Quantitative Fluorescein Angiographic Analysis of Choroidal Neovascular Membranes: Validation and Correlation with Visual Function

Abstract

purpose. To compare computerized analysis with traditional grading methods in the analysis of fluorescein angiograms from patients with choroidal neovascularization (CNV) due to age-related macular degeneration (AMD) and to examine the clinical relevance of parameters generated by computerized analysis by testing their relationships with clinical measures of vision.

methods. Custom quantitative fluorescein analysis (QFA) software was used to analyze 62 angiograms from patients with CNV for whom distance visual acuity (DVA) data were available. On applying QFA, we obtained three mathematical parameters for each lesion component: pixel area (PA), integrated intensity (II), and positive fluorescence (PF). Quotients (Q) were derived for the latter two parameters by correcting against background (b) or optic nerve (o) fluorescence (IIQ^b, IIQ°, PFQ^b, and PFQ°). The new metrics were compared with traditional grading parameters of classic CNV and lesion area. The relationships of both sets of angiographic data with measures of vision were explored by regression analyses.

results. Weighted κ between QFA and traditional grading for lesion subtype assessment was high (κ = 0.7). Regression analyses with PA, IIQ^b, IIQ°, PFQ^b, and PFQ° for each lesion descriptor (leakage, classic CNV, occult CNV, total lesion) as independent variables and DVA as the dependent variable showed that in every case PFQ^b exhibited the most significant relationship with vision (adjusted r ² = 0.26). Parameter estimates showed that for a change of 30 units on the PFQ^b for classic CNV, a loss of 20 letters of DVA may be expected. No parameters from traditional grading methods showed statistically significant relationships with DVA.

conclusions. The markers of dynamic change in area and intensity of fluorescence exhibited stronger relationships with visual function than did area measurements alone.

The scientific literature contains many reports on the interpretation of fundus fluorescein angiograms to diagnose and classify choroidal neovascularization (CNV), a devastating complication of age-related macular degeneration (AMD). Clinically, the presence of CNV is diagnosed when an exudative lesion is seen in the macular region of the posterior fundus of the eye. ¹ There is much interest in the quantification of the angiographic features of CNV, as these parameters are used as markers for monitoring a therapeutic response. To date, the techniques used in angiographic analysis are based on subjective interpretation by experienced clinicians or trained graders of the patterns of fluorescence depicted in the angiographic sequences. ^{2 3 4} Normal and abnormal patterns of fluorescence have been described, and the latter are considered to represent pathologic processes in specific tissues or compartments within the exudative lesion. Notably, abnormal spatial localization of hyperfluorescence and temporal alterations in patterns of hyper- and hypofluorescence are used to describe the presence of leaking new choroidal vessels. ^{2 5} Quantification of the many components of an exudative lesion was usually accomplished by analyzing single angiographic frames with graded categorical methods (MPS circles) or through the application of image analysis to angiographic images. ^{2 5 6} Although these methods take into account the temporal changes in fluorescence, the rate of area change and the alterations in the intensity of fluorescence, which are likely to be important markers for severity of leakage from the abnormal vasculature, are not routinely taken into consideration in traditional angiographic grading. ^{2 7 8 9 10 11}  

Recent marked improvements in computational power coupled with advances in imaging and digital photography have revolutionized the field of retinal imaging. This has permitted both easier access and improved ability to manipulate large volumes of data, thus enabling more innovative approaches to be used in describing fluorescein angiographic parameters. 

The objective of the present study was to compare the grading output from newly developed custom software (computerized analysis) with that of traditional grading and to assess the clinical relevance of the former by examining their relationships with measures of vision. 

Materials and Methods

The image database in the Royal Group of Hospitals Belfast was the source of the angiograms. The database, which was compiled between January 1996 and March 2004, contained the angiograms of the fundus from more than 1000 patients with a clinical diagnosis of exudative AMD. The research was approved by the Research Ethics Committee of the Queen’s University of Belfast and adhered to the guidelines set forth in the Declaration of Helsinki. 

For the purposes of the present study we included angiograms that fulfilled the following criteria.

1.  
   A diagnosis of CNV due to AMD.
2.  
   Baseline (before intervention) angiograms that had been acquired with standardized image-capture protocols. The protocol specified acquisition of color and red free stereo pairs followed by stereo capture of multiple early (up to 30 seconds), mid (31 seconds to 2 minutes), and late (5 to 10 minutes) frames of the eye with CNV (study eye), and mid and late frames of the fellow eye.
3.  
   Corresponding color images of the fundus available in either 35-mm film or digital format.
4.  
   Measures of visual function in the eye with CNV assessed within 1 week of the angiogram. Full details of the vision-testing protocol have been published. ¹² Briefly, best corrected distance visual acuity (DVA) was measured after protocol refraction with ETDRS (Early Treatment Diabetic Retinopathy Study) charts at a testing distance of 4 meters. Near acuity was recorded with Bailey-Lovie near-reading cards at a testing distance of 40 cm, and contrast sensitivity was measured on the Pelli-Robson chart at a distance of 1 meter. All tests were performed on right and left eyes individually. Vision-testing lanes were clearly marked, and tests were performed under distance and illumination conditions as per the manufacturers’ instructions for the various charts.

We exported the image files from the first 134 subject entries that fulfilled these criteria, stripping them of all personal identifiers. Two digital acquisition systems were used (in 79 of the cases, the Ophthalmic Imaging System [OIS] Sacramento, CA; and in 55 of the cases; ImageNet, Topcon, Tokyo, Japan). The color fundus images accompanying the OIS angiograms had been captured on 35-mm film and were digitized (Coolpix 9000ED; Nikon, Tokyo, Japan) to a resolution of 1.7 megapixels. Each angiographic frame on the OIS was 0.26 megapixels. ImageNet color images were 0.44 megapixels, and angiographic frames were 1.1 megapixels. All images, except for the scanned color images, were stored in a lossless file format. All images were downsampled to a resolution of 0.26 megapixels, the lowest resolution of any of the source images. 

Of the 134 sets of color and fluorescein angiograms, we randomly selected 71 (40 OIS and 31 ImageNet) for analysis by our custom software. 

All the 134 sets of angiograms had been subjected to traditional grading, which involved delineation of the lesion boundary, area of total CNV, and area of classic CNV, by displaying the color and angiographic sequence for each case on a computer screen. The size of the CNV lesion was defined as the area of CNV and any features that obscured the boundaries of CNV. ^{1 13 14 15 16}  

As the OIS software had no built-in tools for image analysis, the total area of CNV alone (classic plus occult) and that of classic CNV alone was estimated by using standardized categorical MPS disc circles that were superimposed on the computer screen. The circle areas (in MPS units) used were 1, 2, 3, 3.5, 4, 5, 6, 9, 12, and 16. ^{13 14} Angiograms captured on the Topcon acquisition system were analyzed using the manufacturer’s own precalibrated software and therefore absolute area measurements were available from the gradings. However, to permit analysis conjointly with OIS generated data, the areas of classic CNV, all CNV and total lesion from Topcon gradings were converted into MPS disc areas. Based on angiographic grading, the lesions were classified as predominantly classic, minimally classic, or occult with no classic. 

Custom quantitative fluorescein angiography (QFA) analysis software was written in C# (Microsoft, Redmond, WA) and executed on computer (Dell Dimension 9100, with a Pentium IV 3.0-GHz processor and 1-GB RAM; Round Rock, TX). This software allows a human grader to perform advanced quantitative analyses on digital angiogram sequences by guiding them through four tasks: image selection, preprocessing, registration, and classification. 

The first task completed with the assistance of this QFA software was to scrutinize each angiogram to ensure that an adequate frame existed to represent the (1) arterial, (2) arteriovenous/laminar, (3) venous, (4) late venous, and (5) recirculation phases of each angiogram. Although stereoscopic pairs were available, only one member of a pair was selected for analysis, and wherever possible, frames including the optic nerve wholly or in part were chosen, because this structure is helpful in the subsequent calculation of fluorescence-intensity parameters. Next, a representative frame from each of these five temporal phases was selected and added to a proprietary multipage file. The data within the files were then preprocessed by automatically detecting the black image aperture mask for each fundus image frame and correcting for lighting irregularities (Fig. 1) . The individual pages within the files were then aligned, or registered, to ensure that each pixel location in every image corresponded to the same tissue location in the eye. This was accomplished by using a supervised method based on a quadratic transformation function. A minimum of four points (each at a vessel crossing) was selected on the color image, with all subsequent images registered on the basis of the same four points. Registration was good in 61 sets. In 10 sets, it was deemed inadequate (6 OIS and 4 Topcon) because of misregistration between pairs. 

The QFA software was then used to grade the angiograms by using the expertise of the human observer. A specialized segmentation routine allowed the user to delineate morphologic landmarks and abnormal features rapidly in color and angiographic frames of the fundus by employing a classification tool. Boundaries of lesion components were not delineated in the traditional sense; instead, the software allowed the grader to identify areas with similar characteristics to a coarser or finer definition. 

Although the feature classification process can be performed in almost any order, the grader would typically commence by filling in the optic nerve and large vessels using either the color image or any suitable angiographic frame. The resolution of the classification tool can be changed to allow the grader to fill in larger or smaller groups of pixels to identify and fill in fundus and lesion features as accurately as possible. Regions corresponding to blood, exudate (which are easily recognized on the color image), areas of hyperfluorescence displaying characteristics of classic or occult CNV, atrophy (both geographic and nongeographic), fibrosis, and blocked fluorescence not due to blood were identified and filled in. An option to allow involuted CNV (nonperfused scar) to be identified was also available. The user then completed the process of classification by identifying background for the purposes of normalization. Background was selected by filling in peripheral retinal areas of normal fluorescence that lay between major vessels. During the classification process, the user is able to scroll through the six registered frames using the mouse wheel. Accumulated classified (or graded) data are hidden or revealed, to facilitate refinements in the identification of subtle changes in fluorescence patterns. An example of a case before and after grading is shown in Figure 2 . 

For the purposes of analysis, we derived several metrics from the classified composites. First, pixel areas (PA) were assigned to each component (e.g., classic CNV) identified in the angiographic sequence. Using an empiric camera magnification conversion factor, we calculated the area occupied by each of the normal anatomic features (optic nerve, artery, and vein) and each of the pathologic components of AMD (classic and occult CNV, serous pigment epithelium detachments (PED), blood, blocked fluorescence not due to blood, fibrosis, and nongeographic and geographic atrophy). Next an integrated intensity (II) for each lesion component was obtained (Fig. 3A) . The II is a measure of the total (summed) intensity of fluorescence over the course of the angiogram. Third, the positive fluorescence (PF), which is defined as the summed positive change in fluorescence in only those angiographic frames taken after the arteriovenous laminar phase was calculated (Fig. 3B) . As neither the II nor the PF takes into account changes in background fluorescence and flash intensity, all the generated parameters were normalized (by division) to background (b) and the optic nerve (o). We termed these normalized readings the IIQ^b, PFQ^b and IIQ°, PFQ°, where Q represents the quotient and the superscript indicates the normalizing structure. 

Statistical Analysis

Data were analyzed on computer (SPSS ver. 11; SPSS, Chicago, IL). The results of traditional grading into the three angiographic subtypes were cross-tabulated against that obtained by QFA analysis. The κ statistics were calculated, and the Wilcoxon signed rank test was used to examine for bias between methods. 

Pearson’s correlations between the seven quantitative parameters generated by QFA grading (pixel area, II, IIQ°, IIQ^b, PF, PF° and PFQ^b) for each lesion component were examined to assess colinearity. Pearson’s correlations coefficients were also generated, to examine the relationships for QFA parameters from CNV leakage, classic CNV, occult CNV, blocked fluorescence, blood, and atrophy with each clinical measure of vision. The QFA parameters for single lesion components showed high colinearity with each other. Visual function parameters also showed high colinearity with each other. Therefore, only one QFA parameter per lesion component and one clinical measure of vision with which the strongest and most consistent relationships were observed were entered into the regression model. A backward stepwise elimination algorithm was applied to identify the lesion components that reached significance in the regression model. 

Standard linear regression techniques were then used to select the optimum combination of independent variables (age, gender, QFA lesion components that reached significance in the previous step and the metrics generated by traditional grading) that explained most of the variance in the dependent variable which was DVA. 

Results

Lesion Subtype by Traditional Grading and QFA

Data were complete on 61 angiograms. On the basis of traditional grading a spread of the different angiographic subtypes was observed, with predominantly classic accounting for 48%, minimally classic for 33%, and occult with no classic for 20% (late leakage of undetermined source or fibrovascular pigment epithelial detachment). For cross-tabulating lesion classification by traditional methods versus that by QFA, agreement was good, with a weighted κ statistic of 0.70 (Table 1) . For lesions classified as predominantly classic by traditional grading versus QFA, there was agreement in 25 (74%) of 34 cases. Agreement was lower in lesions classified as minimally classic (13/24, 54%). The Wilcoxon signed rank test showed that there was a statistically significant difference in the assignment of CNV subtype by the two methods. Traditional grading assigned more cases to the predominantly classic subgroup, whereas use of QFA software resulted in more cases being assigned to the minimally classic subtype (P = 0.03). 

Correlations between QFA Metrics

Examination of the correlation matrices for PA, II, PF, IIQ^b, IIQ°, PFQ^b, and PFQ° for each of the lesion components—classic CNV, occult CNV, blocked fluorescence, blood, atrophy and leakage—revealed that in each case II was highly correlated with the IIQ^b and that PF was highly correlated with PFQ^b. An example of the matrix for classic CNV parameters is shown in Table 2 . 

Relationships between QFA Lesion Parameters and Visual Function

On examining correlations between the QFA metrics and clinical measures of vision, only the PFQ^b exhibited statistically significant relationships (data not shown). Table 3shows a sample correlation matrix of PFQ^b from the various lesion components (leakage, classic CNV, occult CNV, blocked fluorescence, atrophy, and blood) with the three clinical measures of vision. The most consistent relationships were seen between DVA and QFA metrics and of these only the relationship between classic CNV and DVA was significant at the 0.01 level. 

As the most consistent relationships occurred between DVA and PFQ^b metrics of the various lesion components, we used regression analysis to explore and quantify the relationships further. With DVA as the dependent variable and PFQ^b from leakage, classic CNV, occult CNV, blocked fluorescence, blood, and atrophy as independent variables, the backward stepwise elimination algorithm excluded all except PFQ^b classic and PFQ^b blood (Table 4) . The parameter estimates from model 5 showed that for a change of 30 on the PFQ^b for classic CNV a loss of 20 letters or four lines of DVA may be expected and this is also illustrated in the scatterplots of DVA and PFQ^b classic (Fig. 4) . 

Relationship between PFQ^b Classic, Traditional Grading, and DVA

The final regression analysis with DVA as the dependent variable and the independent variables age, gender, PFQ^b classic, PFQ^b blood, traditional grading parameters of classic, all CNV and total lesions showed that only PFQ^b classic was retained in the model (β = −0.402, t = −3.281, P < 0.01). 

Discussion

The automated analysis of images from in vivo examinations of human organs is now a reality in many fields of medicine, with notable advances having been made, particularly in radiologic disciplines. This has not been the experience in retinal disorders, and the interpretation and analysis of fundus fluorescein angiograms by automated means is not yet used as an alternative to traditional human grading methods. The findings of the present study challenge these views and show that the complex and subtle changes in fluorescence patterns in CNV are better described by the metrics generated through application of computerized algorithms than by currently used descriptors and area measurements. Notably, more robust relationships with measures of visual function were demonstrated for the dynamic metrics generated by the QFA software (i.e., the background-corrected positive fluorescence quotients from CNV lesion components PFQ^b). 

In the context of a fluorescein angiogram in an eye with exudative AMD, it is intuitive that the dynamic characteristics of the abnormal spatial and temporal fluorescence representing the CNV will be a reflection of the state of health of the tissues and cells in its vicinity. The finding that the PFQ^b measured from classic CNV showed stronger relationships with visual function than do clinical descriptions or area measurements of the lesion is therefore in full accord with our understanding of the underlying disease process. The PFQ^b has two important features. First, it reflects changes in the intensity of fluorescence in the structure; and, second, it takes into account the relative intensity of the structure with respect to the background fluorescence. Accounting for background fluorescence is critical, particularly in the digital era, where photographers frequently adjust the flash and gain settings in an attempt to optimize image quality. Traditional grading, by contrast, is based on planimetric area measurements without quantitative assessment of the temporal changes in the intensity of fluorescence. Thus, it is not surprising that studies show disappointingly low levels of correlation between the descriptive morphology and metrics from the currently used angiographic grading method with visual function. This is particularly the case with newer antiangiogenic pharmacotherapeutic agents, with which an expansion of CNV lesion area has been observed by angiography, despite stabilization or improvement in visual acuity. ⁸ Our results suggest, that quantitative extraction of the dynamic attributes (e.g., PFQs) would result in improved correlation with functional outcomes. Despite this, the correlation between the PFQ^b and visual acuity was modest (adjusted r ² = 0.26), indicating that a significant proportion of the variability in visual function remains unexplained. This finding is not surprising as many clinically relevant attributes of CNV lesions are not taken into account by the PFQ. For example, the PFQ, as it is currently calculated, does not account for the location of the fluorescence with respect to the foveal center. It is likely that lesions with greater activity and leakage near the foveal center will cause worse acuity. It is likely with additional study, QFA parameters can be further refined to take into account factors such as foveal proximity, and thus yield even better correlations. However, QFA parameters alone are unlikely to explain all the variability, as other factors that may not be reflected in the angiographic data such as chronicity of the lesion and extent of photoreceptor loss are likely to be important. Combining QFA data with these other clinical variables, including possibly optical coherence tomography (OCT) data may ultimately yield the most predictive models. 

In the present study, we used the QFA software with a human grader who used a classification tool to delineate the different morphologic features of the retinal fundus. Although there is still a heavy reliance on human grader input in this process, the QFA software has additional functionality that can refine incomplete segmentations generated by a grader to create a more complete map of retinal topography and exudative AMD lesion components. In summary, we have demonstrated the improved functional relevance of the metrics generated by the QFA. We are currently testing its reproducibility and reliability and extending our studies to explore further automation of the technique and in particular the use of algorithms that will minimize grader input and maximize automated analysis. 

 Supported by National Eye Institute Grants EY015914 and EY03040, and a Sabbatical Award to from The Research and Development Office Northern Ireland (UC).

 Submitted for publication April 28, 2006; revised June 22 and August 22, 2006; accepted November 10, 2006.

 Disclosure: U. Chakravarthy, None; A. C. Walsh, None; A. Muldrew, None; P. G. Updike, None; T. Barbour, None; S.R. Sadda, None

 The publication costs of this article were defrayed in part by page charge payment. This article must therefore be marked “advertisement” in accordance with 18 U.S.C. §1734 solely to indicate this fact.

 

Figure 1.

 

Correction for irregularities in lighting. Top row: the image sequence before processing; bottom row: the sequence after detection of the black image aperture mask for each fundus image frame and correction for lighting irregularities. Note that, after processing, the lighting is even across the images.

Figure 2.

 

A case of choroidal neovascularization before (A) and after (B) grading. Although only one angiographic frame is shown, the software allows the user to rapidly move between the aligned sequence of images to facilitate grading decisions. Various lesion components are indicated by different colors (B).

Figure 3.

 

Color maps of (A) the II obtained by summing the pixel intensity at each pixel location across all five phases of the angiogram, and (B) the PF obtained by summing the positive change in fluorescence in only those angiographic frames taken after the arteriovenous laminar phase. As shown in the color bar, red-purple indicates higher values and blue-black are lower values.

Table 1.

 

Lesion Subtype Classification

CNV Type by Traditional Grading	CNV Type by Computerized Analysis (QFA)
		Predominantly Classic	Minimally Classic	Occult, No Classic
Predominantly classic	25	7	2
Minimally classic	1	13	1
Occult no classic	0	2	9

 Weighted κ = 0.70.

Table 2.

 

Correlation Matrix of Relationships for the Metrics Describing Classic CNV

	PA	II	IIQ°	IIQb	PF	PFQ°
PA	1	—	—	—	—	—
II	0.284*	1	—	—	—	—
IIQ°	0.269*	0.726, **	1	—	—	—
IIQb	0.241	0.829, **	0.839, **	1	—	—
PF	−0.073	0.453, **	0.163	0.107	1
PFQ°	0.093	0.304*	0.119	0.025	0.572, **	1
PFQb	−0.031	0.342*	0.144	0.043	0.637, **	0.556, **

 Correlations were significant at

*   P ≤ 0.05 and

**   P ≤ 0.01. The correlation between II and its quotients and PF and its quotients are significant, indicating colinearity.

Table 3.

 

Correlation Matrix of Clinical Measures of Vision and the QFA Metric Positive Fluorescence Quotient Corrected to Background

Lesion Component	DVA	NVA	CS
PFQb leakage	−.315*	272	−.213
PFQb classic	−.402, **	277	−.233
PFQb occult	−.256	285*	−.258
PFQb blocked fluorescence	−.280*	070	−.107
PFQb blood	−.320*	245	−.288*
PFQb atrophy	−.220	185	−.170

 The strongest correlations are seen between DVA (letters read) and the metric from classic CNV. The correlation is negative for DVA and CS, which are measured as the number of letters read, and positive with NVA, which is measured as the log of the minimum angle of resolution. Correlations were significant at

*   P ≤ 0.05 and

**   P ≤ 0.01.

Table 4.

 

Linear Regression with Backward Stepwise Elimination of Independent Variables

Model	Excluded Variables	F	Adjusted R 2	Significance
1	—	2.633	0.15	.027
2	Atrophy	3.199	0.16	.017
3	Atrophy and leakage	4.028	0.18	.006
4	Atrophy, leakage and occult CNV	5.058	0.18	.004
5	Atrophy, leakage, occult CNV and blocked fluorescence	6.969	0.17	.002

 The dependent variable is DVA. The final model, number 5, included the two variables PFQ^b classic and PFQ^b blood and was statistically significant.

Figure 4.

 

Scatterplot of relationships between PFQ corrected to background (PFQ^b) for leakage and DVA (letters read). As leakage increased the number of letters read was reduced. BG, background.