Measurement of Fluorescein Angiograms of the Optic Disc and Retina Using Computerized Image Analysis PAUL NAGIN, PhD, BERNARD SCHWARTZ, MD, PhD, GEORGE REYNOLDS, PhD
Abstract: Computerized image analysis was used to quantify objectively fluorescein angiograms of the optic disc, peripapillary choroid, and retina. Techniques were developed to measure fluorescein filling rates of the optic disc and the retinal vessels and the area of fluorescein filling defects within the optic disc. Two subjects, one with glaucoma and the other with ocular hypertension, showed increases of areas of fluorescein filling defects of the optic disc on follow-up and are presented here as examples of the application of these techniques. This methodology can be applied to the longitudinal follow-up of individual patients with glaucoma and retinal diseases, as well as to cross-sectional studies of patient populations. [Key words: computers, filling defects, filling rates, fluorescein angiography, glaucoma, image analysis, ocular hypertension, optic disc, reproducibility, retina.] Ophthalmology 92:
547-552, 1985
Several studies have demonstrated that fluorescein angiography of the optic disc reveals localized areas of filling defects (Fig 1) or hypofluorescence in glaucomatous, ocular hypertensive, and normal eyes. 1,2 The number and size of the filling defects of the optic disc are greater in ocular hypertensive and glaucomatous eyes than in normal eyes. 2,3 Decreased fluorescein appearance From the Department of Ophthalmology, New England Medical Center and Tufts University School of Medicine, Boston. Supported by research grant 2 R01 EY00936 from the National Eye Institute, National Institutes of Health, Bethesda, Maryland. Presented in part at Meeting of the Association for Research in Vision and Ophthalmology May 3, 1984, Sarasota, Florida, and the Annual Meeting of the American Academy of Ophthalmology, November 15, 1984, Atlanta, Georgia. Reprint requests to Paul Nagin, PhD, Department of Ophthalmology, New England Medical Center, Inc., 171 Harrison Avenue, Boston, MA 02111.
times4 and filling rates 5 have been observed in glaucomatous and ocular hypertensive individuals. The present study is an extension of an attempt to use microdensitometry to observe the relationship between time and density of fluorescein at each time or within each photographic frame of the fluorescein angiogram. 5 The technique was partially objective and quantitative, but was laborious to use. The methodologies described here are both objective and quantitative and rely on computer image processing for the analysis of fluorescein angiograms.
MATERIALS AND METHODS IMAGE ANALYSIS
Computer image analysis is a tool for reading and interpreting visual information present in photographs. A typical system consists of a mainframe computer
547
OPHTHALMOLOGY
•
APRIL 1985
•
VOLUME 92
•
NUMBER 4
Fig 1. Auorescein angiogram of optic di~, late arteriovenous phase, with filling defects visible within most disc areas.
Fig 2. Two digitized images superimposed of two frames of the fluorescein angiogram in Figure I, indicating their lack of alignment with respect to each other.
connected to a television camera used as a scanning device. A photograph is scanned and then converted to digital form by a video digitizer. For each (x, y) point in the image, the digitizer converts the analog video brightness signal of transmittance values to produce a digital value. Once inside the computer, the digital values are arranged as a 2-dimensional array of brightnesses. The image is then analyzed by software programs designed to extract significant details or features in the Image.
the sequence of photographic slides comprising a given angiogram. This was accomplished using the television camera and attached video digitizer. The adjustments for focus, lighting, and slide placement were made for a single photograph taken from the sequence and then fixed during the digitization of the remaining photographs. The digitized images were stored on disk memory for later recall.
HARDWARE
The sequence of digitized images was aligned or registered using a software algorithm described by Glazer, Reynolds, and Anandan. 6 The necessity to register the images can be demonstrated by taking a pair of digitized images from a single angiogram sequence and superimposing them on the display monitor (Fig 2). The images are clearly out of phase with respect to each other, which is caused by several factors such as eye movement and variations in slide mounting under the television camera prior to the scanning/digitizing process. Image registration consisted of three distinct phases: preprocessing, matching, and repositioning. The preprocessing phase was designed to enhance the images so that features important in the matching phase, such as edges, corners, and vessel crossings, would stand out. After preprocessing, the computer selected the 200 points from the preprocessed image that had responded maximally to the enhancement process, ie. the 200 strol1gest edge or corner points. These points were used in the matching phase. The methodology was to place a window around a point to be matched in the first image, and then find its best match in the succeeding image. The matching was based on a similarity measure computed for the two windows. The similarity measure
Our computer system consisted of a Digital Equipment Corporation VAX (Maynard, MA) 11/780 computer with one megabyte of on-line memory and 700 megabytes of on-line magnetic disk used for storage of programs and digital images. A Spatial Data Systems (Goleta, CA) EYECOM television camera with a chalnicon scanning tube was used for image acquisition. The video digitizer was capable of obtaining a digital resolution of 512 X 512 pixels with 8 bits (256) of gray level per pixel. To save computation time the spatial resolution of the images was reduced in software to 128 X 128 pixels. For an optic disc diameter of 1.5 mm, the reduced resolution was approximately 12.0 J.Lm. The image display system was a DEANZA Systems (San Jose, CA) Model IP-5000, with three channels of 512 X 512 points each. The system could distinguish and display 256 gray levels at each pixel, in each of the three channels. SOFTWARE
Fluorescein angiograms were analyzed on the computer in the following manner: first, the operator digitized 548
IMAGE REGISTRATION
NAGIN, et al
•
MEASUREMENT OF FLUORESCEIN ANGIOGRAMS
Fig 3. Vector field for the images in Figure 2, indicating translation and rotation required to shift the first image into registration with the second image.
Fig 4. Two images from Figure 3, superimposed after shifting and rotating according to the mean vector in the field .
was the value of the correlation coefficient, defined as the sum of the pairwise product of the pixels in the two windows. The output of this function reached its maximum as the gray levels in the two windows grew closer. The window placement in the second image was varied across an ever-increasing area until the correlation coefficient reached its maximum. In general, this point was reached after only a few steps. The matching process was repeated independently for each of the 200 points. To view the results of the matching process, a vector field was drawn on the display monitor (Fig 3). In this figure, the tail of each vector lies on a pixel in the first image that is to be matched. The head points to its most similar mate in the second image. A translation and rotation factor for the whole image was determined by computing the average vector in the field. In the final phase of the alignment algorithm, a program was run to shift and rotate the first image into alignment with the second image based on the computed average vector (Fig 4). The registration process was repeated for all successive pairs of images in the angiogram. FILLING RATES
Once the image sequence had been registered, the computer was used to measure several parameters of the angiogram. The first set of measurements related to fluorescein filling and circulation times. As shown in Figure 5, the operator selected, via a joystick, the locations of clinically interesting points in and around the optic disc. A total of 80 points was selected from four distinct areas as follows: 20 points from inside the
Fig 5. Operator-selected points of clinical interest in a fluorescein angiogram, as indicated by white or black SQuares on disc, retinal vessels, and peripapillary retina or choroid.
optic disc (avoiding major vessels), 20 points from the retina (avoiding major vessels), 20 points along the retinal arteries, and 20 points along the retinal veins. All points that were selected outside of the optic disc were located within about a half-disc diameter from the disc. 549
OPHTHALMOLOGY
•
APRIL 1985
•
VOLUME 92
•
NUMBER 4
Table 1. Fluorescein Filling Defect Areas Expressed as a Percentage of Total Disc Area and Reproducibility of Duplicate Measurements
250
>-
I-
en
Percentage Filling Defect
Z
W
0
VEINS
200
>-
I-
en Z
W
IZ
150 ~ ~rf
z
w u
en a:: o W
"
100
"
ARTERIES
, D,
D
' .... ---:.::~-=:..._.:._-=~ DISC ... ,
'
CHOROID
.....J
50
~ .....J W
a:: I!
25
4/78 9/82 3/79 2/82
1 2 2
Measurement 2
8
10
15
13 19
37
20
40
30
35
40
the disc boundary and multiplied by 100, to generate the total percent area of filling defect for that disc. Overall, the algorithm required about 15 minutes of central processor time on the VAX computer to analyze a fluorescein angiogram sequence consisting of 20 slides. This figure includes time for digitization, registration, and all measurements of filling times and filling defect areas . SUBJECf SELECfION AND PHOTOGRAPHY
,!
I
17 20
1
Measurement 1
"......
\
IL
W
Date
..
'\
:::>
>
Patient No.
45
50
TIME (SECONDS) Fig 6. Graph of fluorescein intensity over time of the angiogram showing curves for the various structures.
After the operator selected the points, the computer recorded the intensity level at each point over time. Note that each photographic image in a given sequence was labeled automatically by the fundus camera according to the time in seconds from the angiogram injection. At each time frame the computer calculated the average fluorescein intensity across each of the four sets of 20 points. An intensity graph was then plotted.
Fluorescein angiograms of two patients were selected from the photographic records of the Outpatient Service of the Department of Ophthalmology at the New England Medical Center. The first was a 16~year-old patient with juvenile glaucoma with no visual field loss at the time of the first fluorescein angiogram, but who developed field loss by the time of the second fluorescein angiogram. The second subject was a 55~year-old ocular hypertensive whose ocular pressures ranged from 17 to 30 mmHg in the left eye without any visual field loss, as measured by kinetic and static means with the Goldmann perimeter. The techniques used for fluorescein angiography of the optic disc have been described elsewhere. 2 Informed consent was obtained from all patients and subjects undergoing fluorescein angiography.
FILLING DEFECfS
ANALYSIS OF DATA
The registered sequence of images was also used in the detection of filling defects within the optic disc. The computer first delineated the optic disc boundary using a boundary tracking program.? Next, the computer recorded an intensity-time vector for each point contained within the optic disc, that is, for any given point, the computer maintained a list of the associated intensity values at each time frame of the angiogram. Linear regression analysis was then applied independently to each of the vectors. The P-values indicating the probability that the slopes of the regression lines were significantly different from zero were determined. A threshold P-value was selected empirically and all points having P-values greater than the threshold were counted as filling defects, ie. filling defects were points that did not show a tendency to change in gray level over the time of the angiogram. The number of points showing no change of intensity of fluorescein with time was then divided by the total number of points within
To test the reproducibility of the computer methods, each angiogram was digitized and analyzed twice. Linear regression analysis was done using the BMDP software package. 8
550
RESULTS Figure 6 is an example of the computer-generated plot of fluorescein intensity versus time for the optic disc, retinal veins, retinal arteries, and peripapillary choroid of case 1. The results of the measurements of filling defects, expressed as a percentage of the total disc area, are given in Table 1. Figure 7 shows the analysis for filling defects for patient 1 for the two time periods that were selected. The upper left image is a frame taken at the late arteriovenous phase from a 1978 fluorescein angiogram. The upper right image shows, in black, those points that
NAGIN, et al
•
MEASUREMENT OF FLUORESCEIN ANGIOGRAMS
Fig 7. Result of analysis for filling defects for patient I. Top left, the late arteriovenous phase from a 1978 angiogram. Top right, filling defect areas in black (8%). Bottom, a 1982 angiogram (left) reveals 37% area of filling defects (right).
Fig 8. Results of analysis for filling defects for patient 2. Top, a 1978 angiogram (left) reveals 15% area of defects (right). Bottom, a 1982 angiogram (left) reveals 20% area of defects (right).
were determined to be filling defects of the optic disc (8% area of defect). The lower left image is from the late arteriovenous phase of a 1982 angiogram. The lower right image shows the analysis for filling defects (37% area of defect). Similarly, in Figure 8 for patient 2, the upper images are from a 1978 angiogram (15% area of defect) and the lower images were from a 1982 angiogram in a similar phase (20% area of defect). An increase in filling defect area can be seen in this ocular hypertensive eye, particularly in the temporal quadrant.
The computer analysis was objective and required very little human intervention. Moreover, the measurements produced by the analysis were all relative and invariant to linear changes in the overall brightness level from one fluorescein sequence to another. Any calculations of circulation and transit times or slopes would be based on intensity differences, not absolute density levels. s Most clinicians evaluate angiograms of the optic disc subjectively. They typically look only for filling defect areas and cannot readily measure filling rates or circulation times. Filling defects often are too subtle to be seen, and observer variation can be an important factor. Several ' studies have used quantitative or objective systems to evaluate fluorescein angiograms. Schwartz and KernS used a rotating scanning Q-stage microdensitometer to obtain density readings from fluorescein angiogram sequences. They found that increased age correlated with decreased rates of filling in all areas of the retina and optic disc. Other factors that correlated with decreased filling rates included increased diastolic blood pressure, increased intraocular pressure, and decreased tonographic outflow. Significant differences of circulatory changes in the retinal arteries and veins were obtained between the normal, ocular hypertensive, and glaucomatous eyes that they measured. Others have used television ophthalmoscopy and video recording to store images for later analysis. 9- 11 Moses' 2 measured blood flow in fluorescein angiograms using a photometer to measure relative light transmission. Piccolino et al '3 ,14 used image analysis in retinal and
DISCUSSION Our results showed that computerized image analysis is a reproducible technique for quantifying fluorescein angiograms. The graph of filling times (Fig 6) is similar to the graphs obtained by manual measurements in our previous study.s These graphs can be produced easily by the computer, and the associated measurements of transit times, filling times, and slopes of the curves can also be calculated. The measurements of areas of fluorescein filling defects showed an increase over time in each of the patients tested. In the first patient this was associated with visual field loss. In the second patient there was no associated visual field loss, but the patient stilJ showed elevated ocular pressure and no apparent changes in the optic disc. These changes corresponded well to the subjective interpretation of the angiograms.
551
OPHTHALMOLOGY
•
APRIL 1985
macular fluorescein angiography. They obtained graphs of overall fluorescein density changes in the retina through the course of an angiogram and used these graphs to detect several types of retinal leakage. These authors also produced macular densitomograms for the evaluation of macular pigmentation. None of these studies offer an objective or quantitative method for overcoming misregistration of successive frames of an angiogram. Nor do any other studies attempt to quantify objectively any filling defect areas of the optic disc. For analysis of filling defect areas, a linear model was used to represent fluorescein intensity changes over time. As can be seen from the graph of disc filling (Fig 6), this assumption is not quite valid and a secondorder fit would be more appropriate. The usefulness of this approach is presently under investigation. The amount of computer time required to analyze a given angiogram was a function of the number of photographs selected from the sequence (typically, 20 were selected) as well as the digital resolution used (ie. the number of pixels). A resolution of 128 X 128 was used in the present study. Both factors affect the computer time linearly. In future studies we will determine the minimum number of photographs per sequence as well as the minimum number of pixels per photograph required to generate comparable results. Of particular interest is the effect of either factor on the reproducibility of the overall analysis. A great advantage of computerized image analysis over human evaluation of photographs is the uniformity of sensitivity of the evaluation process. The software programs used to extract the measurements from the digitized photographs were not altered in any manner during the course of the measurements. The reproducibility of analyzing an identical digitized image was zero. The only sources of errors with the use of fixed software programs were external to the computer, in the photographic and scanning processes. The clinical application of objective, quantitative, and reproducible methods using computerized image analysis can establish data bases that accurately relate measurements of fluorescein angiograms to other clinical and demographic parameters. Several such studies are currently in progress. Furthermore, these computer techniques can be applied to a wide range of related problems in diagnostic ophthalmology, including the measurement of optic disc pallor 15 and the measurement of optic cup volume by means of digital stereophotogrammetry.16
552
•
VOLUME 92
•
NUMBER 4
ACKNOWLEDGMENT The authors thank Susan Glick, MS, for editing the manuscript.
REFERENCES 1. Spaeth GL. Fluorescein angiography: its contributions towards understanding the mechanisms of visual loss in glaucoma. Trans Am Ophthalmol Soc 1975; 73:491-553. 2. Schwartz B, Rieser JC, Fishbein SL. Fluorescein angiographic defects of the optic disc in glaucoma. Arch Ophthalmol 1977; 95: 1961-74. 3. Loebl M, Schwartz B. Fluorescein angiographic defects of the optic disc in ocular hypertension. Arch Ophthalmol 1977; 95:1980-4. 4. Spaeth GL. The Pathogenesis of Nerve Damage in Glaucoma: Contributions of Fluorescein Angiography. New York: Grune and Stratton, 1977. 5. Schwartz B, Kem J. Age, increased ocular and blood pressures, and retinal and disc fluorescein angiogram. Arch Ophthalmol 1980; 98:1980-6. 6. Glazer F, Reynolds G, Anandan P. Scene matching by hierarchical correlation. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattem Recognition, 1983; 432-41. 7. Nagin P, Schwartz B. Approaches to image analysis of the optic disc. In: IEEE Computer SOciety Fifth International Conference on Pattern Recognition, 1980. New York: Institute of Electrical and Electronic Engineers; 948-56. 8. Dixon WJ. BMDP Statistical Software. Berkeley: University of California Press, 1981; 264-77. 9. Haining WM. Video funduscopy and fluoroscopy. Br J Ophthalmol 1981; 65:702-6. 10. Preussner PR, Richard G, Darrelmann 0, et al. Quantitative measurement of retinal blood flow in human beings by application of digital image-processing methods to television fluorescein angiograms. Graefes Arch Clin Exp Ophthalmol 1983; 221: 11 0-2. 11. Jung F, Kiesewetter H, Korber N, et al. Quantification of characteristic blood-flow parameters in the vessels of the retina with a picture analysis system for video-fluorescence angiograrns: initial findings. Graefes Arch Clin Exp Ophthalmol 1983; 221: 133-6. 12. Moses RA. Intraocular blood flow from analysis of angiograms. Invest Ophthalmol Vis Sci 1983; 24:354-60. 13. Piccolino FC, Zingirian M, Parodi GC. Electronic image analysis in retinal fluoroangiography. Ophthalmologica 1979; 179:142-7. 14. Piccolino FC, Zingirian M, Parodi GC. Densitometric analysis of macular fluorescein angiography. Ophthalmologica 1981; 182: 171-4. 15. Nagin P, Schwartz B. Detection of increased pallor over tirne; computerized image analysis in untreated ocular hypertension. Ophthalmology 1984; 92:252-61 16. Falconer DG. Digital stereophotogrammetry of the ocular fundus. Appl Opt 1973; 12:1388-9.