Professional Documents
Culture Documents
Specular Microscopic Ancillary Study
Specular Microscopic Ancillary Study
Copyright
c Taylor & Francis Group, LLC
ISSN: 0271-3683 print / 1460-2202 online
DOI: 10.1080/02713680500536738
INTRODUCTION
Received 21 March 2005 The Specular Microscopy Reading Center (SMRC), which is the centralized
Accepted 13 December 2005
reading center for the Specular Microscopy Ancillary Study (SMAS), determined
Correspondence: Jonathan H. Lass, M.D.,
c/o Cornea Donor Study Coordinating
the central endothelial cell density (ECD) for eye bank baseline donor endothe-
Center, Jaeb Center for Health Research, lial images in the SMAS. Throughout the 10-year Cornea Donor Study follow-up
15310 Amberly Drive, Suite 350, Tampa,
FL 33647. Tel: (813) 975-8690; Fax: (813)
period, the SMRC will also evaluate clinical site postoperative images for par-
975-8761; E-mail: cds@jaeb.org ticipating patients, in order to assess the effect of donor age on endothelial cell
319
loss after penetrating keratoplasty. Methods specific to certify readers in the SMAS image quality classification
ECD determination for clinical images will be described system. Criteria for this image quality classification
in a later paper at the conclusion of the SMAS. are described in Figure 1. To be certified, readers were
The SMAS employed a central reading center be- required to correctly classify five out of six images
cause previous studies demonstrated significant vari- as analyzable or unanalyzable and four out of six
ability in both image quality assessment and ECD from analyzable images as excellent, good, or fair.
multiple readers.2,3 To provide the best estimate of
ECD, the SMRC developed, an image quality classi- ECD Determination
fication system, a dual-grading procedure, and an adju- The SMRC established a set of 25 representative
dication process. In this paper, we describe the SMRC eye bank images of varying analyzable image quality
certification procedures and the methods developed to (excellent to fair), varying ECD [high cell density
address and minimize the effect of intra- and inter- uniform across the field (2800–3400 cells/mm2 ), mod-
observer variability on image quality and ECD deter- erate cell density uniform across the field (2200–2800
mination for donor corneal endothelial images. To our cells/mm2 ), low cell density uniform across the field
Curr Eye Res Downloaded from informahealthcare.com by University of Southern California on 04/06/14
knowledge, this study marks the first time that this (1600–2200 cells/mm2 )], and various image media (dig-
methodology has been applied to ECD determination ital, VHS tapes and 35-mm negatives) to certify readers
in a multi-eye bank study. for ECD determination. Ten of the 25 images, derived
from various image media, fell into the low ECD
MATERIALS AND METHODS category, as these images were expected to yield greater
Staff Certification Procedures inter- and intra-observer variability due to potentially
greater endothelial polymegethism and pleomorphism.
The SMRC staff includes readers (CR, SE, LK).
An experienced technician established ECD ranges
a technical director/adjudicator (BB). and a medical
for each of the certification images by analyzing the
director (JL). After being trained, the readers and
images repeatedly on three separate days to determine
For personal use only.
FIGURE 1 The Specular Microscopy Reading Center corneal endothelial image classification. Analyzable—Excellent: All cell borders,
boundaries, and centers across a single image of the endothelium are distinct excluding the peripheral edges of the image. The single
image has a sufficient number of cells to count at least 50 and as many as 150 cells contiguous to each other. Good: A sufficient
number of distinct cell borders, boundaries, and centers from a single image of the endothelium to count at least 50 and as many as150
cells from variable frames encompassing a minimum of 15 cells contiguous to each other for each variable frame. Fair : A sufficient
number of cell borders, boundaries, and centers from a single image of the endothelium to count at least 50 cells from variable frames
encompassing a minimum of 15 cells contiguous to each other for each variable frame. The borders, boundaries and centers of up
to 25% of analyzed cells within the variable frames may be indistinct, but sufficient to estimate their location to conduct the analysis.
Unanalyzable—Less than 50 cells with distinct borders, boundaries, and centers from a single image of the endothelium to count from
variable frames encompassing a minimum of 15 cells contiguous to each other of each variable frame. The borders, boundaries and
centers of greater than 25% of potentially analyzable cells within the variable frames are indistinct, and therefore insufficient to estimate
their location to conduct the analysis.
FIGURE 2 Decision-making flow chart for the Specular Microscopy Reading Center corneal endothelial image classification.
The senior reader set the calibration for all eye banks traced and counted a minimum of three variable frames
by measuring all clearly focused hash marks to a min- and a maximum of six variable frames. Each reader at-
imum 100-µm and maximum 300-µm distance along tempted to count a minimum of 50 contiguous cells in
the X and Y directions. If measurements fell within 2% a single variable frame. In some cases, however, read-
of each of the measured distances, the reader saved the ers marked fewer than 50 cells in a given variable frame
calibration in the database, and the ECD determina- due to greater image magnification of some specular mi-
tion for all subsequent images from that eye bank were croscopes, lower ECD, or a variation of image quality
based on this calibration setting. If at any time in the within a given image. In those cases, the readers counted
For personal use only.
study the specular microscope at a SMAS participating multiple non-overlapping variable frames, with a min-
eye bank was serviced, repaired, or replaced, the SMRC imum of 3 frames of 15 cells each and a maximum of
required the eye bank to repeat the calibration process 6 frames, until they counted at least 50 total cells, as
before submitting additional study images. shown, for example, in Figure 1.
Images were flagged for adjudication when the ECD
Image Quality Classification
analyses varied by ≥5.0% between readers (Fig. 3). The
The two readers independently classified each image adjudicator independently determined the ECD of all
as unanalyzable or analyzable (excellent, good, or fair) flagged images while remaining masked to the readers
utilizing predefined criteria, sample images (Fig. 1), and gradings, and the adjudicator’s ECD was compared with
a decision-making flow chart (Fig. 2). If the two read- the ECDs of both readers. Final data consisted of the
ers differed on the image quality classification, or if average of all ECDs that were <5% different.
both readers classified an image as unanalyzable, the If the adjudicator’s ECD was not <5% of either of
image was flagged for adjudication (Fig. 3). The adju- the readers ECDs, the image was selected for final re-
dicator’s independent image quality classification was view. For these images, the adjudicator reviewed each of
considered final for all flagged records. the analyzed frames to confirm whether the frames were
drawn properly, whether the image was sampled prop-
ECD Determination erly across the image, and whether all cells were counted
The two readers independently determined the ECD accurately within each variable frame. If the adjudica-
for each donor image using the variable frame anal- tor judged readings to be accurate, those ECDs were
ysis method. To perform variable frame analysis, the accepted. Otherwise, the adjudicator rejected the read-
reader traced the boundary of the largest possible area ings and documented the reason for rejection. The final
of cells from the central endothelial specular image, ECD was the average of all ECDs from the accepted
incorporating a minimum of 15 and as many as 150 reading(s). If the adjudicator rejected all three analyses,
contiguous cells with distinct cell borders, boundaries, the data were disregarded and the image was returned
and centers within a variable frame (Fig. 1). The readers for a complete regrading.
FIGURE 3 The Specular Microscopy Ancillary Study dual grading and adjudication of eye bank images by the SMRC. RAD = relative
absolute difference calculated as [absolute value of (second – primary)/primary] or as [absolute value of (reader – adjudicator)/adjudicator].
using SAS version 8.2 statistical software (SAS Institute reader had originally categorized as fair. The adjudicator
Inc., Cary, NC, USA). classified 25 images as unanalyzable, agreeing with both
readers for 16 images and only one of the two readers
RESULTS for 9 images.
tics Inc., Portland, OR, USA (331 images submitted) class correlation of 0.88 (95% CI: 0.86, 0.90). The ECDs
CooperVision (no longer manufactured: 31 images sub- were <5% different from each other for 442 (69%) im-
mitted); HAI Laboratories, Inc., Lexington, MA, USA ages. The adjudicator independently determined the
(254 images submitted); Konan Inc., Phoenix, AZ, USA ECD for the remaining 203 images. Among the adjudi-
(61 images submitted); or Tomey, Phoenix, AZ, USA (11 cated images, the adjudicator’s ECD was <5% from the
images submitted). ECD determined by both readers for 44 (22%) images
and <5% different from one of the two reader’s as-
Image Quality Assessment sessments for 149 (73%) images. Ten (5%) images were
The two readers agreed on the image quality assess- selected for final review because the ECD determined
ment of 441 total images (64%); they categorized 424 by the adjudicator was ≥5% of both of the readers’ as-
images as analyzable and 17 as unanalyzable (Table 1). sessments. During final review, the adjudicator rejected
Image quality agreement was defined as the readers se- nine ECD assessments by a reader because of inaccurate
lecting the same grade. The observed agreement be- counts (i.e., missed or double-counted cells, seven im-
tween readers was fair, with a weighted kappa statistic of ages), improper frame construction (i.e., entire cell not
included in frame, one image), or poor sampling (i.e.,
TABLE 1 Comparison of Image Quality Classifications Between entire countable area not represented in frames, one im-
Readers age). None of the adjudicator’s ECD assessments were
Second reader rejected. Among images considered analyzable by only
one reader, the adjudicator’s ECD was <5% different
Primary reader Excellent Good Fair Poor from the ECD determined by the reader for all images
Excellent 21 30 1 categorized as analyzable by the adjudicator.
Good 12 244 164 ECD difference between readers increased with de-
Fair 14 159 6 creasing image quality; median 2% difference between
Poor 20 17 readers for excellent images, 3% for good images, and
Weighted kappa = 0.49 (95% confidence limits: 0.44, 0.54). 4% for fair images (p < 0.001). The relative absolute
75,000 to <90,000 95 3% (1%, 5%) 70 (74%) image by the same reader, where both assessments are analyzable.
b Defined as the absolute difference between the two assessments di-
µm2
≥90,000 µm2 119 2% (1%, 3%) 103 (87%) vided by their mean.
c Number of pairwise comparisons from duplicate readings.
Curr Eye Res Downloaded from informahealthcare.com by University of Southern California on 04/06/14
difference was <5% for 93% of the excellent images, number of cells that are counted3 ), errors inherent in
77% of good, and 56% of fair images (p < 0.001). There the method of ECD determination,7,8,13−21 or observer
were no donor cornea or slit-lamp characteristics identi- error that may result from intra-observer or interob-
fied that related to the difference in ECDs determined server variability.2,3,22−25 Factors unique to the donor
by the readers. A significant trend was observed with corneal endothelium that potentially influence spec-
a lower difference in ECD between readers when the ular image quality and reliability of ECD determina-
readers analyzed a greater cell area and counted more tion include cause of death, death to preservation time,
cells (p < 0.001 for both) (Table 2). state of the epithelium, degree of striae in the Descemet
membrane, and tissue temperature at the time of image
capture.8,26,27
Quality Control Images Given all of the variables that may affect ECD
Two readers made independent assessments on 78 determination, the SMAS was designed to provide the
individual images submitted between two and four most reliable ECD determination by (1) establishing
times for quality control purposes. There were 126 a central reading center with certified readers to
pairwise comparisons on images assessed more than address variability in training and methods for ECD
once by the same reader, and 122 were analyzable. determination by multiple eye banks; (2) using variable
The readers demonstrated good intra-observer agree- frame analysis with a defined approach for frame
ment on image quality (73% to 100% agreement for construction to address known errors with fixed frame
quality control images reviewed by the two readers analysis and difficulties of morphometric analysis with
with corresponding weighted kappa values of 0.56 to varying image quality7,8,28 ; (3) providing dual assess-
1.0). The most frequent difference in image quality ment of ECD with adjudication of all study images to
classifications among quality control images occurred address inter-observer variability; and (4) establishing
when a reader inconsistently classified images as good a retesting procedure by the Coordinating Center to
or fair. Differences in ECD for quality control images address intra-observer variability given the length of
ranged from 0% to 4% for one reader and 0% to 12% the enrollment period and the possibility of multiple
area of this population of cells.7,8,13−21 The majority ECD determination over an extended period of time by
of studies assessing donor ECD have utilized one eye the readers. This is particularly important with images
bank with donor specular images obtained and deter- of fair quality, in which sampling and frame construc-
mined by multiple technicians within that single eye tion issues become more apparent. Despite these fac-
bank.29−34 They employed fixed-frame or morphome- tors, SMAS readers achieved 73–100% quality criteria
tric manual approaches or a semiautomated approach agreement and 92% ECD agreement (<5% difference)
for ECD determination by multiple readers from a sin- with the quality control images.
gle image. These reports do not address the effect of This study demonstrates a reproducible and reliable
intra- or inter-observer variability on their results. A central reading center method for the assessment of
limited number of studies utilized a reading center for donor endothelial image quality and ECD, including
For personal use only.
determination of the donor ECD from multiple eye the use of certified readers, a defined image quality clas-
banks.28,35,36 sification scheme, dual grading and adjudication, and
The SMRC used the variable frame approach for the masked quality control procedures, all of which are crit-
SMAS, because in a previous study, 27 of the 105 donor ical for the conduct of large, multicenter, long-term eye
pairs (26%) could not be analyzed by a morphomet- bank studies. Future studies including image classifi-
ric approach due to poor image quality28 The variable cation of donor images should employ a dual grad-
frame approach for the SMAS was successful because ing and adjudication system. This recommendation is
only 25 of the 688 donor images (4%) were unanalyz- consistent with other studies of ocular images that rou-
able; if a morphometric approach had been utilized, tinely involve a dual grading and adjudication process
the 321 fair images would also have been excluded, and for interpretation.37,38 Future studies will determine if
half of the images would have been unanalyzable. reproducibility and reliability can be improved if eye
The value of the dual grading procedure with adju- bank image quality is improved and if more than one
dication for donor images, even with certified readers, donor image is available for central reading center eval-
was affirmed for both image quality classification uation.
and ECD determination. Despite a rigorous training
process and employment of the variable frame analysis
ACKNOWLEDGMENTS
method, the two readers’ figures differed by a clinically Technical assistance in the SMRC was provided by
significant amount (≥5%) for 203 of the 645 analyzable Shannon Edwards (SE) and Lori Karpinecz (LK). A
images (31%). Differences were more common in cases listing of the eye banks, eye bank directors and staff,
of poorer image quality. A similar finding of increased clinical centers, investigators and clinical center staff
difference with decreased image quality was observed who participated in the enrollment phase of the Cornea
in ECD determination between the SMRC final Donor Study has been previously published (see Ref.
adjudicated ECD and the eye bank ECD employing 1). This work was supported by a cooperative agreement
varying ECD determination methods and multiple with the National Eye Institute, National Institutes of
readers.27 Adjudication of the images found an ECD Health, U10 EY12728 and U10 EY12358. Additional
EJ, Cornea. Fundamentals of Cornea and External Disease, Vol. I. St. penetrating keratoplasty. Ophthalmology. 1998;105:1855–65.
Louis: Mosby; 1997:313–334. [30] Musch DC, Meyer RF, Sugar A. Predictive factors for endothelial cell
[8] Laing RA. Specular microscopy. In: Brightbill F. Corneal Surgery. The- loss after penetrating keratoplasty. Arch Ophthalmol. 1993;111:80–
ory, Technique, and Tissue. St. Louis: Mosby; 1999:101–112. 83.
[9] Doughty MJ. Toward a quantitative analysis of corneal endothelial [31] Lass JH, Reinhart WJ, Bruner WE, et al. Comparison of corneal stor-
cell morphology: A review of techniques and their application. Op- age in K-sol and chondroitin sulfate corneal storage medium in hu-
tome Vision Sci. 1989;66:626–642. man corneal transplantation. Ophthalmology. 1989;96:688–697.
[10] Sperling S, Gundersen HJ. The precision of unbiased estimates of [32] Lass JH, Reinhart WJ, Skelnik DL, et al. An in vitro and clinical com-
numerical density of endothelial cells in donor cornea. Acta Oph- parison of corneal storage with chondroitin sulfate corneal storage
thalmol. 1978;56:793–802. medium with and without dextran. Ophthalmology. 1990;97:96–
[11] Hirst LW, Yamauchi K, Enger C, et al. Quantitative analysis of wide- 103.
field specular microscopy. II. Precision of sampling from the cen- [33] Claerhout I, Beele H, Van Den Abeele K, Kestelyn P. Therapeutic pen-
tral corneal endothelium. Invest Ophthalmol Vis Sci. 1989;30:1972– etrating keratoplasty: clinical outcome and evolution of endothelial
1979. cell density. Cornea. 2002;21:637–642.
[12] American Academy of Ophthalmology. Corneal endothelial photog- [34] Lass JH, DeSantis DM, Reinhart WJ, et al. Clinical and morphometric
raphy. Ophthalmology. 1991;98:1464–1468. results of penetrating keratoplasty with one-piece anterior-chamber
[13] Modis L Jr, Langenbucher A, Seitz B. Corneal endothelial cell den- or suture-fixated posterior-chamber lenses in the absence of lens
sity and pachymetry measured by contact and noncontact specular capsule. Arch Ophthalmol. 1990;108:1427–1431.
microscopy. J Cataract Refract Surg. 2002;28:1763–1769. [35] Lass JH, Bourne WM, Musch DC, et al. A randomized, prospective,
[14] Landesz M, Siertsema JV, Van Rij G. Comparative study of three double-masked clinical trial of Optisol vs DexSol corneal storage
semiautomated specular microscopes. J Cataract Refract Surg. media. Arch Ophthalmol. 1992;110:1404–1408.
1995;21:409–416. [36] Lass JH, Gordon JF, Sugar A, et al. Optisol containing streptomycin.
[15] Vecchi M, Braccio L, Orsoni JG. The Topcon SP 1000 and Image- Am J Ophthalmol. 1993;116:503–504.
NET systems. A comparison of four methods for evaluating corneal [37] Maguire M. Complications of Age-Related Macular Degeneration
endothelial cell density. Cornea. 1996;15:271–277. Prevention Trial (CAPT). Baseline characteristics, the 25-item Na-
[16] Ohno K, Nelson LR, McLaren JW, et al. Comparison of record- tional Eye Institute Visual Functioning Questionnaire, and their as-
ing systems and analysis methods in specular microscopy. Cornea. sociations in the Complications of Age-Related Macular Degen-
1999;18:416–423. eration Prevention Trial (CAPT). Ophthalmology. 2004;111:1307–
[17] Benetz BA, Diaconu E, Bowlin SJ, et al. Comparison of corneal en- 16.
dothelial image analysis by SP8000 Konan non-contact and BioOp- [38] Early Treatment Diabetic Retinopathy Study Research Group.
tics BAMBI6 systems. Cornea. 1999;18:67–72. Grading diabetic retinopathy from stereoscopic color fundus
[18] Geroski DH, Edelhauser HF. Techniques for evaluating endothelial photographs—An extension of the modified Airlie House classifi-
cell function and cell loss. In: Brightbill FS. Corneal Surgery. Theory, cation. ETDRS Report Number 10. Ophthalmology. 1991;98:786–
Technique, and Tissue. St. Louis: Mosby; 1999:72–84. 806.