| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
From the Computer Vision Laboratory and Department of Ophthalmology, Scheie Eye Institute, University of Pennsylvania, Philadelphia.
 |
Abstract |
|---|
 Top Abstract IntroductionMethodsResultsDiscussionReferences |
|---|
METHODS. Fluorescein angiographic images were selected from 5 patients enrolled in a clinical trial for which three follow-up visits were available. Thirty- and 600-second images were digitized at 1000 dots/in and registered (aligned) with polynomial warping algorithms. Custom-developed software allowed for coarse, automated identification of CNV. An easy-to-use graphical user interface facilitated supervision and refinement of the lesion boundaries by a skilled reader based on standard stereoscopic viewing of the fluorescein angiography study. Capabilities for boundary delineation in both early and late phases, and animation to allow for image correlation and evaluation of temporal changes in fluorescence of spatially corresponding pixels, were included. Two metrics for CNV characterization were generated. First, the lesion area based on the lesion boundaries was identified after supervision. Second, an integrated lesion intensity (ILI) reflecting the integrated, normalized lesion hyperfluorescence was calculated.
RESULTS. Area and ILI measures were calculated for each of 5 patients for three or more visits. Facile supervision based on the stereoscopic angiogram permitted arbitrarily close concordance with CNV identification using standard methods. Changes in area and ILI measurements between visits correlated closely with clinically observed changes in each case.
CONCLUSIONS. Interactive image processing permits efficient, accurate, computer-assisted CNV quantitation that may be useful for the support of clinical trials and preclinical studies.
|
Introduction |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Laser photocoagulation of CNV, as described by the Macular Photocoagulation Study Group (MPS), is the only treatment of long-term proven benefit.2 However, lesion geometry renders most lesions ineligible for treatment. Moreover, among eligible lesions, persistence and recurrence limits the efficacy of laser photocoagulation for preservation of visual function, whereas laser photocoagulation of subfoveal lesions often results in an immediate, irreversible loss of central vision. Accordingly, major efforts are being directed toward development of novel treatment and prophylaxis strategies, with attention focused on development, refinement, and evaluation of therapeutic options that more selectively obliterate new vessels while preserving functional retina. Encouraging results at 1 year were recently reported by the Treatment of AMD with Photodynamic Therapy (TAP) Study Group for treatment of subfoveal CNV in AMD with verteporfin-mediated photodynamic therapy.
Although visual function and quality of life are the major outcomes in clinical trials evaluating new approaches for treatment and prophylaxis of the complications of late AMD, fundus feature evaluation, characterization, and quantitation are routinely incorporated into trial design. These measures provide secondary outcomes that allow for improved interpretation of trial data, allow for subgroup analysis based on fundus feature characteristics, and permit exploration of the prognostic significance of baseline and follow-up fundus characteristics. In animal studies, where visual function testing is far more difficult, CNV characterization often serves as the primary study outcome.
Our group previously reported a robust computerized approach for identification and quantitation of macular drusen for the support of epidemiologic studies and clinical trials.3 We have also acknowledged the increased complexity and near intractability of the development of a completely automated approach for CNV identification4 but have made progress toward evaluating the spatiotemporal characteristics of fundus features in AMD based on fluorescein angiography image sequences.5 6
Current methods for CNV evaluation and quantitation in clinical trials differ little, if at all, from the original methods described by the MPS group in 1989.7 Frames of the angiogram are projected, and manual tracing with overlay of templates of known sizes permits rough estimation of lesion size. No prior effort has been made to capture a measure of the intensity of the CNV hyperfluorescence.
In this report, we describe a computer-assisted method for deriving measures of the area subtended by a CNV complex, and the fluorescence intensity associated with that lesion.
|
Methods |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Clinical evaluation of a routinely acquired fluorescein angiogram requires inspection of multiple frames of the angiography image sequence after dye administration. Ordinarily, "early-," "middle," and "late" phase frames are acquired, with much information derived by comparing early (on the order of 30 seconds after dye injection) and late (approximately 600 seconds) images for the eye in question. Clinical intuition suggests, and a limited number of studies confirm, that inspection of stereoscopic image pairs is required, particularly for judging the extent of occult CNV (Berger, unpublished observations, October 1999).8 While spatial extent of occult CNV is best determined by inspection of the late-phase frames, classic CNV fluoresces early, and is usually well-circumscribed in the early-phase frames.
We sought to develop a platform to allow for coarse computer identification of CNV, with extensive yet facile capabilities for supervision and refinement of computer-generated output. Acknowledging the difficulties and inadequacies of desktop personal computer stereoscopic rendering, particularly for subtle stereoscopic features often present in CNV lesions, we chose not to rely on monitor-based stereoscopic viewing. Instead, it is assumed that the observer will supervise the CNV identification process simultaneous with light-box inspection of the angiogram with a magnified stereoscopic viewer.
Angiography image sequence evaluation requires the observer to superimpose in his or her mind’s eye the individual frames of the fluorescein angiogram to judge the temporal fluorescence behavior of spatially corresponding areas. By comparing angiography frames, and correlating this information with stereoscopic inspection of the color fundus photographs, areas of depigmentation and geographic atrophy can be distinguished from CNV, for example.
To minimize the number of frames processed, yet allow for lesion boundary identification in both early- and late-phase frames, and to assist in the process of allowing correlation and comparison of early and late fluorescence at precisely corresponding regions, early- and late-phase frames of the angiogram are selected, digitized, and aligned.
Metrics
For the support of clinical trials, quantitative parameters
describing primary and secondary outcomes are defined, measured, and
correlated with study variables. For example, one might examine the
relationship between visual acuity, reading speed, and contrast
sensitivity as a function of treatment group or dose. An appropriate
metric for CNV should capture lesion size and, in accordance with
clinical principles and intuition, some measure of the activity or
leakage associated with a CNV complex. It appears reasonable, although
it remains to be proven, that hyperfluorescent lesion components
reflect more aggressive clinical behavior than blocking components. One
might further divide lesion size by measuring individual components of
the lesion such as blood, blocked fluorescence, and classic and occult
CNV. The designed software allows for facile lesion size and component
measurements.
An appropriate metric capturing both size and leakage of a CNV complex has not been previously explored with rigor. Intuitively, discernible differences in lesion fluorescence or brightness reflect differences in lesion activity. However, the absolute brightness of the lesion as measured on a frame of the fluorescein angiogram depends not only on dye leakage but also on other sources of variability, including camera flash intensity, film exposure, dye bolus timing, camera settings, and ocular light scattering. Accordingly, we attempted to normalize the lesion fluorescence intensity by a measure of background, specifically the median pixel brightness of the fundus background. The normalized pixel brightness was calculated by subtracting the median of the fundus background (those pixels not identified as part of the CNV lesion) from the raw pixel brightness. Although one might identify the optic disc and blood vessels (either automatically or manually) so as not to include these pixels that may be brighter than the true fundus background, the total number of pixels in these features is far less than those representing true background. So as to minimize the potential contribution of these pixels, the median rather than the mean is used for normalization.
The integrated lesion intensity (ILI), then, is the integral over all pixels identified as a component of the CNV complex, which are hyperfluorescent with respect to fundus background, of the pixel brightness in the 600-second frame of the fluorescein angiogram. Areas of blocked fluorescence will contribute to lesion area measurements. Alternately, these components will not contribute toward an integrated brightness measure. In addition, areas of scar, which are judged as inactive, may contribute to the area measurement but not to the ILI. Data are reported for the ILI and normalized ILI to explore the validity of these measures.
Case Selection
For development, evaluation, and validation, we selected 5
patients enrolled and followed in a clinical trial evaluating the
efficacy of a pharmaceutical agent for treatment and recurrence
prevention after laser treatment for CNV in AMD. Appropriate
institutional review board approval with adherence to the tenets of the
Declaration of Helsinki, and informed consent had been
obtained. Cases were selected to represent stability and
progression, and to contain the spectrum of lesion components including
classic and various forms of occult CNV, and blood and scar. Eyes with
at least three follow-up visits were selected to measure quantitatively
lesion progression to correlate with qualitative assessment of lesion
progression or stability.
Image Digitization and Processing
Thirty- and 600-second angiographic images for each eye at each of
the visits were digitized on a Nikon SuperCoolScan II at 1000 dots per
inch and stored as low-compression (~30:1) JPEG files. Previous
studies in our laboratory have demonstrated the equivalence of
uncompressed TIFF images and low-compression JPEG images for derivation
of qualitative and quantitative fundus feature data.9
The
scanned image files were then registered to align early- and late-phase
angiography images. Several approaches for image registration were
attempted (including correction for translation, rotation, and scale)
based on intensity correlation metrics. We have previously demonstrated
that highly accurate image registration sufficient to permit change
detection requires correction for nonglobal, warp-related image
misalignment.10
However, our automated registration
algorithms were not sufficiently robust to detect and align sometimes
subtle vessel landmarks as are present at 600 seconds after dye
injection. Accordingly, six corresponding points (usually vessel
bifurcations or crossings) were identified by the supervisor, and the
images were then registered by computer using polynomial warping
algorithms5
6
Highly accurate image registration allowed
for rendering as an animation to assist in evaluation of the temporal
change in fluorescence for spatially corresponding pixels in the
angiogram. Each pair-wise registration was accomplished in
approximately 2 minutes.
CNV Identification and Supervision
Identification of those pixels corresponding to CNV depends on
complex decisions based on temporal fluorescence behavior through the
phases of the angiogram. The various lesion components have very
different and sometimes highly variable fluorescence behavior.
Therefore, automated, precise lesion determination is a major
challenge. With only monoscopic early- and late-phase fluorescein
angiography frames available, precise lesion delineation is impossible
except in only those most straightforward cases.
Accordingly, only coarse efforts are directed toward computerized CNV determination. A region of analysis ("region of interest," ROI) is defined, and those pixels with a change in fluorescence above an arbitrary threshold are preliminarily identified as CNV.
Extensive supervision capabilities permit accurate computer-assisted CNV delineation. First, the ROI can be selected to confine attention to the area of known CNV. Second, a slider allows threshold adjustment to be concordant with supervisor inspection. Finally, simultaneous with stereoscopic magnified angiography review on a light box, the supervisor can add pixels not identified by the computer as CNV or remove pixels erroneously identified (Fig. 1) . The user is assisted by several features including the ability to draw on both early- and late-phase frames, the ability to animate the early- and late-phase frames to explore the temporal fluorescence behavior of spatially corresponding regions, and the ability to confirm identified pixels by animating the images with and without identified pixels.
|
|
Results |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
|
|
|
|
Case 4
Initially, a small area of classic CNV is noted just nasal to the
center of the macula. Inferiorly is an area of subretinal hemorrhage
occupying 0.3 disc areas, and a large area of occult ooze and blocked
fluorescence through and temporal to the center of the macula (Figs. 5A
5B
; Table 1
). At 4 and 8 months, the classic component of the lesion,
in particular, was noted to increase markedly in size (Figs. 5C
5D
5E
5F)
, accompanied by an increase in the normalized ILI out of proportion
to the increase in area. This change reflects the marked increase in
overall hyperfluorescence attributable to the disproportionate increase
in the more intensely hyperfluorescent classic component.
|
|
|
Discussion |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
In human trials, visual function (as measured by visual acuity, reading speed, contrast sensitivity, and other parameters) serves as the major outcome variable, but treatment effect is interpreted and correlated with effects on CNV. Characterization of CNV provides an important secondary outcome for these trials. In animal studies where measurement of visual function is more difficult, characterization of CNV assumes even greater significance.
The MPS investigators approximated lesion size by overlaying a template onto the appropriate frames of the fluorescein angiogram. An approximation of the area subtended by CNV could be derived and coarse judgments regarding change could be attempted. We sought to improve on current methods for CNV quantitation through the application of digital image processing methodologies.
In previous reports, we communicated our philosophy that a completely automated, computerized approach was of no value unless generated results were in arbitrarily close agreement with human grading.3 4 An automated, supervised approach incorporating, when necessary, superior human recognition skills, resulted in a highly useful and accurate tool for drusen identification, quantitation, and change detection. CNV quantitation is a far more challenging problem as described above. In previous work, we developed and evaluated nonglobal image registration algorithms to permit highly accurate image warping to produce a three-dimensional (two dimensions in space, and one dimension in time) angiogram image sequence vector.5 6 Characteristic temporal profiles for spatially corresponding fundus features of importance (including CNV, geographic atrophy, vascular structures, drusen, and background) were identified. It was soon recognized that the components of a CNV lesion exhibited highly variable temporal fluorescence behavior. In addition, there is extensive overlap in fluorescence behavior between lesion components and structures not part of the CNV complex. For example, active CNV may behave somewhat similarly when compared with retinal pigment epithelium atrophy or disciform scar, and blocked fluorescence may behave similarly when compared with background fluorescence or subretinal blood.
A number of groups have applied image analysis methodologies to fluorescein angiography image sequences. Phillips et al.17 projected early and late FA images onto a white screen and digitized these images with a CCD camera, frame grabber, and digitizer. Late- and early-phase frames were aligned manually during projection, and "leakage" was identified by inspecting for pixels with a positive brightness change. Saito et al.18 acquired scanning laser ophthalmoscope images and registered these images correcting for translational misalignment, only. In attempting to investigate the temporal fluorescence behavior of fundus features in AMD, these authors measured the maximum fluorescence in a ROI with time. Lack of precise image registration did not allow for pixel-resolution characterization of fluorescence behavior. Nagin et al.19 used a semiautomated global registration approach to investigate optic disc and peripapillary fluorescein filling rates. Data were reported for the average of 20 pixels reflecting the temporal behavior for the structures of interest. Importantly, inadequate registration accuracy and lack of efforts to incorporate supervision and refinement strategies have precluded widespread adoption of these early approaches.
We have demonstrated that nonglobal image registration algorithms are required to achieve the highly accurate registration required for image change analysis and image sequence analysis.6 10 Our methods permit pixel-resolution evaluation of fluorescence intensity change not previously possible with global registration techniques. Moreover, appropriate quantitation with correction for image brightness variability as might arise from a large number of factors has not been previously explored.
Acknowledging that autonomous CNV delineation is not possible except in the most straightforward lesion configurations (e.g., isolated, classic CNV), the supervisor must have access to the fluorescein angiography image sequence with capabilities for stereoscopic viewing. With this in mind, we have chosen to process the minimum number of frames that are likely to be helpful in the vast majority of cases.
Animation and flicker of early- and late-phase frames allows for correlation of spatially corresponding regions to help discriminate among possible features, most notably CNV and atrophy; the ability to outline portions of the lesion in the early phase, in which blood and classic CNV are most well-defined; and the ability to outline other portions of the lesion in the late phase in which occult components are evident. The two frames are not intended to capture and depict the information contained in the entire stereoscopic, angiographic image sequence but merely to provide a platform for coarse CNV identification followed by facile supervision and refinement of lesion boundaries based on the entire angiogram.
Cases were selected to represent a broad range of CNV configurations. The utility of the graphical-user-interface platform for area measurements is demonstrated in that the user could define lesion components in both early- and late-phase frames and capitalize on image correlation to investigate areas suspicious for CNV (as in case 3). These cases demonstrate that the platform allows for CNV lesion and component quantitation with facility and accuracy heretofore not possible. Importantly, supervision capabilities are native, allowing for arbitrarily close concordance with human grading.
The validity and utility of the intensity metrics are proposed but not proven rigorously. Calculations are presented for the cases presented to verify that the generated metrics are to some extent consistent with clinical impressions. It seems reasonable to speculate that a bright classic lesion might be more aggressive (potential for growth) and more visually significant (more leakage to influence visual function) than a less bright occult lesion, and that a brighter occult lesion may be more aggressive and significant than a subtle ooze, even if these lesions subtend the same area. This concept resonates with clinical experience. However, quantitative characterization of CNV lesion fluorescence has neither been explored rigorously nor evaluated formally.
In separate studies,20 we observed that intensity fluctuations for typical fundus features in routinely acquired digital and film-based fluorescein angiograms were on the order of 20%. Large horizontal camera movements as required for acquisition of stereo pairs and small camera movements toward or away from the eye may result in readily discernible brightness fluctuations, with intensity variability approaching 100%. We noted that large intensity fluctuations in hyperfluorescent lesions were almost always in phase with smaller amplitude fluctuations in vessels and background, and we are now attempting to derive and validate optimum normalization strategies based on these parameters.
It may be that highly accurate fluorescence intensity calculation and comparison requires protocol standardization during angiographic image acquisition. Variability in flashbrightness, camera sensitivity, camera position, photographic processing, dye bolus timing, ocular scattering, and the nonlinear relation between fluorescence intensity and pixel intensity values (particularly for brightly fluorescent regions) undermine our ability to measure accurately the lesion hyperfluorescence in true absolute units. Here, we present two such metrics that are calculated automatically after lesion delineation. The potential utility and prognostic significance of these measures await further assessment.
|
Acknowledgements |
|---|
|
Footnotes |
|---|
Submitted for publication November 17, 1999; revised January 18, 2000; accepted February 7, 2000.
Commercial relationships policy: N.
Corresponding author: Jeffrey W. Berger, Computer Vision Laboratory, Scheie Eye Institute, University of Pennsylvania, 51 North 39th Street, Philadelphia, PA 19104. jwberger{at}mail.med.upenn.edu
|
References |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  U. Chakravarthy, A. C. Walsh, A. Muldrew, P. G. Updike, T. Barbour, and S. R. Sadda Quantitative Fluorescein Angiographic Analysis of Choroidal Neovascular Membranes: Validation and Correlation with Visual Function Invest. Ophthalmol. Vis. Sci., January 1, 2007; 48(1): 349 - 354. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. M. Shah, S. Tatlipinar, E. Quinlan, J. U. Sung, H. Tabandeh, Q. D. Nguyen, A. S. Fahmy, I. Zimmer-Galler, R. C. A. Symons, J. M. Cedarbaum, et al. Dynamic and Quantitative Analysis of Choroidal Neovascularization by Fluorescein Angiography Invest. Ophthalmol. Vis. Sci., December 1, 2006; 47(12): 5460 - 5468. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
