Acta Ophthalmologica 2008

Evaluating the effect of the new alignment algorithm for longitudinal series of Heidelberg retina tomography images
Ciara Bergin,1 David F. Garway-Heath2 and David P. Crabb1
1 2

Department of Optometry and Visual Science, City University, London, UK Glaucoma Research Unit, Moorelds Eye Hospital, London, UK

ABSTRACT. Purpose: To evaluate the impact of a new image-alignment algorithm on the repeatability of longitudinal measurements obtained from Heidelberg retina tomograph (HRT) images. Methods: HRTI and HRTII image series from 124 patients with glaucoma or ocular hypertension were made available from previously reported studies and were reprocessed with the old and new image-alignment algorithms. Improvements afforded by the new alignment algorithm were examined by considering statistically signicant improvement in repeatability of specic stereometric parameters (SP), namely rim area (RA), rim volume (RV), cup volume (CV) and cup shape measure (CSM). A further comparison was made by examining reduction in the variability of pixel-by-pixel height measures within image series. Results: In some HRT image series, the new algorithm automatically corrected obvious misalignment events that occurred with the previous algorithm. However, average improvement in repeatability of the SP in HRTI image series was not statistically signicant (P 0.13) and there was no statistically signicant reduction in pixel-by-pixel height measurement variability (P 0.73). In HRTII image series, there was evidence of improvement, on average, in the repeatability of some parameters (RA, P 0.01; RV, P 0.02; CSM, P 0.05), but not in CV (P 0.22). There was a large reduction in pixel-bypixel variability in HRTII image series (P < 0.001). Conclusion: There was no evidence to show that the new algorithm improved repeatability, on average, in HRTI images. However, the application of the new algorithm to HRTII image series marginally improved repeatability in stereometric measures and yielded a signicant reduction in pixel-by-pixel variability.
Key words: glaucoma image processing imaging optic nerve head repeatability

Introduction
Evaluating disease progression reliably is essential in the management of patients with established or suspected glaucoma. Glaucoma progression is detected by either (or both) a loss of function (usually detected by automated perimetry) or a change in structure of the optic nerve head (ONH) or retinal nerve bre layer (RNFL) surrounding the ONH (Strouthidis et al. 2006). Changes in structure, such as the thinning of RNFL or progressive damage to the ONH, indicate glaucoma progression; these changes can be measured by imaging technology such as scanning laser polarimetry, optical coherence tomography and scanning laser ophthalmoscopy (SLO). The Heidelberg retina tomograph (HRT; Heidelberg Engineering, Heidelberg, Germany) is an established SLO technology used for the examination of the ONH and surrounding RNFL (Zinser et al. 1989; Mikelberg et al. 1995; Chauhan 1996; Zangwill et al. 2004). HRT is a useful tool for diagnosis and is valuable in patient follow-up because it assesses the structural change of the ONH by considering a series of images acquired over time. There is evidence that structural changes to the ONH may occur before function is lost in some eyes (Kamal et al. 2000), whilst other evidence suggests that identiable functional and structural progression occurs almost independently (Artes &

Acta Ophthalmol. 2008: 86: 207214

2007 The Authors Journal compilation 2007 Acta Ophthalmol Scand

doi: 10.1111/j.1600-0420.2007.01035.x

207

Acta Ophthalmologica 2008

Chauhan 2005; Strouthidis et al. 2006). There is a general consensus that monitoring structural changes is useful to detect glaucomatous disease progression as early as possible (AGIS 1994; Bathija et al. 1998; Wollstein et al. 1998; Kwartz et al. 2005). Details of the HRT image acquisition process can be found elsewhere (Zinser et al. 1989; Alfonso et al. 2005); in short, each HRTI single scan acquires a three-dimensional stack of 32 confocal images of the ONH, to a depth of approximately 14 mm. The HRTII and HRT3 acquire a stack of 1664 confocal images, depending on scan depth. A three-dimensional topography image is formed by nding the peak reectance for each point (pixel) throughout the stack of twodimensional scans; this peak reectance measure is an estimate of the surface height at this position, thus yielding a topography image (a height map of the ONH and retinal surface). The technique used to form these images attempts to compensate for eye movement during acquisition and typically three single scans (of good quality) are taken in immediate succession. The topography image derived from each scan stack is averaged to form a mean topography image in which each pixel has a mean and standard deviation (SD) of surface height estimates. Analysis of a series of HRT images is performed clinically to help distinguish between progressing glaucoma patients and non-progressing patients. However, these images are not identically scaled versions of the true ONH and are affected by measurement variability during image acquisition (intrascan noise) and between the visits in follow-up (inter-scan noise). For example, scan resolution is limited by the optics of the human eye: the scanning laser must pass through the tear lm, cornea, lens and vitreous humour to reach the ONH; these media produce wave aberrations that degrade the resulting image (Artal et al. 2001) and add measurement noise. A number of other factors also contributes to the measurement noise including eye movements, pupil size (Zangwill et al. 1997), optic disc size (Iester et al. 1997), changes in intraocular pressure (IOP) (Nicolela et al. 2006), placement of the contour line (Orgu l et al. 1995) and the use of a reference plane in

generating stereometric parameters (Burk et al. 2000; Strouthidis et al. 2005b) and even the cardiac cycle (Chauhan & McCormick 1995). One method for reducing inter-test measurement noise involves processing the images so that they are correctly aligned in space to correct for eye movements that are more complex than simple horizontal and vertical displacements. Heidelberg Engineering has recently introduced a new type of alignment algorithm in the latest HRT software. The aim of this study was to examine the impact of this new alignment algorithm on the inter-scan variability in series of HRT images, focusing on the improvements in repeatability of global stereometric parameters (such as neuroretinal rim area measurements) and also on the variability in the estimated height measurements at each pixel in a series of images.

lar surgery. Images were acquired on two test sessions within a 6 week period, with three HRTI scans and three HRTII scans obtained at the rst session and two HRTI and two HRTII scans at the second session. This yielded a short series of HRTI data and a short series of HRTII data for each subject. These data are ideal for this study because we assume that no physiological change could have occurred between any of the images taken because of the short times between test and retest.
Long series

Materials and Methods

Data

The longitudinal HRT images series used in this study were obtained from previously reported studies (Strouthidis et al. 2005a). These studies adhered to the tenets of the Declaration of Helsinki, had local ethical committee approval and informed patient consent was obtained.
Short series

Longitudinal HRT series from all the patients reported by Strouthidis et al. (2005a) were retrospectively examined (Strouthidis et al. 2005b). Briey, this data set comprised ve images from each of 74 patients [43 ocular hypertensive (OHT); 31 primary open-angle glaucoma (POAG)]. OHT patients had IOP of 22 mmHg on two or more occasions and two initial reliable visual elds. POAG patients had visual eld defects quantied, as dened by AGIS (1994) scores of 1 on three consecutive reliable visual elds. The data were originally collected to evaluate the testretest variability of the 10 HRTI (or 256 256 pixels) images and 15 HRTII images (384 384 pixels). Patients were not excluded on the basis of ONH appearance but were excluded on the basis of myopia greater than 12 D of spherical power, or any history of intraocu-

Longitudinal HRT series data collected for a clinical trial of betaxolol against placebo in patients with OHT were also re-examined. This study, reported in detail elsewhere (Kamal et al. 1999, 2000), consisted of 216 patients who had HRTI images performed at 416 (median 10) visits over 28 (median 6) years . In the course of follow-up, 44 OHT patients developed early glaucomatous eld loss. This conversion to early glaucoma was dened on the basis of visual eld change by AGIS criteria. For the purpose of this study we randomly selected 50 non-converter patient series, making the assumption that these patients were less likely to have experienced signicant change in the appearance of their ONH. The followup for this sample of patients was 3.27.7 (median 6) years with imaging taken at 415 (median 9) visits. In this case, physiological change may have occurred because of ageing or other factors, but we attempt to correct for this in the methods.
New alignment algorithm

The new image-alignment algorithm is part of the new hrt3 proprietary software (version 3.0.2.0). This algorithm uses a landmarking technique that is derived from a general methodology developed by Capel (2004). Briey, the algorithm forms a homogenous linear translation, or map from one point in an image to a unique point in the next (meaning no other point gets mapped to this point). Image alignment is a problem of many variables and although these images appear two-dimensional they are a representation of a three-dimensional object (ONH). During image acquisi-

208

Acta Ophthalmologica 2008

tion, movement through six directions can occur: x, y and z displacements and rotation about all three of these axes. The method uses a novel algorithm, based on a numerical method, for solving the system of equations that are generated by this movement. In addition, the algorithm reduces the effect of noise by grading points using a principle of best tting, with the highest graded points within a predened distance grouped to form landmarks in every image. This algorithm then aligns these landmarks throughout the series of images. This is simply an outline of the methodology used by the alignment algorithm (Heidelberg Engineering, unpublished) but the exact computational detail is proprietary information.
Analysis

The same version of hrt heyex software (version 1.5.1.0) was installed on two separate computers, with the only difference being that one machine had the old alignment algorithm (software version 1.7.1.0) whilst the other had the new algorithm (software version 3.0.2.0). All the images and testing material were imported into both machines with the single topographies, the mean topographies and the progression analysis computed independently in each version. The contour lines were then redrawn by an expert observer on both machines and the stereometric parameters were then exported from the heyex software in the usual way. So-called .RAW les were also exported from the progression analysis menu at this point and these contained the aligned image information, in binary form, for each series. These data were transferred into an independent statistical program r, version 2.2.0 (The R Foundation for Statistical Computing, Vienna, Austria). In r, these binary data were redrawn back into the familiar images found in the heyex software. In r it was possible to extract estimates of topographic height at each pixel for analysis. To quantify the difference in the performance of the algorithms, we rst examined the stereometric analysis found in the hrt software. This stereometric analysis yields summary measures of structural properties of the ONH that are typically used to

aid diagnosis and can be monitored to detect progression. From the stereometric parameters, we selected rim area (RA), rim volume (RV), cup volume (CV) and cup shape measure (CSM). RA has been shown to be the most repeatable parameter and is a good candidate for identifying clinically meaningful change (the neural rim contains ganglion cell axons and distinguishes moderately well between normal and glaucomatous eyes) (Balazsi et al. 1984; Airaksinen et al. 1985; Wollstein et al. 1998; Strouthidis et al. 2005a). RV and CV are closely related but are more representative of the physiology of the ONH, and CSM is related to the steepness of the cup edges and, therefore, is considered by some to be meaningful to glaucomatous damage (Mikelberg et al. 1993; Vihanninjoki et al. 2000). These parameters were exported directly from the hrt heyex software using a modied report le to obtain a list in comma-delimited le format. The analysis of the repeatability of the stereometric parameters is best explained taking RA as an example. For the short series data, we computed the RA for each of the ve images. A repeatability score for RA for each series was taken simply as the SD of these ve values and was recorded with the old alignment algorithm and then separately with the new algorithm, with greater SD inferring worse repeatability. We examined the difference in repeatability graphically using a type of Bland & Altman (1986) plot. The plots were merely used as a device to show improvement (points above the line of unity) against non-improvement (points below the line of unity). We then used a paired Wilcoxon nonparametric test to see if the number of image series that demonstrated improved repeatability (above the line) was beyond what would be expected by chance alone. For the analysis of the short series, we assumed there to be no physiological or structural change. In the long series, however, we needed to account for potential change because there might have been some structural change in the ONH because of disease or ageing effects. Therefore, with the long series data, we computed the RA for each of the available images and then plotted these against time. We removed any

trend (because of potential structural changes or ageing effects) by tting a simple linear regression and taking the SD of the residuals as a repeatability score. We then proceeded as with the short series data. This analysis was then performed separately for the four stereometric parameters. We also evaluated potential improvement in measurement repeatability at a pixel-by-pixel level. Each pixel value is representative of the topographic height at that position in the image. By considering one pixel position of the mean topography images throughout a series, we had several values for the topographic height at that position. Because these images have been aligned, these should be several measures of the surface height at exactly the same location in the ONH. The SD of these topographic height measurements is considered to be a measure of variability or measurement noise at that pixel. Each mean topography image is a matrix of topographic height measurements and for any given series there is corresponding SD for each entry in the matrix; these SDs form a distribution that is representative of the noise in the series. Fig. 1 shows a series with a clear alignment improvement and the corresponding distribution of SDs. In this example, the new alignment algorithm greatly and obviously reduced the spread of the distribution. Note, the rotational alignment sometimes gave blank areas in the margin of the image: all pixels in these areas were considered missing values in the analysis. By using this pixel-by-pixel variability method, we established an alternative measure for the inter-series noise. We applied this method to the short series images only, where it is assumed that no structural change could have occurred between scans, obtaining a measure for the noise within the series as processed with both the old and the new algorithms. A comparison between the performances of the algorithms across all the subjects was made using a type of BlandAltman plot as before: the points (subjects) above the line of unity indicated a reduction in mean SD of topographic heights (noise) throughout a series with the new alignment algorithm as compared to the old algorithm; points below the line represented a worsening

209

Acta Ophthalmologica 2008

in repeatability. The statistical signicance of the difference in variability was tested as before with the paired Wilcoxon non-parametric test.

Results
Some patient series demonstrated obvious and marked improvements in image alignment when using the new
(A) Baseline exam

algorithm as compared to the old version (Figs 1 and 2). Despite these cases, where the new algorithm automatically corrects for obvious misalignment, the overall effect on the repeatability of global stereometric parameters was less marked. There was no evidence of average improvement in the repeatability of stereometric parameters in any of the HRTI long or short series data (see
Followup exams

Fig. 3). In the short HRTII series, there was some evidence of an average improvement in RA (P = 0.006), RV (P = 0.024) and CSM (P = 0.054), but not in CV (P = 0.218). The lack of improvement in repeatability of images with the new alignment algorithm when images are acquired by the HRTI was also apparent with the pixel-by-pixel analysis (Fig. 4). Conversely, there

Topography 13/Feb/1996 (136)

Topography 18/Aug/1997 (137)

Topography 02/Feb/1998 (138)

Topography 01/Mar/1999 (139)

Topography 21/Jun/1999 (140)

Correction Baseline exam Followup exams

Misalignment event

Topography 13/Feb/1996 (153)

Topography 18/Aug/1997 (154)

Topography 02/Feb/1998 (155)

Topography 01/Mar/1999 (156)

Topography 21/Jun/1999 (157)

Underlying distribution of variability

(B)

Computed with the old algorithm

Mean SD = 99.8 8000 10 000 12 000

(C)

Computed with the new algorithm

Mean SD = 39.3

Number of pixels

6000

4000

2000

50

100

150

200

250

300

0 0

2000

4000

6000

8000 10 000 12 000

50

100

150

200

250

300

The magnitude of the SD per pixel of inter-visit images in microns

Fig. 1. (A) The same short image series processed with the old algorithm (top line) and the new algorithm (bottom line). There is an obvious improvement with the new algorithm, highlighted in the fourth image in the series. The corresponding distributions of standard deviation (SD) of the pixel-by-pixel height measurements for the series as calculated with the old and new algorithms respectively are shown in (B) and (C).

210

Acta Ophthalmologica 2008

(A)

Baseline exam

Followup exams

Topography 26 Mar 1996 (129)

Topography 20 Aug 1997 (130)

Topography 18 Feb 1998 (131)

Topography 02 Mar 1999 (132)

Topography 06 July 1999 (133)

Baseline exam

Followup exams

Topography 26 Mar 1996 (156)

Topography 20 Aug 1997 (157)

Topography 18 Feb 1998 (154)

Topography 02 Mar 1999 (155)

Topography 06 July 1999 (154)

(B)

Fig. 2. (A) An example of an alignment correction, which is quite signicant but initially hard for an observer to detect; crosshairs denoting the centre of the image make it easier to detect the improvement (for example, in the position of the fourth image in the bottom row of the newly aligned series). For this subject, the improvement was conrmed by a reduction in the pixel-by-pixel variability. (B) The algorithm seemed to have a signicant impact on an artefact commonly referred to as image doubling. Again, this manifested a signicant reduction in pixel-by-pixel variability in this subject. Nevertheless, instances of this effect were quite rare in the sample analysed as whole.

was a considerable improvement, using the same analysis, in average repeatability of the HRTII image series (P = 0.001). The mean pixel height standard deviation (MPHSD) metric reported by the hrt software gives an indication of the quality of each individual image that constitutes a series. Table 1 shows the distribution of this MPHSD metric for all the images in the HRTI

and HRTII short series. Generally, and independent of the effect of the alignment algorithm, the individual HRTII image quality was better than that of the HRTI series.

Discussion
The HRT is an example of a clinically useful imaging device that aids the diagnosis of glaucoma and the

monitoring of structural glaucomatous deterioration over time. However, the images that need to be quantied and interpreted are still subject to measurement noise, including the movements during image acquisition that displace the image in a complicated fashion. It is therefore desirable, certainly when examining series of images from the same patient over time, to make use of image-processing techniques that attempt to align the images in the series so that estimates of the exact position at each pixel remain constant over the follow-up. The developers of the HRT have recently introduced an image-alignment algorithm based on landmarking techniques previously used in face recognition and satellite tracking in an attempt to achieve this goal (Capel 2004). In this study, we assessed the impact of this algorithm on previously acquired longitudinal series of HRT data. In some individual cases, the effect of the algorithm on series of images was very clear, with cases of obviously misaligned images automatically registered with the rest of the series. When we examined the improvement in repeatability of the stereometric parameters commonly used to interpret series of images, any improvement afforded by the new algorithm across a large sample of series was less apparent. There was, however, a signicant improvement in average repeatability of RA, RV and CSM measurements when we considered series of HRTII data. Whilst statistical signicance was achieved in average improvement, it is pertinent to note that this meant repeatability of the measures worsened in some cases. In addition, only global parameters, rather than sectoral values, were considered in this study. There are currently two broad statistical approaches to detect changes in series of HRT images over time: the rst involves summary measures for example, the neural rim area (RA) being tracked over time. For the HRT, RA has been shown to be one of the most reproducible parameters and is a good candidate for identifying clinically meaningful change (Strouthidis et al. 2005b). Other methods act on the actual height estimates at the pixels themselves; this is exemplied by the topographic change analysis (TCA), which is incorporated

211

Acta Ophthalmologica 2008

Rim area
0.10 0.04

Rim volume
0.04

Cup volume
0.04 P = 0.322 0.00 0.10 0.20 0.30 0.04 0.00

Cup shape measure

Short series HRTI

0.00

0.05

0.00

0.10 0.05

0.04

P = 0.359 0.00 0.10 0.20 0.30

P = 0.852 0.00 0.10 0.20 0.30

0.04

0.00

P = 0.605 0.00 0.10 0.20 0.30

Difference in standard deviation (Old New)

0.10

0.04

0.04

Short series HRTII

0.00

0.05

0.00

0.00

0.10 0.05

0.04

0.04

P = 0.007 0.00 0.10 0.20 0.30

P = 0.025 0.00 0.10 0.20 0.30

P = 0.191 0.00 0.10 0.20 0.30

0.04

0.00

0.04

P = 0.054 0.00 0.10 0.20 0.30

0.10

0.04

0.04

Long series HRTI

0.05

0.00

0.00

0.10 0.05

0.00

0.04

0.04

P = 0.064 0.00 0.10 0.20 0.30

P = 0.158 0.00 0.10 0.20 0.30

P = 0.125 0.00 0.10 0.20 0.30

0.04

0.00

0.04

P = 0.598 0.00 0.10 0.20 0.30

Average in standard deviation between old and new

Fig. 3. BlandAltman-type plots summarizing the differences in repeatability of stereometric parameters between the old and new alignment algorithms. If the standard deviation (SD) of the stereometric parameters were identical then all the symbols would fall on the line of unity. Improvements in repeatability manifested by the new algorithm are highlighted by symbols above the line of unity. The rst row of plots contains all the results for HRTI images from the short series images; the second and third rows show the results from the HRTII short series and the HRTI long series, respectively. The rst column indicates the performance on RA in each of these series types. The second, third and fourth columns contain the plots for rim volume (RV), cup volume (CV) and cup shape measure (CSM), respectively. The corresponding P value is the result of the statistical test (paired Wilcoxon non-parametric test) of the number of subjects showing improved repeatability against those not showing improved repeatability.

into the hrt software (Chauhan et al. 2000), and a more recent approach, adopted from quantitative methods applied to MRI images, called statistic image mapping (SIM) where the signicance of change at each pixel is determined from the rate of surface height change and the measurement variability inherent in each patients series of images (Patterson et al. 2005). Common to all these techniques, however, is the principle of

separating clinically relevant effects (signal) from measurement variability (noise) in the images. The new image-alignment algorithm examined in this study is an example of an attempt to reduce this noise. Despite some cases where obvious misalignment is clearly corrected, the ndings suggest that the new alignment algorithm has, on average, little impact when applied to HRTI images. However, the stereometric parameters were

generally more repeatable when the alignment algorithm was used with HRTII images. Furthermore, the greatest gain in repeatability in HRTII images was in the estimation of the height at the pixel level. This result may be of great benet to progression algorithms that detect change at the pixel level, such as TCA- or SIM-type analyses. Different pre- and post-processing techniques have previously been used

212

Acta Ophthalmologica 2008

Pixel variability of 74 HRTI short series images

10

P = 0.73

0
10

20

40

60

80

100

Pixel variability of 74 HRTII short series images

P = 0.001

20

40

60

80

100

Average in standard deviation between old and new

Fig. 4. BlandAltman-type plots summarizing the differences in repeatability of pixel-by-pixel topographic height between the old and new alignment algorithms. The rst and second rows show pixel-by-pixel variability results for image series in HRTI and HRTII formats, respectively. Those points above the line of unity show a reduction in variability when calculated with the new algorithm. Table 1. Image quality (mean pixel height standard deviation in lm) for the individual images that make up the short series HRTI and HRTII data set. Range 010 1020 2030 3040 4050 5060 HRTI (%) 3 126 102 51 34 13 (1) (34) (28) (14) (9) (3) HRTII (%) 12 190 72 38 26 13 (3) (52) (20) (10) (7) (3)

HRT, Heidelberg retina tomograph.

to improve the repeatability of HRT images. In pre-processing, the selection of a standard reference plane in 2000 (Burk et al. 2000) allowed for automatic determination of intrapapillary variables once a disc border contour line had been dened. The use of the 320 lm reference plane was shown to confer a signicant improvement to repeatability (Strouthidis et al. 2005a). In post-processing, the introduction of a new alignment algorithm in the hrt software in 2000 was shown to signicantly improve the image quality of mean topographies (Burk & Rendon 2001). Manual alignment and contour lines drawn by the operator are not only time-consuming but are operatordependent and propagate noise: quantitative methods that are independent

of the contour line have been introduced in the hrt3 software (Swindale et al. 2000; Swindale & Volz 2005; Coops et al. 2006) and may reduce subjective, observer-derived, variability. The new automated landmarking alignment algorithm investigated in this study is an alternative to the manual alignment, and it has similar or better performance compared to previous versions, with no obvious misalignments where these had previously occurred. Other post-processing work focusing on the confocal stacks that make up the HRT images show promise in reducing measurement variability (Patterson et al. 2006). The new alignment algorithm appears to be dependent on the image dimensions: on larger images (and wider eld of view), the number of landmarks used by the algorithm was increased (Capel 2004). As the number of landmarks increased, the t improved between images. HRTII images (384 384 pixels) are more than twice the area size of HRTI images (256 256 pixels). Therefore, HRTII images contain approximately three times as many landmarks as HRTI images (Heidelberg Engineering, unpublished). This may explain the difference in the algorithms performance between the HRTI and HRTII images

observed in this study. It is important to note that the quality of an image probably does have an effect on the amount of improvement possible. That is, a series that contains low-quality or poorly aligned images is more likely to benet from improved alignment. However, if we consider the MPHSD of the individual images (Table 1), then the HRTII image series sample generally had better image quality than the comparable HRTI image series, yet we only found improvements in HRTII data. So while image quality is an important issue it doesnt explain the difference in performance of the alignment algorithm between the image series types we have in seen in this study. A prospective study looking at the performance of the alignment algorithm on consecutive patients and relating improvement to original image quality may be useful. Any assessment of repeatability may be masked by other factors that contribute to the overall measurement variability, such as the placement of the contour line, or rotational alignment. The fact that RA and RV are the only parameters to show overall improvement is reasonable because these parameters have been shown to be the most repeatable of the stereometric parameters (Strouthidis et al. 2005b). In summary, the new image-alignment algorithm for longitudinal HRT data markedly improved some series, as it was primarily designed to do, but it had no impact on the average repeatability of stereometric parameters in HRTI images. There was some improvement, on average, in global RA repeatability in series of HRTII images but the most signicant improvement occurred in the reduction of pixel-by-pixel variability in HRTII images. The results from this study suggest that the new alignment algorithm may have a signicant impact on progression detection methods for the HRT that are based on the topographic heights at pixel and super-pixel level, such as TCA and SIM, and this is the subject of further research.

Difference in standard deviation (Old New)

10

10

Acknowledgements
The authors thank Nick Strouthidis for collecting the HRT data and redrawing the contour lines; Aachal Kotecha, who also helped with the

213

Acta Ophthalmologica 2008

redrawing of the contour lines; and Andrew Patterson for his help with software and technical support. This work was presented in part as a poster at the Association for Research in Vision and Ophthalmology (ARVO) in Fort Lauderdale, Florida, USA, May 2006. The work is supported in part by funding from the Moorelds Special Trustees.

References
AGIS (1994): The advanced glaucoma intervention study (AGIS): 1. Study design and methods and baseline characteristics of study patients. Control Clin Trials 15: 299 325. Airaksinen PJ, Drance SM, Schulzer M et al. (1985): Neuroretinal rim area in early glaucoma. Am J Ophthalmol 99: 14. Alfonso AC, Ciof GA, Garway-Heath DF, Girkin CA, Hudson C, Johnson CA, Weinreb RN & Zangwill LM (2005): The essential HRT primer. San Ramon, California: Jocoto Advertising, Inc. Artal P, Guirao A, Berrio E & Williams DR (2001): Compensation of corneal aberrations by the internal optics in the human eye. J Vis 1: 18. Artes PH & Chauhan BC (2005): Longitudinal changes in the visual eld and optic disc in glaucoma. Prog Retin Eye Res 24: 333354. Balazsi AG, Drance SM, Schulzer M & Douglas GR (1984): Neuroretinal rim area in suspected glaucoma and early chronic open-angle glaucoma. Correlation with parameters of visual function. Arch Ophthalmol 102: 10111014. Bathija R, Zangwill L, Berry CC, Sample PA & Weinreb RN (1998): Detection of early glaucomatous structural damage with confocal scanning laser. J Glaucoma 7: 121 127. Bland JM & Altman DG (1986): Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1: 307310. Burk RO & Rendon R (2001): Clinical detection of optic nerve damage: measuring changes in cup steepness with use of a new image alignment algorithm. Surv Ophthalmol 45: S297S303. Burk RO, Vihanninjoki K, Bartke T, Tuuloren A, Airaksinen PJ, Vo lcker HE & Ko nig JM (2000): Development of the standard reference plane for the Heidelberg retina tomograph. Graefes Arch Clin Exp Ophthalmol 238: 375384. Capel D (2004). Image mosaicing and superresolution, 1st edn. London: SpringerVerlag. Chauhan BC (1996): Confocal scanning laser tomography. Can J Ophthalmol 31: 152 156. Chauhan BC & McCormick TA (1995): Effect of the cardiac cycle on topographic measurements using confocal scanning

laser tomography. Graefes Arch Clin Exp Ophthalmol 233: 568572. Chauhan BC, Blanchard JW et al. (2000): Technique for detecting serial topographic changes in the optic disc and peripapillary retina using scanning laser tomography. Invest Ophthalmol Vis Sci 41: 775782. Coops A, Henson DB, Kwartz AJ & Artes PH (2006): Automated analysis of Heidelberg retina tomograph optic disc images by glaucoma probability score. Invest Ophthalmol Vis Sci 47: 53485355. Iester M, Mikelberg FS & Drance SM (1997): The effect of optic disc size on diagnostic precision with the Heidelberg retina tomograph. Ophthalmology 104: 545548. Kamal DS, Viswanathan AC, Garway-Heath DF, Hitchings RA, Poinoosawmy D & Bunce C (1999): Detection of optic disc change with the Heidelberg retina tomograph before conrmed visual eld change in ocular hypertensives converting to early glaucoma. Br J Ophthalmol 83: 290294. Kamal DS, Garway-Heath DF, Hitchings RA & Fitzke FW (2000): Use of sequential Heidelberg retina tomograph images to identify changes at the optic disc in ocular hypertensive patients at risk of developing glaucoma. Br J Ophthalmol 84: 993998. Kwartz AJ, Henson DB, Harper RA, Spencer AF & McLeod D (2005): The effectiveness of the Heidelberg retina tomograph and laser diagnostic glaucoma scanning system (GDx) in detecting and monitoring glaucoma. Health Technol Assess 9: 1148. Mikelberg FS, Wijsman K & Schulzer M (1993): Reproducibility of topographic parameters obtained with the Heidelberg retina tomograph. J Glaucoma 2: 101103. Mikelberg FS, Partt CM, Swindale NV, Graham SL, Drance SM & Gosine R (1995): The ability of the Heidelberg retina tomograph to detect early glaucomatous visual eld loss. J Glaucoma 4: 242247. Nicolela MT, Soares AS, Carrillo MM, Chauhan BC, LeBlanc RP & Artes PH (2006): Effect of moderate intraocular pressure changes on topographic measurements with confocal scanning laser tomography in patients with glaucoma. Arch Ophthalmol 124: 633640. Orgu l S, Kaiser HJ, Flammer J & Gasser P (1995): Systemic blood pressure and capillary blood-cell velocity in glaucoma patients: a preliminary study. Eur J Ophthalmol 5: 8891. Patterson AJ, Garway-Heath DF, Strouthidis NG & Crabb DP (2005): A new statistical approach for quantifying change in series of retinal and optic nerve head topography images. Invest Ophthalmol Vis Sci 46: 16591667. Patterson AJ, Garway-Heath DF & Crabb DP (2006): Improving the repeatability of topographic height measurements in confocal scanning laser imaging using maximum-likelihood deconvolution. Invest Ophthalmol Vis Sci 47: 44154421. Strouthidis NG, White ET, Owen VMF, Ho TA, Hammond CJ & Garway-Heath DF

(2005a): Factors affecting the testretest variability of Heidelberg retina tomograph and Heidelberg retina tomograph II measurements. Br J Ophthalmol 89: 14271432. Strouthidis NG, White ET, Owen VMF, Ho TA & Garway-Heath DF (2005b): Improving the repeatability of Heidelberg retina tomograph and Heidelberg retina tomograph II rim area measurements. Br J Ophthalmol 89: 14331437. Strouthidis NG, Scott A, Peter NM & Garway-Heath DF (2006): Optic disc and visual eld progression in ocular hypertensive subjects: detection rates, specicity, and agreement. Invest Ophthalmol Vis Sci 47: 2904. Swindale NV & Volz D (2005): Automated identication of early glaucoma with the Heidelberg retina tomograph. The Association for Research in Vision and Ophthalmology, Inc. Arvo Abstract Invest Ophthalmol Vis Sci 2005; 46: E-Abstract 2528. Swindale NV, Stjepanovic G, Chin A & Mikelberg FS (2000): Automated analysis of normal and glaucomatous optic nerve head topography images. Invest Ophthalmol Vis Sci 41: 17301742. Vihanninjoki K, Teesalu P, Burk ROW, La a ra E, Tuulonen A & Airaksinen PJ (2000): Search for an optimal combination of structural and functional parameters for the diagnosis of glaucoma. Graefes Arch Clin Exp Ophthalmol 238: 477481. Wollstein G, Garway-Heath DF & Hitchings RA (1998): Identication of early glaucoma cases with the scanning laser ophthalmoscope. Ophthalmology 105: 15571563. Zangwill L, Irak I, Berry CC, Garden V, de Souza Lima M & Weinreb RN (1997): Effect of cataract and pupil size on image quality with confocal scanning laser ophthalmoscopy. Arch Ophthalmol 115: 983990. Zangwill LM, Weinreb RN, Berry CC et al. (2004): Racial differences in optic disc topography baseline results from the confocal scanning laser ophthalmoscopy ancillary study to the ocular hypertension treatment study. American Medical Association. Arch Ophthalmol 122: 2228. Zinser G, Wijnaendts van Resandt RW, Dreher AW, Weinreb RN, Harbarth U, Schroeder H & Burk RO (1989): Confocal laser tomographic scanning of the eye. Proc. SPIE 1161: 337340.

Received on January 30th, 2007. Accepted on July 14th, 2007. Correspondence: David Crabb Department of Optometry and Visual Science City University London EC1 V 0HB UK Tel: +44 20 7 0400191 Fax: +44 20 7 0408494 Email: d.crabb@city.ac.uk

214