- Email: [email protected]
Computer Analysis of Visual Field Loss and Optimization of Automated Perimetric Test Strategies
Computer Analysis of Visual Field Loss and Optimization of Automated Perimetric Test Strategies CHRIS A. JOHNSON, PhD,* JOHN L. KELTNER, MDt
Abstract: Preliminary investigations of optimal target distribution patterns for automated perimetry were conducted. In the first study, frequency distributions of visual field defects were determined in glaucoma (260 scotomata) and optic nerve disease (110 scotomata) for approximately 30,000 visual field locations. The frequency distributions in glaucoma and optic nerve disease were different, and suggest guidelines for development of target distribution patterns to achieve optimal detection rates. The second study included computer simulation of 20 target distribution patterns currently used in manual and automated perimetry. Each distribution pattern was processed through the 370 glaucoma and optic nerve disease scotomata to assess detection performance. Both the density and distribution of target locations affected overall detection rates. In addition, some target configurations were more susceptible to false alarms than others. These data provide qualitative information for optimization of target distribution patterns in automated perimetry, and serve as a foundation for future quantitative studies. [Key words: computer simulation, glaucoma, optic nerve disease, perimetry, visual field loss.] Ophthalmology 88: 1058-1 065, 1981
The field of automated perimetry began about 10 to 15 years ago with theoretical evaluations of its feasibility for performing clinical visual field testDepartment of Ophthalmology, * and Departments of Ophthalmology, Neurology and Neurological Surgery,t School of Medicine, University of California, Davis California. Presented at the Eighty-fifth Annual Meeting of the American Academy of Ophthalmology, Chicago, Illinois, November 2-7, 1980. Supported in part by National Eye Institute Academic Investigator Award #K07-EY00095.* Reprint requests to Chris A. Johnson, PhD, Department of Ophthalmology, University of California, Davis, CA 95616.
1058
ing. Since that time, it has progressed from a laboratory research phenomenon to the development of almost a dozen different commercially available devices. Most of these commercial devices have now been subjected to one or more clinical validation and comparison studies .1-15 The consensus from such investigations is that automated perimetry can be an effective procedure for detection (and in some instances detailed evaluation) of visual field loss, provided that specific test procedures, stimulus conditions, and methods of interpreting the test results are employed. We are currently entering a new phase of auto-
0161-6420/8111000/1058/$00.90
© American Academy of Ophthalmology
JOHNSON AND KELTNER. COMPUTER ANALYSIS OF VISUAL FIELD LOSS
mated perimetry, in which refinements are being developed to optimize existing test procedures. One approach to this problem consists of direct clinical comparison studies of different automated devices using rigorous manual procedures as a reference standard. 2,8,12,15 This approach has many desirable features, although there are also several disadvantages. First, it is not possible for patients to perform in a consistent, error-free manner for successive test procedures. Thus, there will always be some differences in performance among various devices that are due to patient variability, fatigue, and other factors not directly related to the test procedure. A second problem with clinical comparison studies is the interpretation of results. Since various devices differ according to their test strategies, stimulus conditions, and other relevant features, it is difficult to evaluate the contribution of individual aspects of the test to overall performance characteristics. In addition, such comparison studies are limited to evaluation of the features available on each device. Not only does this initially constrain the possible solutions for optimizing test procedures, but it also requires that revalidation studies be performed following any modifications of the devices or test procedures. As we have discovered through previous experience,1O-15 this can be a very tedious and inefficient use of research efforts. An alternative approach to clinical comparison studies in automated perimetry is the use of digital computer processing and simulation techniques. 16-20 With this approach, it is possible to have "patient' , data and performance characteristics that are the same for all devices or procedures evaluated. In addition, one can isolate individual components of the test procedure, and rapidly evaluate many potential solutions for optimal procedures. This makes it possible to quickly eliminate unsuccessful alternatives, define relevant properties of individual components of the test, and suggest the most effective test procedure. Clinical comparison studies can then be performed for this procedure to validate its effectiveness as a diagnostic tool. Fankhauser and his associates 16 ,17,19,20 have demonstrated the efficacy of this approach to define optimal threshold determination strategies for automated static perimetry with subsequent implementation in a clinical test device. The present study describes our preliminary efforts to utilize digital computer processing and simulation techniques to define optimal target distribution patterns for detection of visual field defects. Two studies were conducted: (1) using a high resolution grid (0.75 by 0.75 degree elements), frequency distributions of scotomata in glaucoma and optic nerve disease were determined for approximately 30,000 locations in the visual field; and (2) detection rates for target distribution patterns associated with 20 existing manual and
automated test procedures were determined through computer simulation of tests performed on 370 scotomata from glaucoma and optic nerve disease.
MATERIALS AND METHODS The visual field data for this study were obtained from previous kinetic visual field examinations performed on the Goldmann perimeter. Each visual field chart was selected from our patient records, and had been performed within the past four years. Most of these patients had participated in previous clinical comparison studies of manual and automated perimetry. A total of 370 scotomata were entered into the computer, 260 from glaucoma patients and 110 from patients with optic nerve disease. fur glaucoma, a distribution of early, moderate, and advanced visual field defects was chosen. fur optic nerve disease patients, a variety of anomalies at various stages of progression were selected. The population included ischemic optic neuropathy, optic atrophy, optic neuritis, papilledema, optic nerve drusen, and other optic nerve abnormalities. Only scotomata were considered for this analysis; isopter-related deficits (eg, nasal steps) were not included. An LSI-II minicomputer system, a Tektronix® Model 4953 digitizing graphics tablet and a Tektronix Model 4010 graphics terminal were used to process the visual field information. To achieve proper alignment, each visual field chart was positioned on the digitizing tablet and the coordinates of the center and right horizontal axis (0 degree meridian, 90 degree radius) were entered by a graphics pen to check the centering, orientation (rotation), and scaling of the visual field chart with respect to the tablet. fullowing this procedure, the scotoma data were entered. If more than one target had been used to plot the scotoma originally, the largest representation of the visual field defect was used. Information pertaining to the intensity of the scotoma (depth and slope characteristics) was not considered in this investigation. Each scotoma was entered by tracing its outer boundary on the digitizing tablet using the graphics pen. This boundary trace was simultaneously displayed on the graphics terminal to check for proper data input. A "window" of 0.75 by 0.75 degrees was used to control the amount of data generated by the tracing procedure. That is, the graphics pen was required to move a distance of 0.75 degrees in either the X (horizontal) or Y (vertical) direction to record each successive data point. The X- Y coordinates of each data point were then ordered, and a software algorithm was used to regenerate the total area within the scotoma, using only the boundary information. The algorithm defines the area within the scotoma by successive "slicing" along the X axis. 1059
OPHTHALMOLOGY. OCTOBER 1981 • VOLUME 88 • NUMBER 10
For a given X coordinate value, the lowest Y coordinate value is determined, and locations between it and the next Y coordinate value are filled in. If additional Y coordinates are present for this X coordinate, the procedure is repeated; otherwise, the algorithm selects the next X coordinate value and then repeats the process. The algorithm continues until the entire scotoma has been .. sliced." Under some circumstances (vertical or near-vertical boundaries, complex shapes with multiple convolutions, and regions of partial over-lap during tracing), this algorithm becomes confused as to what is inside and what is outside the scotoma. These ambiguities are resolved by keeping track of changes in direction of travel of the tracing pen. This program provides a reliable method of reconstructing the complete scotoma from boundary information, which allows us to retain the essential features of the scotoma while storing a minimum amount of data. To check for missing or erroneous data, the" filled-in" scotoma generated by this algorithm was displayed on the graphics terminal. Once the information had been checked and stored, it was used for two purposes in the present study. First, each scotoma was entered into a composite visual field representation of 30,000 locations (0.75 by 0.75 degree grid size) to obtain frequency distributions for visual field defects in glaucoma and optic nerve disease. The second purpose of the study was to simulate the target distribution patterns of most of the existing manual and automated test procedures used to detect visual field characteristics. All target distribution patterns were entered through the digitizing tablet and checked by procedures similar to the process used for scotomata. Each distribution pattern was then processed through the 370 scotomata, the number and location of targets that fell within each scotoma were recorded, and the information was then used to determine the detection performance of each simulated test procedure.
RESULTS AND DISCUSSION Before describing our findings, it is important to point out the limitations of the present study. First, our investigation only evaluates scotomatous visual field loss. Visual deficits such as nasal steps and other isopter-related abnormalities are not considered in this phase of our work. Secondly, this study is concerned solely with the location of visual field defects; depth and slope characteristics of scotomata are not examined. Finally, our sample sizes for glaucoma and optic nerve disease are rather small. Since this is a preliminary study, our data should not be interpreted as reflecting comprehensive popUlation characteristics for scotomata in glaucoma and optic nerve disease. 1060
Figure 1 displays the visual field locations that were involved in scotomata associated with glaucoma at six different frequency levels. The data are presented in the format for a right eye. The top left plot shows a composite distribution of the location of all 260 glaucoma scotomata. Visual field locations that were involved in glaucomatous scotomata at least 5%, 8%, 12%, 16%, and 20% of the time are also presented in Figure 1. These specific values were selected because they reflected meaningful transition points along the frequency distribution spectrum. Note that for the 5% frequency level, all of the scotomata are located within the paracentral arcuate region between 5 and 20 degrees radius from fixation. At successively higher frequency levels, there is a systematic progression from the nasal paracentral visual field (horizontal meridian) towards the blind spot region. Since the blind spot is often included in both superior and inferior arcuate defects, it is not surprising that it exhibits the highest frequency of occurrence for glaucomatous scotomata. The frequency distribution shown in Fig 1 is somewhat similar to those described by Aulhorn and her colleagues, 20,21 although some distinct differences are also apparent. These differences are probably due to two primary factors: (1) The frequency distribution reported by Aulhorn and colleagues 21 ,22 was for the first (earliest) glaucomatous visual field defect in their patient population, whereas our study included all forms of glaucomatous scotomata ranging from early to rather advanced stages. Since automated perimeters are used to test patients with glaucomatous defects at all stages of progression, the frequency distribution obtained in the present study is probably more applicable to this problem. (2) Aulhorn and her associates 21 ,22 used a rather coarse grid (approximately 400 visual field locations) to determine their frequency distributions, while our study was based on a much higher resolution grid (approximately 30,000 visual field locations separated by 0.75 degrees). Frequency distributions for 110 scotomata in optic nerve disease are presented in Figure 2 according to the format for a right eye. The patient popu1ation included cases of papilledema, ischemic optic neuropathy, optic atrophy, optic neuritis, optic nerve drusen, and other types of optic nerve dysfunction. Separate evaluations of each type of optic nerve disease were not performed because of the small sample size used in this investigation. Again, specific 'frequency levels were selected on the basis of reflecting distinct transition points in the frequency distribution spectrum. The frequency distribution pattern for optic nerve disease is different than that presented for glaucoma. At successively higher frequencies, there is a systematic progression of scotomatous locations towards the central 30 degrees, and then to the centrocecal region of the
JOHNSON AND KELTNER. COMPUTER ANALYSIS OF VISUAL FIELD LOSS
GLAUCOMA ALL DEFECTS
5%
-i<}
,e . . . . . .
, ,
. "" . . ,
, ,
... -3&
'
, ,
.. . , ,
-3&
••
-68
8%
,.,
,
12%
. .. ,.:) . .. , ,
, ,
, ,
"
16%
".
• t;e ,
.
-3&
, ,
,
."
. ,)., ,
,"
,.,
Fig. 1. Frequency distribution for 260 glaucoma scotomata. Early, moderate, and advanced visual field defects were included in the sample. Specific levels (all defects, 5%, 8%, 12%, 16%, and >20%) were selected to represent transition points along the frequency distribution spectrum.
>20%
,
J0,
, :)
=-, ,
t;e
I
I
JIll
".
I
t;e ,
,
visual field. This progresses to encompass only the central 5 degrees and the blind spot, and eventually only the blind spot region alone. The frequency distributions for scotomata in glaucoma and optic nerve disease provide useful preliminary information for the development of optimal target distribution patterns for detection of various types of visual field defects. As an additional prelude to defining optimal target patterns, we were interested in the capabilities and limitations of target distribution patterns for existing manual and automated peri metric tests. A computer simulation process was used to examine the detection capabilities oftarget distribution patterns from 20 automated and manual peri metric test pro-
J0,
,
•
t,
=-
I
I
t;e
I
I
JIll
cedures, as evaluated under equivalent, idealized conditions. Each target configuration was processed through the 370 individual scotomata from glaucoma and optic nerve disease to determine their maximum detection rate, based solely on the location of test stimuli. Detection rates for each of the 20 target configurations are presented by the cross-hatched portions of the histograms shown in Fig 3. The data are plotted from left to right in order of increasing number of locations in the target pattern, as indicated above each histogram. We have purposely avoided identification of specific devices or test procedures to discourage individuals from making inferences about their clinical performance on the 1061
OPHTHALMOLOGY • OCTOBER 1981
• VOLUME 88 • NUMBER 10
OPTIC NERVE DISEASE ALL DEFECTS
Itt.
5%
-Ol'
I.,
••
18
"i .....
...
,
.
••
•• 8%
. .•.
18%
.. '
••
,
••
,..'
-. . . . . . . . . . .
t
••••
Fig 2. Frequency distribution for 110 optic nerve disease scotomata. All types of optic nerve disease were included in the sample. Specific levels (all defects, 5%, 8%, 16%, 26%, and >30%) were selected to represent transition points along the frequency distribution spectrum.
-Ill
28%
10 • • li8 • •
-Ill
lie . . . . .
-Ill
as ••
118 • •
*'
10 • • .;e • • lie • •
';Je
-;Je
-. basis of these data alone. Most of the devices and test procedures also exhibit differences among features other than target configuration; our present findings thus provide only a partial analysis of overall performance capabilities. Other aspects may either improve or degrade the performance capabilities found in our study of target distribution patterns. Several features are of interest in Fig 3. First, there is a general increase in detection rate as the number of target locations increases, indicating that target density is an important factor in determining optimal detection rates. However, there are some variations in detection rate that cannot be attributed to target density, as shown by the irregularities (peaks and valleys) present in the histo1062
gram representations. This indicates that the pattern in which the targets are distributed is also an important factor underlying the detection of visual field defects. When more than 140 to 150 targets are employed, both the distribution pattern and the target density appear to have little or no additional effect on detection rate. The above evaluation is based upon idealized conditions, in which the' 'patient" does not make mistakes, become fatigued, or unreliable. If we assume that a single missed target could represent a , 'false alarm" due to patient error, and thereby require at least two targets to define a scotoma, the detection rate histograms for some test procedures change considerably (stippled regions of histograms in Fig 3). The effects of both target density
JOHNSON AND KELTNER • COMPUTER ANALYSIS OF VISUAL FIELD LOSS
rzz2Il or more spots 100
(::::::::12 or more spots Fig 3. Detection rates (%) for target distribution patterns associated with 20 existing manual and automated peri metric test procedures. To discourage comparisons on the basis of this factor alone, devices and test procedures are not identified by name . The number of targets for individual presentation patterns are indicated above each histogram. The cross-hatched region indicates the detection rate for scotomata identified by one or more targets, and the stippled area represents the detection rate for a criterion of two or more targets needed to identify scotomata (see text for explanation).
90
w
80
I-
<[
a::
70
Z 0
60
I-
U
w
IW 0
50 40 30
2
3
4
5
6
7
8
9
10 II
12 13 14 15 16 17 18 19 20
TARGET DISTRIBUTION PATTERNS and distribution pattern become more pronounced under these conditions, indicating that low target densities and certain distribution patterns are more susceptible to false alarms. Again, we note that this evaluation only examines one aspect of the test procedure. Some devices and screening procedures employ additional techniques to minimize the occurrence or significance of false alarms; these factors were not considered in the present analysis. The frequency distribution data in Figs 1 and 2, and the simulation results shown in Fig 3 clearly reflect the need for determining optimal target distribution patterns for detection of visual field defects with automated perimetry. Our present findings reveal several important qualitative aspects of specifying target distribution patterns. First, target distribution patterns should emphasize the paracentral arcuate region of the visual field for glaucoma and the centrocecal region of the visual field for optic nerve disease. Second, both the target density and the configuration of the target distribution pattern affect the detection rate. Third, some target configurations are more susceptible to patient errors or false alarms than others. Fourth, it does not appear that the detection rate is substantially improved by using more than 140 to 150 target locations. Since our preliminary findings have revealed several important aspects of target distribution patterns and detection rate, it is now nec-
essary to expand these studies and perform a systematic quantitative analysis of target density and distribution patterns to define optimal target configurations for detection of visual field defects. The diagnostic efficacy of manual or automated visual field test procedures must ultimately be determined by clinical evaluation studies. However, the use of computer simulation techniques for preliminary investigations can provide several advantages: (1) Individual components of the test procedure can be evaluated in isolation, or various combinations of factors can be analyzed systematically. (2) Desirable and undesirable alternatives can be rapidly differentiated. (3) In computer simulation studies, the visual field characteristics stored in the computer are known and serve as the reference standard. In clinical evaluations, the patient's visual field is inferred and the reference standard is usually another test procedure. (4) With computer simulation, maximum performance limits can be determined by eliminating errors and false alarms. If desired, the frequency and type of errors can be controlled to quantitatively evaluate their effects on test results. (5) All procedures and tests are conducted under equivalent conditions using the same information. Fankhauser and his associates16.17.19,20 have demonstrated the power and efficacy of simulation techniques to define automated perimetric procedures for clinical use. Further 1063
OPHTHALMOLOGY. OCTOBER 1981 • VOLUME 88 • NUMBER 10
utilization of this approach should greatly enhance the standardization of automated visual field testing and improve our understanding of perimetry in general.
ACKNOWLEDGMENTS We are grateful to Jan Horton for her significant contributions to the development of computer software algorithms for this study.
REFERENCES 1. Dannheim F. Clinical experiences with semi-automated perimeter ("Field master"). Presented at the "Weisbadener Tagung" of The German Ophthalmologists, Nov 20-26, 1978; Weisbaden, West Germany. 2. Dannheim F. Perimetry in Glaucoma. Presented at the Glaucoma Symposium, Amsterdam, Netherlands, Sept 21-22, 1979. 3. Fankhauser F, Koch P, Roulier A On automation of perimetry. Albrecht von Graefes Arch Klin Exp Ophthalmol 1972; 184: 126-50. 4. Fankhauser F, Spahr J, Bebie H. Three years of experience with the Octopus automatic perimeter. Doc Ophthalmol Proc Ser 1977; 14:7-15. 5. Fankhauser F, Spahr J, Bebie H. Some aspects of the automation of perimetry. Surv Ophthalmol 1977; 22:131-41. 6. Heijl A Automated perimetry in glaucoma visual field screening. A clinical study. Albrecht von Graefes Arch Klin Exp Ophthalmol1976; 200:21-37. 7. Heijl A Studies on computerized perimetry. Acta Ophthalmol 1977; Supp1132. 8. Heijl A, Drance SM. A clinical comparison of three computerized automatic perimeters in the detection of glaucoma defects. Doc Ophthalmol Proc Ser 1980; 26:43-8.
9. Heijl A, Krakau CET. An automatic static perimeter, design and pilot study. Acta Ophthalmol 1975; 53:293-310. 10. Johnson CA, Keltner JL. Comparison of manual and automated perimetry in 1,000 eyes. In: IEEE Computer Society Proceedings, St. Louis, 1978. Computers in Ophthalmology. New York: IEEE, 1979: 178-81. 11. Johnson CA, Keltner JL. Automated suprathreshold static perimetry. Am J Ophthalmol 1980; 89:731-41. 12. Johnson CA, Keltner JL. Comparative evaluation of the Autofield-I®, CFA-120®, and Fieldmaster Model 101-PR® automated perimeters. Ophthalmology 1980; 87:777-83. 13. Johnson CA, Keltner JL, Balestrery FG. Suprathreshold static perimetry in glaucoma and other optic nerve disease. Ophthalmology 1979; 86:1278-86. 14. Keltner JL, Johnson CA, Balestrery FG. Suprathreshold static perimetry: Initial clinical trials with the Fieldmaster automated perimeter. Arch Ophthalmol 1979; 97:260-72 15. Keltner JL, Johnson CA Capabilities and limitations of automated suprathreshold static perimetry. Doc Ophthalmol Proc Ser 1980; 26:49-55. 16. Bebie H, Frankhauser F, Spahr J. Static perimetry: Strategies. Acta Ophthalmol 1976; 54:325-38. 17. Fankhauser F, Bebie H. Threshold fluctuations, interpolations and spatial resolution in perimetry. Doc Ophthalmol Proc Ser 1979; 19:295-309. 18. Frisen M. Evaluation of peri metric procedures: A statistical approach. Doc Ophthalmol Proc Ser 1979; 19:427-31. 19. Gauger E. Computer simulation of examination procedures for the automatic "TO binger perimeter." Doc Ophthalmol Proc Ser 1977; 14:31-6. 20. Spahr J. Optimization of the presentation pattern in automated static perimetry. Vision Res 1975; 15: 1275-81. 21. Aulhorn E, Harms H. Early visual field defects in glaucoma. In: Leydhecker W, ed. Glaucoma, Tutzing Symposium, 1966. Basel; Karger: 1967: 151-86. 22. Aulhorn E, Karmeyer H. Frequency distribution in early glaucomatous visual field defects. Doc Ophthalmol Proc Ser 1977; 14:75-83.
Discussion by William M. Hart, Jr., MD, PhD There have been numerous recent developments in the application of computer technology to the problems of clinical perimetry. Coincident with these developments there have appeared a number of commercially available perimetric devices that attempt to exploit this technology, and which are generally, although somewhat loosely, referred to as "automated perimeters. " A clear distinction should be made at the outset between two broad classes of test strategies used by these devices. The first includes those that are suprathreshold static testing protocols. These are primarily intended to be used in a screening function to examine large numbers of patients who are not necessarily known to be diseased, but are From Washington University, St. Louis, Missouri.
1064
thought to be at risk. The second class of strategies includes those that actually determine brightness thresholds. Since the theoretical foundations for automated perimetry have been most firmly established for static testing protocols, I think it wisest to eliminate from consideration here those devices that purport to perform automated kinetic testing procedures. The major classes of instruments to be considered then includes, first, those that are essentially screening devices for the detection of previously unrecognized disease states, and second, those that are meant to be used for the high resolution definition of visual field defects for the purposes of comparison and sequential observation. The authors of the present work have undertaken the fundamentally important task of analyzing the theoretical efficiencies of examination patterns, employed by a
JOHNSON AND KELTNER • COMPUTER ANALYSIS OF VISUAL FIELD LOSS
variety of automated perimeters, for the detection of central and paracentral scotomatous defects found in glaucoma and optic nerve disease . That this sort of analysis is needed is reflected in their findings that the distribution of points is at least as important as their number and density in determining their probability of detection. This fact goes to the heart of the problem that the clinical practitioner should consider if he is going to use an automated testing device . fur what clinical purpose is it to be used? If, on the one hand, the need is for a screening device that will examine large numbers of patients who are suspected of having glaucoma, the pattern of suprathreshold testing should necessarily be different than that used for the detection of visual field defects associated with optic nerve disease or peripheral visual field defects, such as those associated with chiasmal or retrochiasmal disease. Equally important, if not more so, is that a screening device cannot be used for high resolution definition of the extent or morphology of a visual field defect. Note that the detection rates reported in this study to exceed 90 % require examination patterns having more than 100 points. While this may
be practical for single suprathreshold tests at each point , it is clearly not so for threshold testing, where the limits of subject performance are frequently exceeded with as few as 50 points during a single examination. Despite the automation of testing strategies, the selection of a strategy for each individual case remains in the realm of the art of perimetry. It is not possible to explore completely the visu al field of a patient during a single examination. Screening strategies may be appropriate for dealing with large numbers of people about whom little advance information is available . However, when history, visual function testing or physical findings suggest disease , or when a screening test has a positive result, the patient must be examined more carefully, and test strategy selection becomes more complex. Then a synthesis of available information must provide an idea of what kind of visual field defect is most likely to be present and where it is most likely to be located, so that an appropriate strategy may b e selected whether it be binocular confrontation testing, tangent screen testing, kinetic perimetry, or some pattern of static perimetry.
1065
Abstract: Preliminary investigations of optimal target distribution patterns for automated perimetry were conducted. In the first study, frequency distributions of visual field defects were determined in glaucoma (260 scotomata) and optic nerve disease (110 scotomata) for approximately 30,000 visual field locations. The frequency distributions in glaucoma and optic nerve disease were different, and suggest guidelines for development of target distribution patterns to achieve optimal detection rates. The second study included computer simulation of 20 target distribution patterns currently used in manual and automated perimetry. Each distribution pattern was processed through the 370 glaucoma and optic nerve disease scotomata to assess detection performance. Both the density and distribution of target locations affected overall detection rates. In addition, some target configurations were more susceptible to false alarms than others. These data provide qualitative information for optimization of target distribution patterns in automated perimetry, and serve as a foundation for future quantitative studies. [Key words: computer simulation, glaucoma, optic nerve disease, perimetry, visual field loss.] Ophthalmology 88: 1058-1 065, 1981
The field of automated perimetry began about 10 to 15 years ago with theoretical evaluations of its feasibility for performing clinical visual field testDepartment of Ophthalmology, * and Departments of Ophthalmology, Neurology and Neurological Surgery,t School of Medicine, University of California, Davis California. Presented at the Eighty-fifth Annual Meeting of the American Academy of Ophthalmology, Chicago, Illinois, November 2-7, 1980. Supported in part by National Eye Institute Academic Investigator Award #K07-EY00095.* Reprint requests to Chris A. Johnson, PhD, Department of Ophthalmology, University of California, Davis, CA 95616.
1058
ing. Since that time, it has progressed from a laboratory research phenomenon to the development of almost a dozen different commercially available devices. Most of these commercial devices have now been subjected to one or more clinical validation and comparison studies .1-15 The consensus from such investigations is that automated perimetry can be an effective procedure for detection (and in some instances detailed evaluation) of visual field loss, provided that specific test procedures, stimulus conditions, and methods of interpreting the test results are employed. We are currently entering a new phase of auto-
0161-6420/8111000/1058/$00.90
© American Academy of Ophthalmology
JOHNSON AND KELTNER. COMPUTER ANALYSIS OF VISUAL FIELD LOSS
mated perimetry, in which refinements are being developed to optimize existing test procedures. One approach to this problem consists of direct clinical comparison studies of different automated devices using rigorous manual procedures as a reference standard. 2,8,12,15 This approach has many desirable features, although there are also several disadvantages. First, it is not possible for patients to perform in a consistent, error-free manner for successive test procedures. Thus, there will always be some differences in performance among various devices that are due to patient variability, fatigue, and other factors not directly related to the test procedure. A second problem with clinical comparison studies is the interpretation of results. Since various devices differ according to their test strategies, stimulus conditions, and other relevant features, it is difficult to evaluate the contribution of individual aspects of the test to overall performance characteristics. In addition, such comparison studies are limited to evaluation of the features available on each device. Not only does this initially constrain the possible solutions for optimizing test procedures, but it also requires that revalidation studies be performed following any modifications of the devices or test procedures. As we have discovered through previous experience,1O-15 this can be a very tedious and inefficient use of research efforts. An alternative approach to clinical comparison studies in automated perimetry is the use of digital computer processing and simulation techniques. 16-20 With this approach, it is possible to have "patient' , data and performance characteristics that are the same for all devices or procedures evaluated. In addition, one can isolate individual components of the test procedure, and rapidly evaluate many potential solutions for optimal procedures. This makes it possible to quickly eliminate unsuccessful alternatives, define relevant properties of individual components of the test, and suggest the most effective test procedure. Clinical comparison studies can then be performed for this procedure to validate its effectiveness as a diagnostic tool. Fankhauser and his associates 16 ,17,19,20 have demonstrated the efficacy of this approach to define optimal threshold determination strategies for automated static perimetry with subsequent implementation in a clinical test device. The present study describes our preliminary efforts to utilize digital computer processing and simulation techniques to define optimal target distribution patterns for detection of visual field defects. Two studies were conducted: (1) using a high resolution grid (0.75 by 0.75 degree elements), frequency distributions of scotomata in glaucoma and optic nerve disease were determined for approximately 30,000 locations in the visual field; and (2) detection rates for target distribution patterns associated with 20 existing manual and
automated test procedures were determined through computer simulation of tests performed on 370 scotomata from glaucoma and optic nerve disease.
MATERIALS AND METHODS The visual field data for this study were obtained from previous kinetic visual field examinations performed on the Goldmann perimeter. Each visual field chart was selected from our patient records, and had been performed within the past four years. Most of these patients had participated in previous clinical comparison studies of manual and automated perimetry. A total of 370 scotomata were entered into the computer, 260 from glaucoma patients and 110 from patients with optic nerve disease. fur glaucoma, a distribution of early, moderate, and advanced visual field defects was chosen. fur optic nerve disease patients, a variety of anomalies at various stages of progression were selected. The population included ischemic optic neuropathy, optic atrophy, optic neuritis, papilledema, optic nerve drusen, and other optic nerve abnormalities. Only scotomata were considered for this analysis; isopter-related deficits (eg, nasal steps) were not included. An LSI-II minicomputer system, a Tektronix® Model 4953 digitizing graphics tablet and a Tektronix Model 4010 graphics terminal were used to process the visual field information. To achieve proper alignment, each visual field chart was positioned on the digitizing tablet and the coordinates of the center and right horizontal axis (0 degree meridian, 90 degree radius) were entered by a graphics pen to check the centering, orientation (rotation), and scaling of the visual field chart with respect to the tablet. fullowing this procedure, the scotoma data were entered. If more than one target had been used to plot the scotoma originally, the largest representation of the visual field defect was used. Information pertaining to the intensity of the scotoma (depth and slope characteristics) was not considered in this investigation. Each scotoma was entered by tracing its outer boundary on the digitizing tablet using the graphics pen. This boundary trace was simultaneously displayed on the graphics terminal to check for proper data input. A "window" of 0.75 by 0.75 degrees was used to control the amount of data generated by the tracing procedure. That is, the graphics pen was required to move a distance of 0.75 degrees in either the X (horizontal) or Y (vertical) direction to record each successive data point. The X- Y coordinates of each data point were then ordered, and a software algorithm was used to regenerate the total area within the scotoma, using only the boundary information. The algorithm defines the area within the scotoma by successive "slicing" along the X axis. 1059
OPHTHALMOLOGY. OCTOBER 1981 • VOLUME 88 • NUMBER 10
For a given X coordinate value, the lowest Y coordinate value is determined, and locations between it and the next Y coordinate value are filled in. If additional Y coordinates are present for this X coordinate, the procedure is repeated; otherwise, the algorithm selects the next X coordinate value and then repeats the process. The algorithm continues until the entire scotoma has been .. sliced." Under some circumstances (vertical or near-vertical boundaries, complex shapes with multiple convolutions, and regions of partial over-lap during tracing), this algorithm becomes confused as to what is inside and what is outside the scotoma. These ambiguities are resolved by keeping track of changes in direction of travel of the tracing pen. This program provides a reliable method of reconstructing the complete scotoma from boundary information, which allows us to retain the essential features of the scotoma while storing a minimum amount of data. To check for missing or erroneous data, the" filled-in" scotoma generated by this algorithm was displayed on the graphics terminal. Once the information had been checked and stored, it was used for two purposes in the present study. First, each scotoma was entered into a composite visual field representation of 30,000 locations (0.75 by 0.75 degree grid size) to obtain frequency distributions for visual field defects in glaucoma and optic nerve disease. The second purpose of the study was to simulate the target distribution patterns of most of the existing manual and automated test procedures used to detect visual field characteristics. All target distribution patterns were entered through the digitizing tablet and checked by procedures similar to the process used for scotomata. Each distribution pattern was then processed through the 370 scotomata, the number and location of targets that fell within each scotoma were recorded, and the information was then used to determine the detection performance of each simulated test procedure.
RESULTS AND DISCUSSION Before describing our findings, it is important to point out the limitations of the present study. First, our investigation only evaluates scotomatous visual field loss. Visual deficits such as nasal steps and other isopter-related abnormalities are not considered in this phase of our work. Secondly, this study is concerned solely with the location of visual field defects; depth and slope characteristics of scotomata are not examined. Finally, our sample sizes for glaucoma and optic nerve disease are rather small. Since this is a preliminary study, our data should not be interpreted as reflecting comprehensive popUlation characteristics for scotomata in glaucoma and optic nerve disease. 1060
Figure 1 displays the visual field locations that were involved in scotomata associated with glaucoma at six different frequency levels. The data are presented in the format for a right eye. The top left plot shows a composite distribution of the location of all 260 glaucoma scotomata. Visual field locations that were involved in glaucomatous scotomata at least 5%, 8%, 12%, 16%, and 20% of the time are also presented in Figure 1. These specific values were selected because they reflected meaningful transition points along the frequency distribution spectrum. Note that for the 5% frequency level, all of the scotomata are located within the paracentral arcuate region between 5 and 20 degrees radius from fixation. At successively higher frequency levels, there is a systematic progression from the nasal paracentral visual field (horizontal meridian) towards the blind spot region. Since the blind spot is often included in both superior and inferior arcuate defects, it is not surprising that it exhibits the highest frequency of occurrence for glaucomatous scotomata. The frequency distribution shown in Fig 1 is somewhat similar to those described by Aulhorn and her colleagues, 20,21 although some distinct differences are also apparent. These differences are probably due to two primary factors: (1) The frequency distribution reported by Aulhorn and colleagues 21 ,22 was for the first (earliest) glaucomatous visual field defect in their patient population, whereas our study included all forms of glaucomatous scotomata ranging from early to rather advanced stages. Since automated perimeters are used to test patients with glaucomatous defects at all stages of progression, the frequency distribution obtained in the present study is probably more applicable to this problem. (2) Aulhorn and her associates 21 ,22 used a rather coarse grid (approximately 400 visual field locations) to determine their frequency distributions, while our study was based on a much higher resolution grid (approximately 30,000 visual field locations separated by 0.75 degrees). Frequency distributions for 110 scotomata in optic nerve disease are presented in Figure 2 according to the format for a right eye. The patient popu1ation included cases of papilledema, ischemic optic neuropathy, optic atrophy, optic neuritis, optic nerve drusen, and other types of optic nerve dysfunction. Separate evaluations of each type of optic nerve disease were not performed because of the small sample size used in this investigation. Again, specific 'frequency levels were selected on the basis of reflecting distinct transition points in the frequency distribution spectrum. The frequency distribution pattern for optic nerve disease is different than that presented for glaucoma. At successively higher frequencies, there is a systematic progression of scotomatous locations towards the central 30 degrees, and then to the centrocecal region of the
JOHNSON AND KELTNER. COMPUTER ANALYSIS OF VISUAL FIELD LOSS
GLAUCOMA ALL DEFECTS
5%
-i<}
,e . . . . . .
, ,
. "" . . ,
, ,
... -3&
'
, ,
.. . , ,
-3&
••
-68
8%
,.,
,
12%
. .. ,.:) . .. , ,
, ,
, ,
"
16%
".
• t;e ,
.
-3&
, ,
,
."
. ,)., ,
,"
,.,
Fig. 1. Frequency distribution for 260 glaucoma scotomata. Early, moderate, and advanced visual field defects were included in the sample. Specific levels (all defects, 5%, 8%, 12%, 16%, and >20%) were selected to represent transition points along the frequency distribution spectrum.
>20%
,
J0,
, :)
=-, ,
t;e
I
I
JIll
".
I
t;e ,
,
visual field. This progresses to encompass only the central 5 degrees and the blind spot, and eventually only the blind spot region alone. The frequency distributions for scotomata in glaucoma and optic nerve disease provide useful preliminary information for the development of optimal target distribution patterns for detection of various types of visual field defects. As an additional prelude to defining optimal target patterns, we were interested in the capabilities and limitations of target distribution patterns for existing manual and automated peri metric tests. A computer simulation process was used to examine the detection capabilities oftarget distribution patterns from 20 automated and manual peri metric test pro-
J0,
,
•
t,
=-
I
I
t;e
I
I
JIll
cedures, as evaluated under equivalent, idealized conditions. Each target configuration was processed through the 370 individual scotomata from glaucoma and optic nerve disease to determine their maximum detection rate, based solely on the location of test stimuli. Detection rates for each of the 20 target configurations are presented by the cross-hatched portions of the histograms shown in Fig 3. The data are plotted from left to right in order of increasing number of locations in the target pattern, as indicated above each histogram. We have purposely avoided identification of specific devices or test procedures to discourage individuals from making inferences about their clinical performance on the 1061
OPHTHALMOLOGY • OCTOBER 1981
• VOLUME 88 • NUMBER 10
OPTIC NERVE DISEASE ALL DEFECTS
Itt.
5%
-Ol'
I.,
••
18
"i .....
...
,
.
••
•• 8%
. .•.
18%
.. '
••
,
••
,..'
-. . . . . . . . . . .
t
••••
Fig 2. Frequency distribution for 110 optic nerve disease scotomata. All types of optic nerve disease were included in the sample. Specific levels (all defects, 5%, 8%, 16%, 26%, and >30%) were selected to represent transition points along the frequency distribution spectrum.
-Ill
28%
10 • • li8 • •
-Ill
lie . . . . .
-Ill
as ••
118 • •
*'
10 • • .;e • • lie • •
';Je
-;Je
-. basis of these data alone. Most of the devices and test procedures also exhibit differences among features other than target configuration; our present findings thus provide only a partial analysis of overall performance capabilities. Other aspects may either improve or degrade the performance capabilities found in our study of target distribution patterns. Several features are of interest in Fig 3. First, there is a general increase in detection rate as the number of target locations increases, indicating that target density is an important factor in determining optimal detection rates. However, there are some variations in detection rate that cannot be attributed to target density, as shown by the irregularities (peaks and valleys) present in the histo1062
gram representations. This indicates that the pattern in which the targets are distributed is also an important factor underlying the detection of visual field defects. When more than 140 to 150 targets are employed, both the distribution pattern and the target density appear to have little or no additional effect on detection rate. The above evaluation is based upon idealized conditions, in which the' 'patient" does not make mistakes, become fatigued, or unreliable. If we assume that a single missed target could represent a , 'false alarm" due to patient error, and thereby require at least two targets to define a scotoma, the detection rate histograms for some test procedures change considerably (stippled regions of histograms in Fig 3). The effects of both target density
JOHNSON AND KELTNER • COMPUTER ANALYSIS OF VISUAL FIELD LOSS
rzz2Il or more spots 100
(::::::::12 or more spots Fig 3. Detection rates (%) for target distribution patterns associated with 20 existing manual and automated peri metric test procedures. To discourage comparisons on the basis of this factor alone, devices and test procedures are not identified by name . The number of targets for individual presentation patterns are indicated above each histogram. The cross-hatched region indicates the detection rate for scotomata identified by one or more targets, and the stippled area represents the detection rate for a criterion of two or more targets needed to identify scotomata (see text for explanation).
90
w
80
I-
<[
a::
70
Z 0
60
I-
U
w
IW 0
50 40 30
2
3
4
5
6
7
8
9
10 II
12 13 14 15 16 17 18 19 20
TARGET DISTRIBUTION PATTERNS and distribution pattern become more pronounced under these conditions, indicating that low target densities and certain distribution patterns are more susceptible to false alarms. Again, we note that this evaluation only examines one aspect of the test procedure. Some devices and screening procedures employ additional techniques to minimize the occurrence or significance of false alarms; these factors were not considered in the present analysis. The frequency distribution data in Figs 1 and 2, and the simulation results shown in Fig 3 clearly reflect the need for determining optimal target distribution patterns for detection of visual field defects with automated perimetry. Our present findings reveal several important qualitative aspects of specifying target distribution patterns. First, target distribution patterns should emphasize the paracentral arcuate region of the visual field for glaucoma and the centrocecal region of the visual field for optic nerve disease. Second, both the target density and the configuration of the target distribution pattern affect the detection rate. Third, some target configurations are more susceptible to patient errors or false alarms than others. Fourth, it does not appear that the detection rate is substantially improved by using more than 140 to 150 target locations. Since our preliminary findings have revealed several important aspects of target distribution patterns and detection rate, it is now nec-
essary to expand these studies and perform a systematic quantitative analysis of target density and distribution patterns to define optimal target configurations for detection of visual field defects. The diagnostic efficacy of manual or automated visual field test procedures must ultimately be determined by clinical evaluation studies. However, the use of computer simulation techniques for preliminary investigations can provide several advantages: (1) Individual components of the test procedure can be evaluated in isolation, or various combinations of factors can be analyzed systematically. (2) Desirable and undesirable alternatives can be rapidly differentiated. (3) In computer simulation studies, the visual field characteristics stored in the computer are known and serve as the reference standard. In clinical evaluations, the patient's visual field is inferred and the reference standard is usually another test procedure. (4) With computer simulation, maximum performance limits can be determined by eliminating errors and false alarms. If desired, the frequency and type of errors can be controlled to quantitatively evaluate their effects on test results. (5) All procedures and tests are conducted under equivalent conditions using the same information. Fankhauser and his associates16.17.19,20 have demonstrated the power and efficacy of simulation techniques to define automated perimetric procedures for clinical use. Further 1063
OPHTHALMOLOGY. OCTOBER 1981 • VOLUME 88 • NUMBER 10
utilization of this approach should greatly enhance the standardization of automated visual field testing and improve our understanding of perimetry in general.
ACKNOWLEDGMENTS We are grateful to Jan Horton for her significant contributions to the development of computer software algorithms for this study.
REFERENCES 1. Dannheim F. Clinical experiences with semi-automated perimeter ("Field master"). Presented at the "Weisbadener Tagung" of The German Ophthalmologists, Nov 20-26, 1978; Weisbaden, West Germany. 2. Dannheim F. Perimetry in Glaucoma. Presented at the Glaucoma Symposium, Amsterdam, Netherlands, Sept 21-22, 1979. 3. Fankhauser F, Koch P, Roulier A On automation of perimetry. Albrecht von Graefes Arch Klin Exp Ophthalmol 1972; 184: 126-50. 4. Fankhauser F, Spahr J, Bebie H. Three years of experience with the Octopus automatic perimeter. Doc Ophthalmol Proc Ser 1977; 14:7-15. 5. Fankhauser F, Spahr J, Bebie H. Some aspects of the automation of perimetry. Surv Ophthalmol 1977; 22:131-41. 6. Heijl A Automated perimetry in glaucoma visual field screening. A clinical study. Albrecht von Graefes Arch Klin Exp Ophthalmol1976; 200:21-37. 7. Heijl A Studies on computerized perimetry. Acta Ophthalmol 1977; Supp1132. 8. Heijl A, Drance SM. A clinical comparison of three computerized automatic perimeters in the detection of glaucoma defects. Doc Ophthalmol Proc Ser 1980; 26:43-8.
9. Heijl A, Krakau CET. An automatic static perimeter, design and pilot study. Acta Ophthalmol 1975; 53:293-310. 10. Johnson CA, Keltner JL. Comparison of manual and automated perimetry in 1,000 eyes. In: IEEE Computer Society Proceedings, St. Louis, 1978. Computers in Ophthalmology. New York: IEEE, 1979: 178-81. 11. Johnson CA, Keltner JL. Automated suprathreshold static perimetry. Am J Ophthalmol 1980; 89:731-41. 12. Johnson CA, Keltner JL. Comparative evaluation of the Autofield-I®, CFA-120®, and Fieldmaster Model 101-PR® automated perimeters. Ophthalmology 1980; 87:777-83. 13. Johnson CA, Keltner JL, Balestrery FG. Suprathreshold static perimetry in glaucoma and other optic nerve disease. Ophthalmology 1979; 86:1278-86. 14. Keltner JL, Johnson CA, Balestrery FG. Suprathreshold static perimetry: Initial clinical trials with the Fieldmaster automated perimeter. Arch Ophthalmol 1979; 97:260-72 15. Keltner JL, Johnson CA Capabilities and limitations of automated suprathreshold static perimetry. Doc Ophthalmol Proc Ser 1980; 26:49-55. 16. Bebie H, Frankhauser F, Spahr J. Static perimetry: Strategies. Acta Ophthalmol 1976; 54:325-38. 17. Fankhauser F, Bebie H. Threshold fluctuations, interpolations and spatial resolution in perimetry. Doc Ophthalmol Proc Ser 1979; 19:295-309. 18. Frisen M. Evaluation of peri metric procedures: A statistical approach. Doc Ophthalmol Proc Ser 1979; 19:427-31. 19. Gauger E. Computer simulation of examination procedures for the automatic "TO binger perimeter." Doc Ophthalmol Proc Ser 1977; 14:31-6. 20. Spahr J. Optimization of the presentation pattern in automated static perimetry. Vision Res 1975; 15: 1275-81. 21. Aulhorn E, Harms H. Early visual field defects in glaucoma. In: Leydhecker W, ed. Glaucoma, Tutzing Symposium, 1966. Basel; Karger: 1967: 151-86. 22. Aulhorn E, Karmeyer H. Frequency distribution in early glaucomatous visual field defects. Doc Ophthalmol Proc Ser 1977; 14:75-83.
Discussion by William M. Hart, Jr., MD, PhD There have been numerous recent developments in the application of computer technology to the problems of clinical perimetry. Coincident with these developments there have appeared a number of commercially available perimetric devices that attempt to exploit this technology, and which are generally, although somewhat loosely, referred to as "automated perimeters. " A clear distinction should be made at the outset between two broad classes of test strategies used by these devices. The first includes those that are suprathreshold static testing protocols. These are primarily intended to be used in a screening function to examine large numbers of patients who are not necessarily known to be diseased, but are From Washington University, St. Louis, Missouri.
1064
thought to be at risk. The second class of strategies includes those that actually determine brightness thresholds. Since the theoretical foundations for automated perimetry have been most firmly established for static testing protocols, I think it wisest to eliminate from consideration here those devices that purport to perform automated kinetic testing procedures. The major classes of instruments to be considered then includes, first, those that are essentially screening devices for the detection of previously unrecognized disease states, and second, those that are meant to be used for the high resolution definition of visual field defects for the purposes of comparison and sequential observation. The authors of the present work have undertaken the fundamentally important task of analyzing the theoretical efficiencies of examination patterns, employed by a
JOHNSON AND KELTNER • COMPUTER ANALYSIS OF VISUAL FIELD LOSS
variety of automated perimeters, for the detection of central and paracentral scotomatous defects found in glaucoma and optic nerve disease . That this sort of analysis is needed is reflected in their findings that the distribution of points is at least as important as their number and density in determining their probability of detection. This fact goes to the heart of the problem that the clinical practitioner should consider if he is going to use an automated testing device . fur what clinical purpose is it to be used? If, on the one hand, the need is for a screening device that will examine large numbers of patients who are suspected of having glaucoma, the pattern of suprathreshold testing should necessarily be different than that used for the detection of visual field defects associated with optic nerve disease or peripheral visual field defects, such as those associated with chiasmal or retrochiasmal disease. Equally important, if not more so, is that a screening device cannot be used for high resolution definition of the extent or morphology of a visual field defect. Note that the detection rates reported in this study to exceed 90 % require examination patterns having more than 100 points. While this may
be practical for single suprathreshold tests at each point , it is clearly not so for threshold testing, where the limits of subject performance are frequently exceeded with as few as 50 points during a single examination. Despite the automation of testing strategies, the selection of a strategy for each individual case remains in the realm of the art of perimetry. It is not possible to explore completely the visu al field of a patient during a single examination. Screening strategies may be appropriate for dealing with large numbers of people about whom little advance information is available . However, when history, visual function testing or physical findings suggest disease , or when a screening test has a positive result, the patient must be examined more carefully, and test strategy selection becomes more complex. Then a synthesis of available information must provide an idea of what kind of visual field defect is most likely to be present and where it is most likely to be located, so that an appropriate strategy may b e selected whether it be binocular confrontation testing, tangent screen testing, kinetic perimetry, or some pattern of static perimetry.
1065