January 2009
Volume 50, Issue 1
Free
Retina  |   January 2009
Segmentation Error in Stratus Optical Coherence Tomography for Neovascular Age-Related Macular Degeneration
Author Affiliations
  • Praveen J. Patel
    From the Moorfields Eye Hospital, London, United Kingdom.
  • Fred K. Chen
    From the Moorfields Eye Hospital, London, United Kingdom.
  • Lyndon da Cruz
    From the Moorfields Eye Hospital, London, United Kingdom.
  • Adnan Tufail
    From the Moorfields Eye Hospital, London, United Kingdom.
Investigative Ophthalmology & Visual Science January 2009, Vol.50, 399-404. doi:10.1167/iovs.08-1697
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to Subscribers Only
      Sign In or Create an Account ×
    • Get Citation

      Praveen J. Patel, Fred K. Chen, Lyndon da Cruz, Adnan Tufail; Segmentation Error in Stratus Optical Coherence Tomography for Neovascular Age-Related Macular Degeneration. Invest. Ophthalmol. Vis. Sci. 2009;50(1):399-404. doi: 10.1167/iovs.08-1697.

      Download citation file:


      © 2017 Association for Research in Vision and Ophthalmology.

      ×
  • Supplements
Abstract

purpose. To describe the rate of automated segmentation error in Stratus optical coherence tomography (OCT) scans in consecutive patients with neovascular age-related macular degeneration (nAMD) receiving treatment and to investigate the effect of the segmentation error on automated retinal thickness measures and whether further imaging reduces the rate of segmentation error.

methods. A retrospective analysis of fast macular thickness map (FMTM) protocol OCT scans of 50 eyes of 50 consecutive patients with nAMD. Each line scan was analyzed for segmentation error with manual measurement of the center-point retinal thickness allowing calculation of the percentage error in automated thickness. OCT scanning was repeated to overcome segmentation error.

results. Segmentation error was detected in 45 (90%) of the 50 patients with 37 (74%) patients having an error affecting the central 1-mm subfield. Scan sets with a high central segmentation error score (two or more line scans affected of six) had a significantly greater error in automated center-point retinal thickness than scan sets with a low error score (20% compared with 3%, P < 0.000005). Central subfield segmentation error persisted in 30 (60%) patients despite repeat scanning.

conclusions. There is a high rate of segmentation error in OCT scans of patients with nAMD who are undergoing treatment, leading to errors in automated central retinal thickness measurement. The authors recommend manual measurement of central macular thickness when two or more line scans are affected by segmentation error in the central 1-mm subfield. Repeated scanning reduced the rate of error but did not eliminate the problem.

Optical coherence tomography (OCT) is a noninvasive technique involving analysis of optical interference patterns of low coherence and wide bandwidth light to provide cross-sectional images of the retina. 1 The development of new intravitreous treatments for neovascular age-related macular degeneration (nAMD) that block the action of vascular endothelial growth factor (VEGF)-A have led to increased interest in OCT analysis of these patients. The ability of OCT to quantify changes in retinal thickness has facilitated assessment of the response to treatment and the need for retreatment. 2 A knowledge of the rate of OCT image artifacts and errors in automated retinal thickness measurement is particularly important in this group of patients, as ophthalmologists must be able to confidently decide on retreatment based on automated quantitative OCT analysis and qualitative features that indicate disease activity. 2  
The most favored protocol for assessing quantitative change in macular thickness in patients with nAMD is the fast macular thickness map protocol (FMTM), which compromises image resolution to achieve fast image acquisition, less prone to movement error. The onboard Stratus OCT analysis mode uses segmentation software for automated detection of inner retinal (vitreoretinal interface) and outer retinal (retinal pigment epithelium; RPE) boundaries. The software then calculates the distance between these layers to estimate retinal thickness. When the Stratus software encounters problems in detecting the inner and outer limits of the retina, segmentation error results, with misidentification of retinal boundaries and consequent errors in the automated retinal thickness measurement. 
Investigators have reported the incidence of image artifact and segmentation error in Stratus OCT for patients with retinal disease, including patients with nAMD, although not specifically in a cohort undergoing anti-VEGF therapy. 3 4 Although complex scoring strategies predicting retinal thickness measurement error have been proposed, actual quantification of the deviation of automated from manual measurement has not been reported. 4 In addition, the effectiveness of repeating acquisition of OCT images (once image quality has been optimized) on reducing segmentation error has not been reported despite the suggestion that this may overcome the problem of segmentation error. 2  
We studied the incidence of segmentation error in consecutive patients being treated for nAMD, investigated the impact of these errors on retinal thickness measurement, and explored whether segmentation error may be overcome with repeated image acquisition in a single session. 
Methods
Equipment
All OCT imaging was performed with a commercially available OCT machine (Stratus OCT, with software version 4.0; Carl Zeiss Meditec, Inc., Dublin, CA). The OCT system is third-generation and provides an axial resolution of less than 10 μm. It is serviced regularly, in line with manufacturer recommendations by authorized technicians from Carl Zeiss Meditec, Inc., to ensure that the machine is calibrated and operating correctly. 
Subjects
OCTs of 50 eyes of 50 consecutive patients with choroidal neovascularization (CNV) due to AMD, who were undergoing evaluation and treatment with anti-VEGF agents, were analyzed by searching the Stratus OCT database at the Clinical Trials Unit at Moorfields Eye Hospital. Approval for the collection and analysis of OCT images was obtained from the Research Governance Committee of Moorfields Eye Hospital. The research adhered to the tenets of the Declaration of Helsinki. 
All patients in the study had subfoveal CNV due to AMD in the study eye and either had had or were about to undergo treatment. For each patient, only images from the eye undergoing treatment were used in the analysis. 
All scans were acquired between May 6, 2007, and June 15, 2007, with a single Stratus OCT machine. All patients had imaging performed after visual acuity measurement and pupil dilation with one drop of 2.5% phenylephrine hydrochloride and 1% tropicamide. They had given consent to OCT imaging as part of their clinical care or as part of their involvement in the clinical trial. The scans were from consecutive patients at different stages of treatment representing a spectrum of disease activity ranging from quiescent to active CNV. 
Image Acquisition
All OCT scanning was performed by a single technician certified by image-reading centers for OCT scanning in pharmaceutical company-sponsored AMD clinical trials. The FMTM protocol was used to assess retinal thickness with six high-speed, 6-mm radial lines (oriented 30° apart) delineating macular anatomy and pathology. The protocol enables all six line scans to be acquired in a continuous, automated sequence within 1.92 seconds, with each of the six lines composed of 128 equally spaced transverse sampled locations (total of 128 × 6 lines or 768 sampled points). At each of these locations, the signal is sampled axially at 1024 equal intervals over a depth of 2 mm. In this study, the term “scan set” refers to all six line scans acquired automatically as part of the FMTM protocol. Because of its short acquisition time, the FMTM protocol is believed to be less prone to errors due to unstable fixation—an important consideration in the assessment of the patient with nAMD. This rapid scan acquisition, however, is at the expense of resolution in the transverse plane. 
Each patient was aligned correctly with the OCT machine and was asked to look at an internal fixation light. If no light was seen, the patient was asked to look straight ahead by use of an external fixation light to ensure that the scans were taken through the fovea. The technician was experienced in identifying common artifacts in OCT images, and scan sets were reacquired as needed to optimize scan quality. The technician was instructed to discard any scans with (1) a signal strength of less than 7, (2) image artifact, or (3) a low analysis confidence message from the onboard Stratus OCT software (ver. 4.0) and to save only focused and centered images with signal strength greater than 6, although such standards were not always achieved because of poor fixation or media opacity (both common in patients with nAMD). Once image quality had been optimized, the technician saved the FMTM scan set. In line with departmental protocol, the OCT technician looked through all six line scans of each scan set for boundary detection accuracy and Stratus software generated low analysis confidence messages. For patients with segmentation error, between 5 to 10 optimized scan sets were acquired and saved in a single imaging session lasting a maximum of 20 minutes. Three scan sets with the fewest line scans affected by segmentation error were selected and saved, and the others were deleted. If there were more than three scan sets with an equal number of line scans affected by segmentation error, the scans sets with the fewest the number of line scans with low analysis confidence messages were saved. If there were more than three scan sets with an equal number of lines scans with a low analysis confidence message, the technician was asked to save the first three optimized scan sets acquired. However, no objective scoring method of central segmentation error was applied at this stage in scan selection. 
With repeated scanning and increasing imaging session time, patient fatigue may develop with deterioration in patient fixation stability, limiting the number of good-quality OCT images that may be acquired in a single session. 
Segmentation Error
Two types of scales were used to quantify the severity of segmentation error. The low analysis confidence message generated by the Stratus OCT software potentially identifies line scans with segmentation error, although it may not be specific for the central 1-mm subfield. As this feature is part of the proprietary software, we were unable to obtain details regarding how the message is generated. The number of line scans with a low analysis confidence message was noted and was termed the low analysis confidence score (LACS), with a range of 0 to 6 for each patient. 
The second scale used to grade severity of segmentation error was the central segmentation error score (CSES), defined as the number of lines scans (of six) with segmentation error affecting the central 1 mm of the scan (range, 0–6). Segmentation error affecting the central 1 mm of the scan was specifically noted, as thickness measures from this subfield are commonly used to guide retreatment in patients with nAMD. In addition, the central 1-mm zone of line scans may have a higher frequency of segmentation errors than more peripheral areas, as most patients with visually significant CNV have fovea involving lesions with more disruption of the photoreceptor-RPE complex in this central 1-mm zone. An experienced ophthalmologist certified for nAMD clinical trials work (PJP) determined the CSES by analyzing each of the six line scans that make up each FMTM protocol OCT scan set for each patient. We defined segmentation error as a visible difference (at least 16 μm) between computer algorithm-determined and observer-determined inner or outer retinal boundaries. 
Identification of segmentation error relating to the inner retinal boundary was relatively straightforward, with an obvious interface between the inner retinal surface and the low-reflectance vitreous. To determine the outer retinal boundary, we adopted the approach used by the Stratus software for straightforward cases (when there is no abnormality or when there is subretinal fluid with no fibrosis) and applied it to the more complex lesions (or suboptimal images) for which the segmentation software was less predictable. In less problematic cases in which a clear inner high-reflectance band (HRB) thought to relate to the junction of the photoreceptor inner and outer segments 5 6 was visible adjacent to a brighter more prominent HRB (thought to relate to the RPE layer), this inner HRB was marked as the outer retinal boundary. In more problematic cases in which the inner HRB was not clear or was displaced or disrupted due to other lesion components (subretinal fluid or tissue), then the brightest continuous HRB (inner RPE surface) was marked as the outer retinal boundary. In cases of lesion components contributing to and indistinguishable from the outer, RPE derived, HRB signal, the inner border of this thickened HRB was used as the outer retinal boundary. This led to inconsistency in the outer retinal boundary between scans, with the signal corresponding to the junction of the photoreceptor inner and outer segments used in some line scans with the inner surface of the RPE used in others in which the inner HRB signal was disrupted or displaced (thereby including subretinal fluid in the retinal thickness measurement); however, this represents an extension of the approach taken by the Stratus algorithm in less-problematic cases. 
A visible deviation (at least 16 μm) of the Stratus segmentation software-generated boundary from this manually determined boundary was identified as a segmentation error. 
Assessment of Interobserver Agreement in Determining Segmentation Error
As determining line scans with central 1-mm zone segmentation error is subject to interobserver variability, a second observer (FKC) assessed 60 line scans (10 consecutive patients) noted by the first observer (PJP) to have central segmentation error. The interobserver agreement for the presence or absence of central 1-mm zone segmentation error was assessed by calculating the κ statistic and determining the percentage agreement. 
Assessing Stratus OCT Automated Retinal Thickness Measurement Error
A manual center-point retinal thickness measurement was taken on each line scan by using a caliper measurement of the vertical distance between the manually determined inner and outer retinal boundaries in line with the 64th A-scan (3 mm from either end of the 6-mm line scan). This manual measure of retinal thickness therefore included any subretinal fluid or discernible subretinal high-reflectance lesion component (blood, CNV, or fibrosis) and was not therefore an accurate measure of retinal thickness. However, this approach was an attempt to mirror the retinal thickness measurement taken by the automated Stratus segmentation software using manually determined inner and outer retinal boundaries. 
A mean manual center-point retinal thickness was then calculated in each patient (by using the manual center-point retinal thickness measurement from all six line scans) and the percentage difference between the automated center-point retinal thickness and the manual one was also calculated. This percentage deviation of the manual center-point retinal thickness measurement from the automated value provided a measure of the retinal thickness measurement error produced by the proprietary Stratus segmentation algorithm. 
Figure 1illustrates one example of segmentation error and two examples of manual retinal thickness measurements in problematic cases. 
Assessment of Interobserver Repeatability of Manual Retinal Thickness Measurement
In view of the potential for variability in manual retinal thickness measurement, a second observer (FKC) performed retinal thickness measurements using the standardized method on 60 line scans from 10 consecutive patients noted by the first observer (PJP) to have central 1-mm zone segmentation error. The interobserver repeatability of the manual retinal thickness measurement was assessed by calculating the 95% coefficient of repeatability in line with methods outlined by Bland and Altman (1.96 × SD of the difference in measurement). 7  
Statistical Analysis
The number of patients with various numbers of OCT line scans with (1) segmentation error affecting any part of the image, (2) segmentation error affecting the central 1 mm (CSES), and (3) LACS were calculated. For further analysis, the ordinal variables, CSES and LACS, were collapsed into categorical variables of high and low scores. When the CSES or LACS was <2, the score was considered low, and when the CSES or LACS was ≥2 the score was classified as high. We have observed that segmentation error or low analysis confidence message appearing in more than one scan may represent a consistent error of the software, resulting in a larger and more significant error in retinal thickness measurement than if a single scan of six was affected. 
To determine whether any factors may be associated with the CSES, we used the Wilcoxon rank sum test to compare CSES for eyes with (1) high or low LACS, (2) high or low average signal strength, (3) good or poor visual acuity, (4) thin or thick foveal height, and (5) repeatable or variable measures of center-point thickness. 
To determine whether any factors may be associated with automated retinal thickness measurement error, we used the Wilcoxon rank sum test to compare the ratio of percentage measurement error in the groups of eyes with (1) high or low LACS, (2) high or low CSES, (3) high or low average signal strength, (4) good or poor visual acuity, (5) thin or thick foveal height, and (6) repeatable or variable measures of center-point thickness. P < 0.05 was chosen as significant. 
The impact of saving a second or third scan set on segmentation error was illustrated with cross-tabulation of the CSES across first, second, and third scan sets. The percentage of patients with OCT scan sets free of central 1-mm segmentation error and the percentage of patients with high CSES after saving three optimized scan sets were also reported. 
Results
Subject and OCT Characteristics
We studied 50 eyes of 50 consecutive patients with nAMD who were attending for treatment. The mean age (±SD) was 80 years (±7) with a range of 64 to 91 years. There were 28 women and 22 men; 49 were Caucasian. The visual acuity of the study eye ranged from an Early Treatment of Diabetic Retinopathy Study (ETDRS) score of 75 letters to 5 letters, with a mean of 48 letters (SD 17) and median of 49 letters (interquartile range, 22). 
The mean and median of the average signal strength in the 50 patients was 7.0, with a range from 4.0 to 9.8 (interquartile range, 1.9). Segmentation error was found in OCTs of 45 (90%) of the 50 patients. Table 1shows the number of eyes with various numbers of line scans affected by segmentation error or a Stratus software low analysis confidence message. The mean CSES for these 50 patients with nAMD was 2.1 (SD 1.9) with a median of 2.0 (interquartile range, 2.8). There were 27 patients with a high CSES. The mean LACS was 1.0 (SD 1.4) with a median of 0 (interquartile range, 1.8). A total of 13 patients had a high LACS. 
The mean percentage automated center-point retinal thickness measurement error was 12% (SD 18%) in the 50 patients. 
Interobserver Agreement in Determining Central 1-mm Segmentation Error and Repeatability of Retinal Thickness Measurements
The κ statistic for interobserver agreement in determining the presence or absence of central 1-mm zone segmentation error was 0.73 (95% confidence interval, 0.54–0.94) suggesting good agreement. The 95% coefficient of repeatability for retinal thickness measurements was 31 μm (11% when expressed as a percentage of mean retinal thickness). 
Factors Associated with CSES
The presence of a Stratus software generated low analysis confidence message in at least two of the six line scans was strongly associated with segmentation error with greater median CSES in scans with a high LACS (Table 2)
Factors Associated with Automated Retinal Thickness Measurement Error
Both the CSES and the LACS were found to be associated with greater automated retinal thickness measurement error (Table 3) . The association was much stronger for CSES than for LACS. There was a poor association between foveal thickness and automated retinal thickness measurement error but with a trend for increased error with variable center-point thickness measures (SD/thickness > 0.1) as shown in Table 3
Effect of Acquiring a Second or Third Optimized Scan Set on Segmentation Error
For the 37 patients with central 1-mm segmentation error in the first scan set, a second scan was saved after repeat OCT scanning and optimization of scan quality. This resulted in an additional four patients with no segmentation error in the central 1-mm subfield and seven patients with low CSES. Acquiring and saving a third optimized scan set in the 33 remaining patients resulted in a further three patients with scans sets free from segmentation error in the central 1-mm subfield and four with low CSES. Tables 4 and 5show the number of patients with low and high CSES after a second or third optimized scan set. After three optimized scans, the number of patients with scan sets with central 1-mm segmentation error decreased from 37 (74%) to 30 (60%), although 18 patients (36%) still had scan sets with a high CSES. 
Discussion
This study reports both the rate of segmentation error and the importance of central 1-mm segmentation error on automated retinal thickness measurement errors in OCTs of patients with nAMD in treatment with anti-VEGF agents. In addition, this study suggests that acquiring up to two further optimized scan sets in a single imaging session may help reduce segmentation error but does not completely eliminate the problem. Analysis of our data suggests that evaluating line scans for a central 1-mm zone segmentation error to report a CSES score for patients may be a better way of identifying scan sets with automated retinal thickness measurement errors than using the onboard Stratus low confidence analysis message. Furthermore, patients with two or more line scans with central 1-mm segmentation error (high CSES) have a mean automated retinal thickness measurement error of 20% (median 12%) and in line with our results, we advise performing manual measurements of retinal thickness in this subgroup. 
Ray et al. 3 reported the different types of artifact and error in OCT of patients with retinal disease and included a small subgroup of patients with nAMD, although not in a specific cohort undergoing anti-VEGF therapy. They found that 62.2% of scans were affected by retinal thickness measurement errors in patients with retinal disease, although no details were given regarding the 27 patients with nAMD. In addition, no analysis was performed to determine the relationship between segmentation errors and the magnitude of error in automated retinal thickness measurement. Others 4 have suggested that the methodology used by Ray et al. to identify image artifacts may have led to an underestimation of the rate of artifact and error. The study by Sadda et al. 4 characterized the rate and magnitude of segmentation error in OCT scans in a series of patients with a range of retinal disorders. They used a grading system with provision for weighting segmentation errors in the central 1-mm subfield and errors that were longer than 1 or 2 mm in length and a deviation of greater than one third of the true retinal thickness. This group reported a high rate of patients with OCT scans with retinal thickness measurement error (92%). Most scans (72.5%) were obtained with Stratus 4.0 software, with the remainder analyzed with the earlier version (ver. 3.0). Sadda et al. found a strong association between Stratus OCT software low analysis confidence and segmentation error. These results are similar to those obtained in this study although we report a better association between our CSES and retinal thickness measurement error than the Stratus low analysis confidence metric. In addition, we found good agreement between observers in determining segmentation error (κ = 0.73) with repeatable interobserver retinal thickness measurements (95% coefficient of repeatability = 31 μm or 11% of the mean thickness) using the standardized definition used in this study. 
Measurement of retinal thickness was not an important consideration in patients with nAMD when the Stratus OCT segmentation software was first developed; however, with the arrival of new intravitreous anti-VEGF therapies, there has been great interest in quantifying changes in retinal thickness in these patients. The quantitative information is used to assess response to treatment, as part of the retreatment criteria for antiangiogenic therapy 2 and as an endpoint in clinical trials. 8 Accurate and repeatable measurements of retinal thickness are now important in this patient group, and segmentation error is an important cause of errors in retinal thickness measurement with Stratus OCT. Identification of the inner retinal boundary is rarely problematic, as there is usually a clear step change in signal between the inner retinal boundary and the low-reflectance vitreous, although epiretinal membrane and signal from the posterior hyaloid face can sometimes lead to segmentation error. Significant problems may arise, however, in the automated identification of the outer retinal boundary where changes in signal may be less definite especially in patients with nAMD with disease affecting the photoreceptor-RPE complex. The Stratus OCT identifies the inner HRB as the outer retinal boundary, even though studies have shown the HRB to correlate better with the photoreceptor inner- and outer-segment junction, 9 Use of the HRB to identify the outer retinal boundary to systematic error in retinal thickness measurement using the proprietary software. However, when disease affects the photoreceptor-RPE complex, there is disruption of the inner HRB, and in these cases the Stratus algorithm often uses the high-reflectance signal from the inner RPE surface to identify the outer retinal boundary. Any moderate-to-highly reflective subretinal lesion components (such as blood, CNV tissue, or fibrosis) in close association with the photoreceptor-RPE complex may lead to further difficulty in identifying the HRB, resulting in erroneous and variable identification of the outer retinal boundary by the segmentation software. In instances in which the Stratus segmentation algorithm uses the signal from the inner surface of the RPE to determine the outer retinal boundary, subretinal fluid may be erroneously included in the retinal thickness measurement. 
Furthermore, calibration of OCT for retinal thickness measurements with optical path measurement using glass targets assumes a group index of refraction change that is based on normal tissue and not a mixture of fluid, lipids, and proteins not normally found in the retina, as may be the case in patients with nAMD, and further limits the accuracy of OCT-derived retinal thickness measurements in nAMD. 
Since quantitative measures of exudation in nAMD are increasingly driving treatment, it is crucial that OCT measures be made accurate in the presence of exudation, where the index of refraction changes are not always sufficient to generate a strong coherence signal. 
Our results suggest that visual acuity and retinal thickness do not influence segmentation error; however, the relationship between scan signal strength and CSES and retinal thickness measurement error is uncertain. Although caution should be applied to conclusions drawn from the subanalysis of factors influencing CSES and retinal thickness measurement error, the data from Tables 2 and 3suggest a trend for scans with signal strength ≥7 for a higher mean retinal thickness measurement error, despite a lower CSES. This finding suggests that a greater magnitude of retinal thickness measurement error may result from fewer line scans with central 1-mm zone segmentation error (a lower CSES) in scan sets with a mean signal strength of ≥7. This result could be due to large-magnitude errors in retinal thickness resulting from areas of subretinal moderate-to-high reflectance lesion components in a few line scans (low CSES) giving rise to errors in identification of the true HRB and outer retinal boundary. 
Fung et al., 2 who used an OCT-guided variable dosage regimen with intravitreous ranibizumab (Lucentis; Genentech Inc., South San Francisco, CA) for nAMD, comment that OCT scanning was repeated until retinal boundaries were accurately identified by the algorithm. One of the purposes of our study was to see whether elimination of segmentation error was possible with repeat scanning within the limits of patient fatigue and imaging time. 
In this study, repeat OCT scanning in patients with central 1-mm zone OCT scan segmentation error led to a further seven patients with central zone segmentation error-free scans after three optimized scans (a total of 40% of patients free of central 1-mm segmentation error). This result suggests that though operator and patient factors may lead to segmentation error in a proportion of patients with nAMD, failure of the Stratus algorithm is the underlying problem in most cases. This conclusion further implies that, although it is an important strategy for reducing segmentation error, repeated scanning may not eliminate the problem. Indeed even after three optimized scan sets were obtained there were still 18 patients (36%) with high CSES requiring manual measurement of retinal thickness. This finding is important, as it contrasts with reports from other studies that suggest that segmentation error may be eliminated with repeated scanning. 2 In addition, this finding is of importance due to the increasing use of quantitative OCT data in the treatment of patients with nAMD with anti-VEGF agents. 
Study Advantages and Limitations
This study highlights the problem of segmentation error in Stratus OCT scans of eyes with nAMD in patients undergoing anti-VEGF therapy and offers a simple way for the ophthalmologist to assess segmentation error. Unlike other studies, 4 our method of assessing segmentation error does not involve quantifying the length of any segmentation error or identifying deviations that are greater than one third of the true retinal thickness. We also report that the magnitude of automated retinal thickness measurement error is far greater if two or more line scans are affected by segmentation error in the central 1-mm zone. This method therefore provides the ophthalmologist with a guide when assessing the OCT scans of patients with nAMD as changes in central retinal thickness are used as part of the retreatment criteria for anti-VEGF agents. 
The study also reports the effect of repeated scanning on segmentation error. Limitations include the retrospective nature of the study and the emphasis on segmentation error and retinal thickness measurement error in the central 1 mm of line scans. Other investigators have highlighted the importance of assessing segmentation error outside the central 1-mm subfield when monitoring retinal thickness changes outside the fovea. 4 However, using scan protocols with radial lines to acquire information about more peripheral areas may be problematic, as the more peripheral the subfield of interest, the fewer the retinal points that are actually sampled and the more the software has to interpolate thicknesses. Other scanning protocols such as raster scanning may be more suitable than radial line scanning protocols when assessing subfields outside the central 1-mm (A1) subfield. A further limitation of our study in assessing the effect of repeat scanning on segmentation error was the limited number of optimized scan sets saved for analysis. The assessment was performed in line with departmental protocols but to determine the effect of repeat scanning further, more than three optimized scans should be saved, although in practice this may be limited by patient fatigue and imaging session duration. 
In summary, our study highlights the problem of segmentation error in OCT scans of patients with nAMD undergoing anti-VEGF therapy. We describe a simplified method to assess segmentation error that may be used to predict errors in automated retinal thickness measurements. In addition, the results suggest that, although it is a valuable strategy in reducing segmentation error, repeat OCT scanning is unlikely to completely eliminate the problem. These findings are important, as quantitative change in retinal thickness measurements is one of the criteria used to guide anti-VEGF retreatment. We therefore advise manual measurement of retinal thickness when making retreatment decisions with anti-VEGF agents for patients in whom segmentation error affects the central 1-mm zone in more than one of the line scans (i.e., CSES of 2–6) that make up an FMTM scan set. 
 
Figure 1.
 
An example of (A) segmentation error affecting the central 1-mm zone with (B) manual measurement of retinal thickness. An additional example of manual retinal thickness measurement (C) is also provided, in a case with a thickened HRB.
Figure 1.
 
An example of (A) segmentation error affecting the central 1-mm zone with (B) manual measurement of retinal thickness. An additional example of manual retinal thickness measurement (C) is also provided, in a case with a thickened HRB.
Table 1.
 
Distribution of Eyes with None to All of the Line Scans in a Scan Set Affected by Segmentation Error or Low Analysis Confidence
Table 1.
 
Distribution of Eyes with None to All of the Line Scans in a Scan Set Affected by Segmentation Error or Low Analysis Confidence
Number of Line Scans
None 1 2 3 4 5 6
Any SE 5 (10) 7 (14) 9 (18) 11 (22) 6 (12) 7 (14) 5 (10)
Central 1 mm SE 13 (26) 10 (20) 12 (24) 5 (10) 4 (8) 4 (8) 2 (4)
LAC message 29 (58) 8 (16) 5 (10) 4 (8) 3 (6) 0 (0) 1 (2)
Table 2.
 
Factors Associated with CSES
Table 2.
 
Factors Associated with CSES
Parameter n Mean CSES (SD), Median P (Wilcoxon Rank Sum Test)
LACS P ≤ 0.005
 0–1 (low) 37 1.6 (1.5), 1
 ≥2 (high) 13 3.5 (2.0), 4
Average signal strength (six line scans)
 ≥7 (high) 25 1.6 (1.6), 1 P ≤ 0.06
 <7 (low) 25 2.6 (2.0), 2.0
ETDRS letter score
 <50 (poor acuity) 26 1.8 (2.0), 1.5 P ≤ 0.57
 ≥50 (good acuity) 24 2.3 (2.0), 2.2
Manual center-point thickness (μm)
 <250 (thin fovea) 22 1.8 (1.5), 2.0 P ≤ 0.49
 ≥250 (thick fovea) 28 2.3 (2.1), 2.0
Manual center-point thickness standard deviation/thickness ratio
 ≤0.1 (repeatable measures) 38 2.0 (2.8), 1.8 P ≤ 0.48
 >0.1 (variable measures) 12 2.4 (2.0), 2.0
Table 3.
 
Factors Predicting Automated Center-Point Retinal Thickness Measurement Error
Table 3.
 
Factors Predicting Automated Center-Point Retinal Thickness Measurement Error
Parameter n % Automated Center-Point Retinal Thickness Measurement Error Mean (SD), Median P (Wilcoxon Rank Sum Test)
LACS
 0–1 (low) 37 9.0 (14.6), 3.2 P ≤ 0.024
 ≥2 (high) 13 21.1 (24.1), 13.8
CSES
 0–1 (low) 23 2.8 (3.3), 1.6 P ≤ 0.000005
 ≥2 (high) 27 19.5 (21.4), 11.9
Average signal strength (six- line scans)
 ≥7 (high) 25 15.2 (22.3), 4.6 P ≤ 0.60
 <7 (low) 25 9.1 (11.3), 5.1
ETDRS letter score
 <50 (poor acuity) 26 12.2 (19.3), 3.7 P ≤ 0.68
 ≥50 (good acuity) 24 12.1 (17.2), 4.7
Manual center-point thickness (μm)
 <250 (thin fovea) 22 14.9 (23.2), 3.9 P ≤ 0.85
 ≥250 (thick fovea) 28 10.0 (12.9), 4.9
Manual center-point thickness standard deviation/thickness ratio
 ≤0.1 (repeatable measures) 38 8.1 (11.0), 3.3 P ≤ 0.07
 >0.1 (variable measures) 12 25.0 (28.6), 16.4
Table 4.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Second Optimized OCT Scan Set
Table 4.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Second Optimized OCT Scan Set
Scan Set 1 Scan Set 2 Total
Low High
Low 8 2 10
High 7 20 27
Total 15 22 37
Table 5.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Third Optimized OCT Scan Set
Table 5.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Third Optimized OCT Scan Set
Scan Set 2 Scan Set 3 Total
Low High
Low 6 5 11
High 4 18 22
Total 10 23 33
HuangD, SwansonEA, LinCP, et al. Optical coherence tomography. Science. 1991;254:1178–1181. [CrossRef] [PubMed]
FungAE, LalwaniGA, RosenfeldPJ, et al. An optical coherence tomography-guided, variable dosing regimen with intravitreal ranibizumab (Lucentis) for neovascular age-related macular degeneration. Am J Ophthalmol. 2007;143:566–583. [CrossRef] [PubMed]
RayR, StinnettSS, JaffeGJ. Evaluation of image artifact produced by optical coherence tomography of retinal pathology. Am J Ophthalmol. 2005;139:18–29. [PubMed]
SaddaSR, WuZ, WalshAC, et al. Errors in retinal thickness measurements obtained by optical coherence tomography. Ophthalmology. 2006;113:285–293. [CrossRef] [PubMed]
DrexlerW, SattmannH, HermannB, et al. Enhanced visualization of macular pathology with the use of ultrahigh-resolution optical coherence tomography. Arch Ophthalmol. 2003;121:695–706. [CrossRef] [PubMed]
GloesmannM, HermannB, SchubertC, SattmannH, AhneltPK, DrexlerW. Histologic correlation of pig retina radial stratification with ultrahigh-resolution optical coherence tomography. Invest Ophthalmol Vis Sci. 2003;44:1696–1703. [CrossRef] [PubMed]
BlandJM, AltmanDG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet. 1986;1:307–310. [PubMed]
CsakyKG, RichmanEA, FerrisFL, III. Report from the NEI/FDA ophthalmic clinical trial design and endpoints symposium. Invest Ophthalmol Vis Sci. 2008;49:479–489. [CrossRef] [PubMed]
CostaRA, CalucciD, SkafM, et al. Optical coherence tomography 3: automatic delineation of the outer neural retinal boundary and its influence on retinal thickness measurements. Invest Ophthalmol Vis Sci. 2004;45:2399–2406. [CrossRef] [PubMed]
Figure 1.
 
An example of (A) segmentation error affecting the central 1-mm zone with (B) manual measurement of retinal thickness. An additional example of manual retinal thickness measurement (C) is also provided, in a case with a thickened HRB.
Figure 1.
 
An example of (A) segmentation error affecting the central 1-mm zone with (B) manual measurement of retinal thickness. An additional example of manual retinal thickness measurement (C) is also provided, in a case with a thickened HRB.
Table 1.
 
Distribution of Eyes with None to All of the Line Scans in a Scan Set Affected by Segmentation Error or Low Analysis Confidence
Table 1.
 
Distribution of Eyes with None to All of the Line Scans in a Scan Set Affected by Segmentation Error or Low Analysis Confidence
Number of Line Scans
None 1 2 3 4 5 6
Any SE 5 (10) 7 (14) 9 (18) 11 (22) 6 (12) 7 (14) 5 (10)
Central 1 mm SE 13 (26) 10 (20) 12 (24) 5 (10) 4 (8) 4 (8) 2 (4)
LAC message 29 (58) 8 (16) 5 (10) 4 (8) 3 (6) 0 (0) 1 (2)
Table 2.
 
Factors Associated with CSES
Table 2.
 
Factors Associated with CSES
Parameter n Mean CSES (SD), Median P (Wilcoxon Rank Sum Test)
LACS P ≤ 0.005
 0–1 (low) 37 1.6 (1.5), 1
 ≥2 (high) 13 3.5 (2.0), 4
Average signal strength (six line scans)
 ≥7 (high) 25 1.6 (1.6), 1 P ≤ 0.06
 <7 (low) 25 2.6 (2.0), 2.0
ETDRS letter score
 <50 (poor acuity) 26 1.8 (2.0), 1.5 P ≤ 0.57
 ≥50 (good acuity) 24 2.3 (2.0), 2.2
Manual center-point thickness (μm)
 <250 (thin fovea) 22 1.8 (1.5), 2.0 P ≤ 0.49
 ≥250 (thick fovea) 28 2.3 (2.1), 2.0
Manual center-point thickness standard deviation/thickness ratio
 ≤0.1 (repeatable measures) 38 2.0 (2.8), 1.8 P ≤ 0.48
 >0.1 (variable measures) 12 2.4 (2.0), 2.0
Table 3.
 
Factors Predicting Automated Center-Point Retinal Thickness Measurement Error
Table 3.
 
Factors Predicting Automated Center-Point Retinal Thickness Measurement Error
Parameter n % Automated Center-Point Retinal Thickness Measurement Error Mean (SD), Median P (Wilcoxon Rank Sum Test)
LACS
 0–1 (low) 37 9.0 (14.6), 3.2 P ≤ 0.024
 ≥2 (high) 13 21.1 (24.1), 13.8
CSES
 0–1 (low) 23 2.8 (3.3), 1.6 P ≤ 0.000005
 ≥2 (high) 27 19.5 (21.4), 11.9
Average signal strength (six- line scans)
 ≥7 (high) 25 15.2 (22.3), 4.6 P ≤ 0.60
 <7 (low) 25 9.1 (11.3), 5.1
ETDRS letter score
 <50 (poor acuity) 26 12.2 (19.3), 3.7 P ≤ 0.68
 ≥50 (good acuity) 24 12.1 (17.2), 4.7
Manual center-point thickness (μm)
 <250 (thin fovea) 22 14.9 (23.2), 3.9 P ≤ 0.85
 ≥250 (thick fovea) 28 10.0 (12.9), 4.9
Manual center-point thickness standard deviation/thickness ratio
 ≤0.1 (repeatable measures) 38 8.1 (11.0), 3.3 P ≤ 0.07
 >0.1 (variable measures) 12 25.0 (28.6), 16.4
Table 4.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Second Optimized OCT Scan Set
Table 4.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Second Optimized OCT Scan Set
Scan Set 1 Scan Set 2 Total
Low High
Low 8 2 10
High 7 20 27
Total 15 22 37
Table 5.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Third Optimized OCT Scan Set
Table 5.
 
Change in Central 1-mm Segmentation Error Score in Patients on the Third Optimized OCT Scan Set
Scan Set 2 Scan Set 3 Total
Low High
Low 6 5 11
High 4 18 22
Total 10 23 33
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×