickness in idiopathic intracranial hypertension

Quantifying papilloedema clinically is subjective and prone to inter–observer variability and inaccuracy during prospective monitoring [1]. Spectral Domain Optical Coherence Tomography (SDOCT) is increasingly used both in the clinical environment and as outcome measures in Idiopathic Intracranial Hypertension (IIH) clinical trials to objectively quantify papilloedema [2]. Commercially available SD-OCT imaging systems, such as the Cirrus HD-OCT (Carl Zeiss Meditec, Dublin, CA) and Spectralis (Heidelberg Engineering, Heidelberg, Germany), have proprietary in-built OCT software logarithms which use the difference in signal intensity between adjacent retinal layers to perform automated segmentation to segment inner and outer retinal boundaries, from which retinal nerve fibre layer thickness (RNFL thickness) measurements can be calculated. Optic disc swelling can be monitored by repeated assessments of the RNFL thickness [2‐4]. Autosegmentation has been found to be inaccurate in some retinal pathologies such as neovascular age related macular degeneration and central serous retinopathy [5] and in optic nerve head pathologies such as glaucoma [6]. In papilloedema, where the interface between the retinal layers is disturbed by oedema, errors in autosegmentation have been noted [7‐9], with large studies using the Cirrus HD-OCT. The aim of this study was to evaluate the extent and location of the RNFL thickness SegE in an IIH cohort, comparing this to normal controls using the Spectralis SD-OCT.

Of the 52 IIH patients, the scans from 6 patients were excluded as they had such severe papilloedema that the optic nerve head elevation was truncated by the scan image and therefore the height could not be visualised and no further accurate refinement of the RNFL thickness could be performed. Forty-six IIH subjects and 14 controls were therefore included in the quantitative analysis (Fig. 1). The reliability between the two independent raters (AA and JH) was 0.732, with 95% CI (0.232–0.926), p < 0.05.

Quantification of the difference between the automated and the manually corrected total RNFL thickness area using ImageJ revealed significantly greater SegE in IIH patients [5% change post segmentation refinement, range = 0–58%] compared to controls [1% change post segmentation refinement, range = 0–6%] p = 0.009 (Table 1; Additional file 1: Table S1, S2). This was particularly evident in IIH patients with moderate to severe papilloedema [10% change post segmentation refinement, range = 0–58%, p < 0.001]. (Additional file 1: Table S3).

The error in automated overall average RNFL thickness values was significantly greater in IIH compared to controls (p = 0.031): median176μm (range 76-581 μm) pre re-segmentation versus 159 μm (range 83-391 μm) post re-segmentation in IIH (4% change post re-segmentation, range 0–58%) this was compared to 98 μm (range 63-125 μm) pre segmentation versus 100 μm (range 65-126 μm) post re-segmentation in controls (2% change post re-segmentation, range = 0–6%). IIH patients with moderate to severe papilloedema displayed significantly greater error in the overall average RNFL thickness values compared to those with mild papilloedema [10% change post re-segmentation (range 0–58% in moderate and severe papilloedema, p = 0.002)] (Table 1 and Fig. 2). In those with moderate to severe papilloedema the SegE was significantly greater in the superior retinal quadrant [11% change post re-segmentation, range = 0–375%, p = 0.001] (Fig. 3).

Qualitative assessment of any error in the RNFL thickness segmentation was more often identified in IIH (any apparent error in 98% of the IIH group, (45/46)) compared to control subjects (error in 79% of patients, (11/14)) (Table 2). It was clearly observed that the magnitude of the SegE was more pronounced in the IIH compared to control subjects. There was no clear pattern between subjects of whether the error was inflation or deflation of their disc height. In IIH patients, the RNFL thickness SegE was predominantly noted in the superior retinal quadrant, but to a lesser degree in the inferior retinal quadrant. In contrast, the control subjects displayed minimal error that had no predominant distribution.

OCT imaging is increasingly utilised for quantification and monitoring of papilloedema in IIH in the clinical setting. The largest prospective controlled cohort in papillodema reporting use of OCT in IIH is data from the IIHTT [2]; this study investigated 126 participants and utilised the Cirrus OCT at multiple centres and found that 3-dimensional scanning was less prone to failures of segmentation than 2-dimensional images. There is limited literature on the accuracy of autosegmentation in papilloedema using the Spectralis OCT. Detection of SegE, and manually refining the interfaces could help improve the accuracy of the RNFL thickness values between consecutive tests, and improve the clinical utility of the SD-OCT in longitudinal monitoring of IIH. This study highlights the issue of significant error in the automated RNFL thickness values generated from peripapillary RNFL thickness circle scans using the Spectralis SD-OCT. Like other SD OCT devices, the Spectralis in built algorithms are not specifically designed to autosegment papilloedema and although we identified significant error in the overall average RNFL thickness value with a 4% change following manual re-segmentation, this is much less error than previously reported using time domain OCT platforms [12]. What is yet to be determined is the clinical significance of the magnitude of this error.

The SegE was most apparent in those with moderate to severe papilloedema (10% error) and particularly in the superior retinal quadrant (11% error). The majority of the RNFL thickness error was accounted for by inaccurate automated identification of the lower boarder of the RNFL at the junction with the ganglion cell layer. It is likely that oedema and vessel artefact lead to error in the average RNFL thickness automated values, as postulated by previous authors [9, 13].

In 6% of the cohort, with severe papilloedema, the extreme elevation of the optic nerve head obscured the upper boarder limit of the RNFL and it was therefore not possible to refine the segmentation in these patients due to the truncation of the image. This truncation artefact has been previously reported by other authors [the type 1–8 paper] [9]. It may be less of an issue with newer OCT systems, using swept source technology, that provide a greater depth of imaging (e.g., Topcon DRI OCT-1 Triton has a depth range of 2.6 mm – greater than the 1.9 mm depth range of the Heidelberg system, based on spectral domain technology) [14].

SegE is not the only cause for erroneous RNFL thickness values; a number of factors have been associated with artifacts in OCT scanning including decentration error, refractive error, posterior vitreous detachment artifacts, reduced visual acuity, small pupils, presence of media opacities, advanced stage of glaucoma and dry eyes [9, 15‐18] Eye tracking on Spectralis ensures better alignment and is reported to decreases error in malalignment [19, 20].

Manual refinement of segmentation has several limitations which include the time taken to perform this accurately; indeed the accuracy of the manual markings, in which experts invariably disagree on where to draw the margins when the boarders between layers are hazy. In this study we performed an inter-user variability check to ensure that there was sufficient agreement between two masked individuals. The clinical impact of this error has not been evaluated in this study but would be a useful area for future investigation. As the Spectralis SD OCT platform was used in this study our results may not be generalizable to results from other types of SD-OCT machines.

RNFL thickness peripapillary scans are not the only OCT scan modality used to assess papilloedema. Other scanning modalities include volumetric analysis of the optic nerve head and macula; and Bruchs Membrane Opening (BMO) rim analysis. However, the degree of oedema in moderate to severe papilloedema is also known to cause error in these scans due to optical penetration. Future solutions include better depth penetration and a wider scan window to account for the elevation found in disc oedema. Polarization-sensitive OCT, which is not currently commercially available, has the potential to delineate the RNFL boundary better based on pigment differences in the retinal layers and be less prone to SegE [21].

In the setting of virtual IIH clinics where patient’s management may be judged exclusively by objective OCT and Humphrey visual field values; there could be clinical risk in the misinterpretation of the degree of papilledema and its course over time if the SegE is not identified and corrected for at the time of acquisition of the scans [9]. Here we have highlighted the limitations of using automated results from OCT RNFL thickness scans, particularly in those with marked papilloedema. We have developed a suggested paradigm to guide healthcare professionals performing OCT RNFL thickness peripapillary scans in IIH (Fig. 4). Scans should be evaluated for error at the time of acquisition to ensure the accuracy of the data at the time of the clinical visit when management decisions may be being made. However, as demonstrated with the results from the IIHTT [2] we would recommend the use of optic nerve volume scans in the routine clinical assessment of papilloedema.

Using the Spectralis SD-OCT, SegE in RNFL thickness values were found to be greater in IIH than controls, with the error increasing with the severity of the papilloedema. Imaging was not useful in very severe papilloedema where the image was truncated. This is the largest cohort assessing SegE in IIH, using the Spectralis SD-OCT. Achieving accurate and reproducible image analysis is important in the longitudinal monitoring in IIH, hence recognition of SegE and manual refinement should be understood by technicians and clinicians alike.