Performance of OCT segmentation procedures to assess morphology and extension in geographic atrophy

Authors


Dr U. Schmidt-Erfurth
Department of Ophthalmology
Medical University of Vienna
Waehringer Guertel 18-20
1090 Vienna
Austria
Tel: + 43 1 40400 7931
Fax: + 43 1 40400 7923
Email: ursula.schmidt-erfurth@meduniwien.ac.at

Abstract.

Purpose:  Investigating segmentation procedures and morphological findings in time domain (TD) and current spectral domain (SD) optical coherence tomography (OCT) devices in patients with geographic atrophy (GA).

Methods:  Fifty eyes of 46 patients with GA secondary to AMD and 15 control eyes were examined in this prospective noninterventional comparative case series. All patients underwent Stratus (model 3000), Cirrus (Carl Zeiss Meditec), Spectralis (Spectralis HRA+OCT; Heidelberg Engineering) and 3D-OCT-1000 (Topcon). Automated segmentation analyses were compared. An overlay of scanning laser ophthalmoscope (SLO) and three-dimensional retinal thickness (RT) maps were used to investigate whether areas of retinal thinning correspond to areas of retinal pigment epithelium (RPE) atrophy.

Results:  Geographic atrophy areas identified in SLO scans were significantly larger than areas of retinal thinning in RT maps. No convincing topographic correlation could be found between areas of retinal thinning and actual GA size as identified in SLO and fundus photography. Spectralis OCT showed significantly more mild and severe segmentation errors than 3D and Cirrus OCT.

Conclusion:  This study showed substantial limitations in identifying zones of GA reliably when using automatic segmentation procedures in current SD-OCT devices. This limitation should be addressed to visualize and document RPE loss realistically in a frequent disease like GA.

Introduction

Age-related macular degeneration (AMD) is the leading cause of severe visual loss in the developed world. Geographic atrophy (GA) characterizes a form of advanced AMD occurring in approximately 3.5% of the population aged 75 or older (Klein et al. 1993; Vingerling et al. 1995). Geographic atrophy is responsible for severe and irreversible visual impairment in approximately 20% of all patients with AMD (Hyman et al. 1983; Ferris et al. 1984). Even though demographic data reveal GA to be a major clinical challenge today, the pathophysiological mechanisms underlying the atrophic process are still not entirely understood (Maguire & Vine 1986; Schatz & McDonald 1989; Holz et al. 1994; Sunness et al. 1999). Unlike the neovascular form of AMD, GA usually progresses slowly and often spares the foveal centre until late during the natural course of the disease (Sunness et al. 1997; Sunness 1999). The primary visual impairment originates from the presence of absolute scotomas limiting the functionality of the available central visual field because of retinal pigment epithelium (RPE) loss and subsequent retinal atrophy (Maguire & Vine 1986; Sarks et al. 1988; Schatz & McDonald 1989; Sunness et al. 1997). This was recently observed in 42% of eyes with GA in a population-based study (Klein et al. 1995).

A recent study has shown that the extension of the atrophic area is a critical determinant for reading performance, implicating that the size of retinal atrophy may play a significant role in patient’s daily vision-related activities (Sunness et al. 1996).

Few data are available concerning specific morphologic changes observed in GA using conventional time domain (TD) and more recently spectral domain (SD) optical coherence tomography (OCT). Fleckenstein et al. (2008) described specific alterations like plaques at the level of Bruch’s membrane, clumps of abnormal RPE cells at different retinal levels, residual deposits and residual drusen material within the area of GA (Fleckenstein et al. 2008), whereas typical intraretinal changes were described by Wolf-Schnurrbusch et al. (2008).

No data, however, exist on quantitative analyses of the assessment of the RPE and retinal thickness (RT) using standard and high-resolution OCT. Moreover, little attention has yet been paid to the precise identification of the size of atrophic areas on the basis of OCT technology in patients with GA. Such quantitative measurements, however, are essential for the documentation of disease progression and evaluation of novel therapeutic strategies.

To close this gap, this study was designed to assess the potential and quality of morphologic segmentation procedures using TD and four different SD-OCT devices (Stratus OCT, Cirrus OCT, Spectralis OCT and 3D-OCT-1000) in patients with GA. Rates of errors in segmentation-based measurements were identified and compared for each device.

Materials and Methods

Study design, inclusion criteria and examination procedures

Fifty eyes of 46 consecutive patients with GA secondary to non-neovascular AMD were included in this prospective noninterventional case series. Patients had to be 55 years or older presenting with GA in at least one eye. A healthy control cohort (n = 15) was evaluated to identify differences in macular volume and RT compared to patients with GA.

Written informed consent was obtained from each patient, and the nature of the study was explained in detail before participants were included. The local ethics committee had approved the study protocol. Ethical tenets of the Declaration of Helsinki were followed.

Primary GA was defined as one or more areas of RPE loss, measuring ≥200 μm in diameter in treatment naive patients prior to patient inclusion, as this dimension seemed to be an appropriate size limit to perform segmentation analysis. Exclusion criteria were the presence of other retinal disease, including neovascular AMD, which was excluded by medical history and a fluorescein angiography (FA) following a standardized procedure.

All patients underwent a comprehensive ophthalmologic examination after pupil dilation with Tropicamide 0.5% and Phenylephrine 2.5%, including measurement of best-corrected visual acuity using early treatment diabetic retinopathy study (ETDRS) protocols, indirect slit lamp biomicroscopy, fundus photography, FA and OCT measurements at a 1-day visit.

In this study, patients with a minimum of five readable letters on ETDRS protocols were included. Further, patients with a cut-off of ±4 spherical equivalents were included.

To evaluate the performance of automatic segmentation procedures, which are the basis for RT analysis, all patients underwent conventional Stratus (model 3000, software version A 4.0), Cirrus (Carl Zeiss Meditec, Dublin, CA, USA), Spectralis (Spectralis HRA + OCT; Heidelberg Engineering, Heidelberg, Germany, Eye Explorer Vers. 1.6.1.0, acquisition Software 4.0.0.0) and 3D-OCT-1000 (Topcon, Tokyo, Japan, v.2.11). For this study, automated segmentation analyses were performed using the standard software analysis provided by the individual manufacturer.

Changes in RT and volume obtained by segmentation analysis were compared between patients with GA and a healthy control group. Quantitative measurements of GA extension measured by SD-OCT retinal mapping were correlated with lesion measurements based on fundus photography and scanning laser ophthalmoscope (SLO) imaging (Cirrus OCT) focussing on lesion size and topography.

OCT imaging procedures and properties of different devices

Stratus OCT (Carl Zeiss Meditec) is still the most frequently used OCT device in clinical practice. It relies on TD-OCT technology, achieving an axial resolution of 10–20 μm. Maximum scanning speed was 400 A-scans per second. For RT analysis, Stratus OCT performs six individual scans, which are arranged in a radial pattern. To confirm correct positioning of the scan, a fundus photograph is taken following the scanning process.

For comparison of different segmentation procedures, conventional macular thickness scans were obtained for this study.

Three-dimensional OCT (Topcon 3D-OCT-1000) is a new SD-OCT system, combining OCT technology with a colour nonmydriatic fundus camera. The device allows combining 3-D OCT imaging and fundus registration using retinal vessels identified in fundus photography as landmarks. Scanning speed reaches 18 000 axial scans per second at a longitudinal axial resolution of 6 μm. A raster scan containing 128 B-scans as well as a high-resolution radial pattern, comprising six B-scans, were performed.

Cirrus OCT achieves an axial resolution of 5 μm at 25 000 A-scans per second. Raster scans are registered to an SLO image which is performed simultaneously with the scan; the scanning area measured 6 × 6 mm. For this study, a 512 × 128 × 1024 as well as the 200 × 200 × 1024 scan was used to image retinal pathology.

The Spectralis OCT obtains high-resolution cross-sectional images of the retina, by computing mean images of a variable amount of B-scans, which are registered to the identical location using an eye tracker function which simultaneously acquires SLO images to correct motion artefacts and ensure correct positioning of the OCT B-scan. Spectralis OCT is also able to combine FA, autofluorescence, IR and SLO imaging in one device. A macular raster scan consisting of 37 B-scans as well as high-resolution scans of the foveal region was performed in this study.

OCT segmentation analysis

Segmentation errors were classified as negligible (≤50 μm), mild (50–200 μm) and severe (≥200 μm) errors. These errors were defined as deviations of the automated segmentation algorithms from the real anatomical structures, identified by an expert OCT grader. Structures relevant for delineation were the internal limiting membrane (ILM) and the RPE, based on which RT values are calculated in Spectralis and Cirrus, whereas 3D-OCT-1000 and Stratus identified the IS/OS junction as RPE. Fifty data sets, including data of all four OCT modalities (Stratus, Cirrus, Spectralis and Topcon OCT) were available for image analysis. Images were graded in a randomized order. The grader was masked to patient identities and scanning dates.

Central retinal thickness was defined as the RT in the middle of the fovea e.g. the point of fixation in an individual patient. Retinal volume (RV) is calculated from the area between the ILM and RPE and interpolated over the scanning area of a complete raster scan.

Identification of GA-lesion size in OCT

To investigate the potential of RT maps to precisely identify atrophic lesions in the central retina, a grader evaluated fundus photography in comparison to the SLO images from Cirrus- and Spectralis OCT.

Depressions in the surface profile of RT maps >50 μm (derived from Cirrus OCT) were marked manually and correlated to the contour of GA and SLO images using Adobe Photoshop CS3 and Autocad Architecture 2008 (US Metric) software (San Rafael, CA, USA). The areas of GA were measured by using the software Autocad Architecture 2008 (US Metric) by manually marking the identified areas of GA, after copying the images from Adobe Photoshop CS3 into Autocad Architecture 2008 (US Metric). Subsequently, the software automatically calculated the areas of atrophy.

For these analyses, the RT maps of the Cirrus OCT were used, because they were proven to have the best performance in this study.

To give an objective impression of different OCT’s segmentation performance, all scans included into the specific macular raster protocol were analysed for segmentation errors (Spectralis n = 37, Cirrus n = 128, 3D OCT-1000 n = 128).

Statistical methods

Statistical analyses were performed using spss software (SPSS Inc., Chicago, IL, USA vs. 15.0) and Microsoft-Excel 2007. Statistical tests used were the independent sample t-test, the paired t-test, the Mann–Whitney U test, as well as bivariate correlation analysis. p < 0.05 was considered to be statistically significant.

Results

Patient demographics

Optical coherence tomography was performed in 50 eyes of 46 patients with unilateral or bilateral GA. Twenty-nine patients were female. Twenty-three right eyes and 27 left eyes were examined. Mean age was 77 ± 9.6.

High-resolution and conventional OCT findings

All patients showed specific morphological changes at different anatomical levels. Perilesional spots of hyperreflective material could be observed at the level of the outer plexiform layer (OPL). Varying degrees of retinal thinning were documented largely corresponding to areas of RPE atrophy.

Focal changes of structures like the OPL, inner plexiform layer, internal (ILM) and external limiting membrane could be differentiated more precisely in SD-OCT than in conventional Stratus OCT imaging.

Segmentation performance of different OCT devices

Stratus OCT showed significantly more severe segmentation errors than Cirrus OCT (p < 0.001) and 3D-OCT-1000 (p = 0.01). Algorithm failures were mainly caused by wrong delineation of the RPE layer, while segmentation of the ILM was generally well performed.

Compared to Cirrus, Topcon’s 3D-OCT-1000 showed significantly more severe errors (p < 0.001). Interestingly, the radial scan comprising higher resolution images demonstrated a better performance with significantly less severe segmentation errors than the 128 by 512 raster scan in 3D-OCT-1000.

Spectralis OCT demonstrated significantly more mild segmentation errors than 3D-OCT-1000 (p ≤ 0.001). Moreover, Spectralis OCT showed significantly more mild and severe segmentation errors than the 3D-OCT-1000’s radial scan pattern (p < 0.001). An overview of segmentation performance of the OCT and SD-OCT devices is given in Fig. 1.

Figure 1.

 Frequency of segmentation errors among different optical coherence tomography (OCT) devices and scanning patterns: Fig. 1 shows the frequency of segmentation errors in all OCT devices investigated (Cirrus, Spectralis, 3D-OCT-1000 and Stratus). Stratus OCT shows more severe segmentation errors than Cirrus and 3D-OCT-1000. Moreover, 3D-OCT-1000 shows more severe segmentation errors than Cirrus OCT. Spectralis OCT demonstrated more mild segmentation errors than 3D-OCT-1000.

Determination of GA-lesion size

Fundus photography performed with 3D-OCT-1000 corresponded to atrophic areas identified in Spectralis and Cirrus SLO images. An example of a fundus photography corresponding to the SLO image is given in Fig. 2.

Figure 2.

 Scanning laser ophthalmoscope image and corresponding fundus photography in a patient with geographic atrophy.

Areas of RPE loss because of GA observed by SLO were significantly larger (p < 0.001) than the area of retinal thinning identified in RT maps (mean area in SLO = 10.2 vs. 6.1 mm2 in RT maps; Fig. 3).

Figure 3.

 Areas of geographic atrophy observed in scanning laser ophthalmoscope scans versus retinal thickness maps.

Despite the relatively mild difference in total quantitative values, the topographic correlation focussing on localization and extension of GA zones in RT maps and SLO was poor (Fig. 4). A consistent topographic correlation could only be found in <10% of our patients.

Figure 4.

 Topographic correlation of retinal thickness maps and retinal pigment epithelium atrophy identified in scanning laser ophthalmoscope images.

Differences in RT and RV in patients with GA and healthy probands

A difference in mean RV could be observed and was statistically significant (p < 0.001) between the group with GA and the control group (9.2 mm3 in GA vs. 10.1 mm3 in the control group). In contrast to these differences observed in the overall RV, mean central RT values differed only marginal between patients with GA (243 μm) and the control group (264 μm). This difference was not statistically significant (p = 0.09), indicating that central RT might not be the ideal parameter to evaluate GA. Original data are listed in Table 1; data are highlighted in Fig. 5. For the evaluation of RT and RV, 512 × 128 × 1024 Cirrus OCT scans were used.

Table 1.   Data of central retinal thickness and retinal volume (RV) in patients with geographic atrophy (GA) and controls. For the evaluation of RT and RV, 512 × 128 × 1024 Cirrus optical coherence tomography scans were used.
 CRT (μm) in GA eyesCRT (μm) in control eyesRV (mm3) in GA eyesRV (mm3) in control eyes
Mean243.2264.19.210.1
Median243.5267.09.310.1
Standard deviation (SD)50.118.90.70.3
Figure 5.

 Central retinal thickness and retinal volume of patients with geographic atrophy and the control group.

Discussion

It was the aim of this study to assess the performance of segmentation procedures provided by current OCT devices in GA. One TD and three current SD-OCT instruments were evaluated and compared. Fundus photography, SLO and RT map imaging were performed and correlated to RT maps to identify the feasibility to pick up the extension of GA lesions from RT maps. Comparison of central retinal thickness (CRT) and volume between GA and control eyes revealed significant differences for RV but not for CRT values. This indicates that RV rather than CRT values should be considered to screen for GA.

Morphologic analysis of OCT devices showed similar findings to a study recently published by Fleckenstein et al. (2008). As expected, SD-OCT, especially the Spectralis and the Cirrus OCT, showed the highest level of details. High-resolution images of SD-OCT devices presented significantly more detail than Stratus OCT allowing to identify even discrete alterations in the affected tissue reliably. Hence, imaging the retinal microstructure of GA-related changes is qualitatively advanced using SD-OCT.

However, this study also showed that automatic segmentation procedures used in current OCT and SD-OCT devices have significant limitations in delineating, visualizing and measuring the area of true RPE atrophy and subsequent retinal thinning. This was generally related to misdetection of the RPE in this study. Interestingly, a higher transversal resolution and shorter scanning time seemed to reduce these segmentation artefacts. This effect became obvious in the significant difference between the segmentation performance of the high-resolution six line scan and the macular raster scan comprising 128 B-scans in Topcon’s 3D-OCT-1000.

Interestingly, Spectralis OCT revealed more mild and severe segmentation errors than Cirrus OCT and 3D-OCT-1000. One reason may be that Spectralis OCT obtains high-resolution cross-sectional images of the retina, by computing mean images of a variable amount of B-scans and does not acquire 128 raster scans as Cirrus OCT and Topcon OCT-1000. On the other hand, both OCT devices might use different segmentation and software algorithms, which could be another reason for different segmentation performance.

Comparison of RPE atrophy, as detected by SLO, revealed that areas of GA seemed to be underestimated in size when RT maps were used to identify GA.

As no consistent topographic correlation between SLO images and retinal map overlays could be found in our patients, retinal mapping procedures seem to be limited in their ability to detect areas of GA precisely. This implicates that RT maps are of limited value when trying to exactly delineate and quantify the area of true RPE tissue loss.

Concluding, SD-OCT, enabling unique visualization of all major intraretinal layers (Schmidt-Erfurth et al. 2005), proved to be clearly superior to conventional TD-OCT in imaging of GA, because distinct morphologic details could be visualized in greater detail, as TD-OCT systems are compromised by a lower scanning speed, consequently allowing only a few B-scans per second (Ahlers et al. 2008b). Advanced analysis and true location imaging because of the raster scanning mode allowed for better detection and reliable visualization of RT and the RPE compared to TD-OCT. Further, the unique capacity of SD-OCT allows for improved localization of affected retinal tissue (Ahlers et al. 2008a,b). However, current automated SD-OCT analyses do not identify the extension of primary RPE atrophy accurately and lack the ability to differentiate discrete zones of RPE atrophy in early disease or different disease patterns. This drawback should be addressed in future developments to be able to take full advantage from higher resolution and true location imaging.

Acknowledgements

None.

Ancillary