- Email: [email protected]
Quantitative coronary cineangiography for the study of atherosclerosis
Phys. Vol. I?, No. 5, pp. 356365, 1995 Copyright 0 1995 Elsevier Science Ltd for BES Printed in Great Britain. All rights reserved 1350-4533/95 $10.00 + 0.00
Med. Eng. ERWORTH EINEMANN
Qu~~~~ve coronary c~e~~o~aphy study of atherosclerosis J.N.H. Brunt*, D.J. Coltart’
G.F. Watts+, B. Lewis+, L.D.R.
Smith’
for the
and
“Dept of Medical Biophysics, Stopford Building, Manchester Universi~, Manchester Ml3 9PT, UK; +Depts of Endocrinology & Chemical Pathology %ardiology, UMDS, St Thomas’ Hospital, London SE1 7EH, UK Received
November
1993, accepted
February
and
1994
ABSTRACT An~o~ap~y is the ~~n~ti~e ~ocedurefor c~ar~ct~ing the extent and course ojcoron~~ artery diseme. We describe the ~thodo~~ retired to measure, with sternal r%~o~v~ngpow~, ~n~~uph~c chenps in canal arty disease. We utilised recent te~hn~~~‘~~~ d~e~p~ts in image di~.t~~tion, storage and analysis. The measures of change quant@ed both diffuse and focal atherosclerosis. Frames from angiographic tine films were digitized at high resolution (1024 x 1024 pixels, 8 bit grq scale) and archived on optical disk. Four radiographic projections were stored to ensure good visualization qfas many as possible of a set qf tenmajor arterial segments. Edges of segments and catheter were automatically delineated by computer using a dynamic programming algorithm involving a costfunction which contained terms based on edge strength and on continuity. For every digitized radiographic projection, delineation, was repeated in three adjacent frames, to improve precision. Edge points for each corona9 segment were stored on disk. From these we computed the mean width along the sent (pixelsj. Scaling to obtain the Mean Absolute Width of the SeFt ~.~~, mm) was achieved using catheter dirn~s~o~s known porn rni~o~~, systematic error due to imaging system line-spread function being corrected using data from computer simulations and phantom studies. Correction for geometric image intensifier distortion was also applied. We used the methodology in a randomized, controlled trial of the effect of lipid-l owering therapy, the St Thomas ’ Atherosclerosis Regression Study. The jundamental measure of change of disease in each segment was the change in MAWS (AMAWS}. Using in-vitro and in-vivo studies we established that the overall resolving power for one segment AMAWS was 0. IO mm at 2 mm width and 0.14 mm at 4 mm width. Subsidia? end+oints were the change (A) in minimum absolute width of segment (~~~nA~~, edge ~~e~~a~ty index (EZZ) and #rcent d~~~t~ stenosis (%DS). A %DS {the &onv~t~o~~L angiographic measure ojc~ona~ disease) was si~~~~~ntly correlated with change in all indices, closest chelation being seen with AEZI (r = 0.94, p
diagnostic
image pro-
1995, Vol. 17, 356-365, July
INTRODUCTION In assessing the natural history of coronary atherosclerosis, and in trials to determine treatment effects, the preferred imaging technique is serial coronary angiography. X-ray projection images of the con~ast~ont~ning coronary arteries are obtained from various directions to ensure good visualization of all the major vessels and of any lesions that may be present. The X-ray image transduction chain incorporates an image Correspondence to: DrJ.N.H. Brunt, Department of Medical Physics, CIat&erbridge Centre for Oncology, Bebington, Wirral, Merseyside, L63 4JY, UK. rPresent address: University Department of Medicine, University of Western Austrafia, Box X2213 GPO Perth, WA 6001, Australia.
intensifier to permit images to be recorded at a rate of 25 orA more per-second while avoiding excessive radiation dose to the patient. The standard recording medium is 35 mm tine film’, though digital recording may replace this in the future. Visual interpretation of coronary angiograms is informative but has limitations: (1) in&a- and inter-observer variability, (2) only relative calibres (e.g. percent stenosis) can be estimated, (3) it cannot form the basis from which to calculate flow resistance and hence assess the physiological significance of the disease. These limitations stimulated the development of computerized techniques for quantitative coronary angiography (QCA). Quantitation is now used at many centres
world~vide, and the similarities and differences among the leading centres have been documented in the reviews by Reibe?. Quantitative analysis most commonly involves measuring changes in arterial segments over a period of time by. comparing results obtained from two films. In this paper we describe our methodology for QCA. It has been used in the St Thomas’ Atherosclerosis Regression study (STARS)” to measure, with optimal resolving power, angiographic changes in coronary artery disease. We utilized recent technological developments in image digitizati(~n, storage and analysis. The changes measured in STARS quantified both diffuse and focal atherosclerosis. METHODS Analysis apparatus
Corona? images on 35 mm film were selected for ~igitizatlon usmg a Tagarno dual-channel projectlon system and subsequently projected into an electronic camera with digital output suitable for processing by computer. Figure 1 shows block cliagrams of the apparatus used. I;iEm ~~~~~(.~~~~~~a~~~ ~~t~~~,~~~. hatching frames from film-pairs were selected using the dual-channel projector, with facilities for taking Polaroid
a) Cine angiographic
b) Frame
selection,
Imaging
using
dual-channel
projection
projection screen
processed films
C) Image digiti~tion
pictures digitizing
of the selected frames.
frames,
to be used when
Film projector for dig-itizaticrn. Because the illumination intensity produced by a normal angiographic projector would be too high to project directly onto the solid state sensor of the digitizing camera, the condenser optics were designed to provide uniform but low intensity illumination. Compared with standard optics, there were three changes. First, the usual lamp of several hundred watts power was replaced with a 10 W quartz halogen lamp. Secondly, no condenser mirror was used. Thirdly, the condenser lens was moved approximately 0.1 m away from the light source: this reduced, by an order of magnitude, the solid angle it subtended while improvitlg the uniformit~7 ~~f~ill~~rni~ation. Thus overall ill~irnin~~ti~~n intensitv at the film was approximatetv 0.5% of normal. If$@&ing ~‘(zr~era. A Videk Megaplus camera (Lastman Kodak Company, Rochester, USA) with a charge coupled device solid state sensor array of 1320 x 1035 pixel resolution and with a-bit digital output, was used. The camera was digitally linked to the analysis computer. Optics were chosen so that each 1 mm square of tine iihn corresponded to a 55 x 55 pixel portion of digitized image. In most tine frames analysed, data storage was limited to the central 1024 x 1024 pixels of the sensor array (i.e. 1 megabyte, MB, per image). This corresponded to a square of film of 18.6 mm side length, and therefore encompassed virtually all the exposed emulsion of the complete (&neformat) frame. The heart was imaged onto the film with a magnification of approximately 0.17; hence each pixel corresponded to a square of side typically 0.11 mm at the heart. ~~~)r~~~Ltirlg environ merit md storage of image c&n. All digitized images, typically twenty to thirty per patient, were stored at fLll1 resolution using both write-once-read-multiple (Optimem) and rewritable (Sony) optical disks, the former providing archiving and the latter being used for day-to-day access. These devices were networked to a Sun workstation (running Pascal programs), with an image manipulation co-processor which could hold multiple images or portions of images. This was convenient for region selection and matching in adjacent film frames and between serial films. When comparing matched angiographic projections from two films, 6 MB of memory stored digitized versions of three adja.cent frames from each film. Analysis protocol
Image analysis computer
Figure s&rtion
1
Block diagrams of (a) tine-film acquisition, and (t‘) image (ii~~ti~~ti~n and anatysis systems
magnetic and optical disks
(b) frame
~a~~~~&~~~~~~~ ~ab~~~~ ~~o~oGo~~s.To rninimise variation between first and second an~ograms a standard approach was employed: Judkins percutaneous femoral artery technique; ‘Cordis’ 8F polyurethane catheters; “Urografin’ contrast medium; no vasodilators. Four views of the right and 7 of the left coronary artery were recorded on ‘Agfa Scopix RPl’ 35 mm tine ftlm, along with
357
Quantitation
of coronary
[email protected]:
J.N.H.
Brunt
et al.
images of a grid (1 cm spacings) attached to the image intensifier to allow correction of pincushion distortion. Radiographic angulations, table height, and the sequence of injections (which predominantly influenced the elapsed time from first injection) were recorded and reproduced as exactly as possible, by the same cardiologist, at repeat angiography. segment selection. We investigated ten major arterial segments: the proximal, middle and distal segments of the right coronary artery, the left main stem, the proximal and distal circumflex artery, the proximal, middle and distal segments of the left anterior descending artery, and the obtuse marginal artery, using the anatomical definitions of the American Heart Association5. Paired, coded films were compared on the dualchannel projector and paired frames (i.e. images matched between serial studies) of the same radiographic views (typically four in each film) were photographed. The frames selected for analysis were taken at enddiastole (as assessed from visual comparison of frames before and after the selected frame). They had to clearly demonstrate particular segments in the same frame as the catheter tip. There had to be no nearly parallel overlying vessels, and the vessels had to be uniformly filled with contrast medium. Polaroid photographs were coded, and marked to identify the segments for image analysis. Coding was done by an investigator who was aware of the film sequence but blinded to the therapeutic group. Films were then transferred to the location of the digitizing projector/camera. Digitization of the coded frames and analysis of the resulting digital images (guided by photographs) were done by an investigator blinded to the sequence and therapeutic group of the angiograms and photographs. Arterial
Segment boundaly delineation and width function. The anatomically defined length of each
analysable segment was input to the analysis computer by pointing to a few points along its centreline. The completely automatic boundary delineation process (Figure 2) then took place along the whole length, as follows. A set of scan rectangles (of length typically 60 pixels and of several pixels Scan rectangles
for sampling
width) perpendicular to the vessel centre line sampled profiles of grey values in the image. The scan rectangles were spaced at intervals of four pixels - approximately 0.45 mm - along the centreline. To find each side of the vessel a computer algorithm searched for the eight most significant correct-direction (grey value rising or falling) edges (i.e. peaks or troughs in the first derivative function) along each entire profile. The edge points from the set of profiles were used as candidate boundary points in a dynamic programming algorithm6,‘. This found the lowest cost path using a cost function which contained terms based on edge strength and on continuity. A width function was derived from the distances of closest approach of the two boundaries and a three-point smoothing function was applied to this function. In measuring widths, we followed a suggestion of Selzer et al* to improve precision by averaging the results derived from three sequential tine frames (the frames themselves were not averaged). Since there were, deliberately, no facilities for manual ‘correction’ of boundaries, if boundary detection was obviously artefactual, the corresponding width function was omitted from the average. Measurements obtained from vessel boundaries: angiographic indices. The measurements obtained
from the width function are described below. All were ‘derived after correcting for pincushion distortion and (for absolute measures) applying a calibration factor. Comparison of change in angiographic indices was carried out by linear regression analysis. (a) Mean absolute width of segment (MAWS). Width measurements expressed in pixels were obtained along an entire arterial segment as described (see Figure 3) and the mean value of width along the segment was computed. The calibration factor (9.00 pixels per mm was a typical value) was used to convert the pixel values to absolute units (mm). Differential magnification between the catheter and the segment (due to X-ray beam divergence) was not corrected for, but the variation of its possible magnitude between similarly angulated views in successive examinations of the same patient was estimated in quoting the overall resolving power of the system.
image grey values width
interpolated at minimum
width
I
position of min. width distance -eeee-
Vessel centreline
-
Vessel
boundaries
to be delineated
Figure 2 Diagrammatic representation ing boundaries of coronary vessels
358
of methodology
for delineat-
along vessel
figure 3 Schematic diagram showing the ‘width function arterial segment. It was from the set of width measurements tuting this function that the changes in the angiographic MAWS, MinAWS, %DS and EII were computed
for an constiindices
(b) Minimum absolute width of sepment (MinAWS). The minimum width was c;mputed as the minimum value among the width values spaced along the segment (i.e. re-using the data required to calculate MAWS). (cl Percentage diameter stenosis (%DS). Percentage diameter stenosis was derived from the ratio of the minimum width along a segment to a computed width representing an approximation to the width that the vessel would have had at that point in the absence of disease. It was obtained by iteratively fitting (by least squares) a straight line to the plot of width values against distance along the vessel. When performing the fit, the weights applied to values corresponding to the part of the segment containing the minimum width were a function of the degree of stenosis computed at each iteration and were progressively reduced during the three iterations, so that the line was ultimately fitted essentially to the widths of the less diseased parts of the segment only. (d) Edge irregularity index (EII). The edge irregularity index used was the ratio obtained by dividing the standard deviation of the (abovementioned) least squares fit by the value of MAWS for the segment. Accounting unvisualised
for data ji-om different views segments. When a segment
and j&r
was visualized and calibrated in more than one view, its MinAWS (and MAWS) values were computed by averaging across the views. The %DS value used was that from the view producing the most severe value. Segments Lln~~sualized because of total occlusion proximal to the measurement site could not undergo edge-detection, but overall assessment would have been seriously incomplete without accounting for them. Therefore we defined a ‘functional width’ for any such segment: if blood did not reach a more distal segment because of an occlusion in a more proximal segment, the functional width of the more distal segment was zero, even though we did not know whether this unvisualized segment was itself occluded. Segments perfused and visualized as a result of coilateral flow were measured in the usual way. Calibration
and correction
of systematic
errors
~a~i~atio~ te&hn~que. The process of calibration converts measurements of vessel widths expressed in pixels to measurements expressed in absolute units (mm). We detected the edges of the catheter using the same automated delineation algorithm used for the vessels, and achieved calibration by comparison with a table of catheter tip diameters, obtained by micrometry. Line stead function. The use of detection of extrema in the first derivative within the edge detector leads to systematic error in width measurement, because the line spread function (LSP) of the imaging system shifts the positions of these maxima. To determine the correction for
this systematic error, we undertook a computational simulation of vessels with diameters 0.53mm, 0,79mm, 1.33mm, 1.97mm and 3.22 mm. The blurring effect of LSFs (of gaussian shape and with full widths at half maxima ranging from 0.6 mm to 1.6 mm) was assessed by convolving them with the simulated vessel profiles. The shifts in the positions of the first derivative extrema in the blurred vessel profiles were measured, and compared with measurements from Perspex phantoms (see Results). Geometric distotiion correction. A grid of wires forming squares of 10 mm side length was placed against the image intensifier input screen and was filmed. For a representative image of the grid, the correction coefficients required to change the image of each square (a geometrically distorted quadrilateral) to a true square were computed by automated image analysis, and were stored. The correction coefficients were subsequently applied to the vessel edge point co-ordinates in all images. This process was greatly facilitated by the fact that for both grid and vessel images, digitization was of virtually complete film frames (to 1024” pixels), and not of portions of frames (for example, appr~~ximately a quarter of a frame to 512’). The latter approach would have entailed correctly recording and applying offsets.
Validation resolving
procedures: power
accuracy,
precision
and
Accuracy was assessed using in z&o studies. A Perspex phantom was constructed, consisting of tubes of diameter 3.22 mm containing stenoses of diameter 1.97 mm, 1.33 mm, 0.79 mm and 0.53 mm. All dimensions of the tubes were accurately known, a travelling microscope being used to check their calibres. The true mean diameters of defined lengths of the tubes were computed. The tubes were filled with contrast medium and filmed immersed in a 25 cm dimension water enclosure. Measurements of MAWS for the defined lengths of tube were compared with the computed true mean diameters. The imprecision of measurement was determined from in Gvo data. Repeated analysis of arterial segments was done as follows, to assess variability. We deliberately digitized different frames {counter discrepancies of 1 and 2)) and repeated centreline definition and automatic detection in each. An analogous procedure was done with the catheters used in calibration. Uncertainties associated with the manufacturing tolerance of the catheters, and with changes in the degree of cardiac expansion between different end-diastolic phases (changing the catheter/segment differential magni~cation) were aIso q~lantified, since these are major contributors to the long-term variability. The overall resolving power (defined as the minimum measurable variat-ion in mean absolute width of segment, MAWS), incorporating all these sources of variability (see Results), was computed.
359
Clhical application iu the St. Thomas’ atherosclerosis regression study STARS was designed to test the effects of a lipidlowering diet or diet plus cholestyramine on the course of angiographically defined coronary artery disease in middle aged men with hypercholesterolaemia. Ninety patients were randomized to receive usual care, diet or diet plus chotestyramine with angiograms performed at baseline and after 3 years. A maximum of 10 pairs of coronary segments” were analysed using the techniques described above. Further details of the study are given elsewhere 4. Here we report on correlations between changes in the angiographic indices and on the correlations between these indices and the in-trial plasma concentrations of low~ensi~ lipoprotein (LDL) cholesterol. RESULTS Effect of line spread measurement
function
on width
The results of the computational simulation of the effect of line spread function on measured width are shown in Figure 4. The systematic error (change in distance between points of maximum slope) can be seen to vary from an overestimation of narrow arteries to an underestimation of broad ones. For comparison the four vertical lines in the figure correspond to the uncertainty values on experimental me~uremen~ from Perspex phantoms (see ‘Validation Procedures’). A correction function was derived from the phantom measurements and applied to all vessel width measurements. Validation resolving
of accuracy, power
precision
and overall
~~~~e 5 shows a composite image from an X-ray film of four Perspex phantom vessels. As explained in more detail in the caption, images from three adjacent film frames are included, as SYSTEMATIC
ERROR
IN WIDTti
Appararus Appnrarur Apparatus
function function function
MEASUREMENT widlh 0.6 m m width 0.8 m m width 1.0 m m
-. Vessel width (mm)
Figure4 The systematic error in width measurement (change in distance between points of maximum slope) can be seen to vary from an overestimation of narrow arteries to an underestimation of broad ones. For comparison the four verticai lines in the figure correspond to the uncertainty values on experimental measurement from Perspex phantoms (see ‘Validation Procedures’)
360
are graphs of widths from the individual frames. A further graph shows the average measurements from all three frames, and the associated reduction in noise. From the measurements of four tubes in three film frames, the accuracy of MAWS was calculated as the mean of unsigned differences between (LSF-corrected) MAWS and true mean diameters. The accuracy (and standard deviation, SD) was 0.03 (0.02) mm. By repeated analysis (see above) of arterial segments, the variability was found to be 0.07 (0.05) mm. The overall resolving power (defined above) is illustrated in Figure 6. An additive contribution to this was made by vessel measurement reproducibility and multiplicative contributions came from man~acturing tolerance, catheter catheter measurement reproducibility and cardiac expansion uncertainty. The catheter manufacturer’s information stated that the tolerance of the catheter tip diameter was *l%. Reproducibility of measuring the same catheter in different film frames was +I%. The contribution of cardiac expansion uncertainty (due to a variation in degree of expansion between cardiac cycles/examinations altering the position of the vessel, relative to the catheter, along the X-ray axis by an amount estimated to be 8 mm) was &l-3%. The description ‘multiplicative’ is applied to these contributions because percentage uncertainties in them scale up absolute uncertainties in vessel width measurements depending on them. Resolving power was computed as the quadrature combination of the individual contributions. Resolving power deteriorates by approximately a factor of 2 between 0.5 and 5 mm width, and the values are 0.10 and 0.14 mm at 2 mm and 4 mm respectively. A mean value of MAWS across a population of patients is approximately 3 mm (mean MAWS in STARS supports this), and an average resolving power (1 SD) for single segment AMAWS is 0.12 mm. We now consider resolving power per patient, estimated by dividing the resolving power for a single segment by the square root of the number of segments analysed in the patient. The worst resolving power that might be encountered would occur if only one, relatively wide (4 mm), segment could be measured. Resolving power would then be 0.14/1°.5, i.e. 0.14mm. The best resolving power would occur if all ten segments were towards the narrow end of the range of widths and were all measurable. Resolving power would then be 0.10/10°.5, i.e. 0.03 mm. We use the mean of the worst case(0.14 mm) and best case (0.03 mm) as the resolving power (1SD = 0.085 mm) for an average patient. This calculation is employed in preference to the following less cautious approach which yields apparently better resolving power. The average resolving power for single segment AMAWS would be divided by the mean number (approximately 6) of measurable segments per patient: resolving power per patient would then be 0.12/6’.“, i.e. 0.05 mm. Ideally, the variation of resolving power with width would always be stated, but for comparison with other work it appears helpful also to quote
(b)
/
IO 20
Vessel length / mm
Vessel length / mm
I
I
30
/
40
50’ 6r
Vessel length / mm
Vessel Iength / mm
Figure 5 (a) A composite image from an X-rav film of a Prrsprx phantonr. I he phantom consisted oi‘ four tubes each rcprrsrrrting a stenosis of different severity. In this Figure the images of the foul- tubes have been orient,tted so that the lube axes are horizontal atrd approximateh contiguous: The three horizorltallporicntated composite images were obtained tram three adjacent film frames. (h) The width measurements from these three individual frames arc’ shown in the left-most three graphs. The hori~otrral scale of thr graphs r’igbt-most ~mph shrnvr the ,ncragr measttrrments is compressed by a factor of approximately four cornpatwith that of the images. ‘hof noise reductictn hy averatfin~ is rrariilv apparettt from all three frames. The effectiveness RESOLVING
A, X
POWER
cardiac expansion uncdnainty catheter measurement reproducibility
progression and regression for patients and for individual segments was based on the resolving power, using a 2SD criterion; overall and segmental progression was thus defined as a change exceeding -0.17 mm and -0.24 mm, respectively, and regression as change exceeding +0.17 mm and +0.24 mm, respectively.
Angiographic findings in STARS: response to lipid lowering treatment and correlations with in-trial LDL cholesterol Vessel width (mm)
Figure 6 Overall resolving power of MAWS: the minimum mrasurable variation in mean arterial width. An additive contribution to this was made by vessel measurement reproducibility and multiplicative contributions c.amr from catheter manufacturing tolerance, catheter measurement nproducihility and cardiac expansion uncertainty
a single, ‘average’ figure for a segment of average width (approximately 3 mm). The ZSD figure for such a segment is 0.24 mm. For similar reasons we also provide a 2SD value for the resolving power per patient (0.17 mm). Change in MAWS was the principal end-point in the STARS trial, and an indication of overall change in coronary atherosclerosis for each patient was given as an overall MAWS change, averaged from all changes obtained from the measurable segments. The definition of disease
In STARS, 489 paired coronary artery segments were available for analysis, with a mean (SD) of 6.4 (2.0) segments per patient. The mean withinpatient changes in the four indices of luminal dimensions all demonstrated benefit with lipid lowering intervention. Per-segment results are detailed elsewhere”. I;igure 7 shows the associations between A%DS (the conventional angiographic measure of coronary disease) and AiMAwS (Fig 7a), AMi~WS (Fig 7b) and AEII (Fig 7c), the closest correlation being seen with AEII. In-trial plasma LDL cholesterol was significantly correlated with changes in all angiographic indices (Fig 8)) but closer correlations were found with AMAWS and AMinAWS (Figs Sa, 8b) than with AEII and A%DS EII (Figs 8c, 8d). Example images (with computer tracings) from a patient whose disease progressed with usual care are shown in Fig 9.
361
@ant&ion
of coronary
cineangiograms:
J.N.H.
Brunt
et al
-2.0~
60 AMhA%% = - 0.01 - ff.02dWS
1.5
pcO.&I?
1.0 0.5 0.0 4.5 -1 .o .I .5 .2.0 -SO
-40
40
-40
-20
0
20
0
A%wsmeter
20
40
64
20
40
60
sienosis
Association between change (A) in % diameter stenosis (%DS) and (a) A mean absolute width of segment (A~WS), (b) A minimum absolute width of segment (AMinAWS), and (c) A edge irregularity index (AEII)
DISCUSSION We have described the methodology required to measure, with optimal resolving power, angiographic changes in coronary artery disease. We utilised recent technological developments in storage and analysis. The image di~~zadon~ changes measured in STARS quantified both dif fuse and focal atherosclerosis. Coronary angiography has a number of limitations in this context. It is invasive, The angiogram is a sequence of Z-dimensional ‘shadowgrams’ of opacified vessels and as such does not visualise the arterial wall. It may underestimate the extent of disease (especially early atherosclerotic changes). Several processes may contribute to the changes seen: atherosclerosis, arterial spasm, thrombosis, post-stenotic dilatation. The visualization does not distinguish between these. Never-
362
theless its use in several trials has proved highly informative and is congruent with the incidence of clinical coronary events. The information contained in the automatically-detected, stored vessel boundaries was expressed in various ways* The primary angiographic index of atherosclerosis we chose was an absolute measurement, mean absolute width of segment (MAWS). This has the important advantage over the commonly used relative measurement percent diameter stenosis that it does not require reference section(s) of vessel of assumed normality (see methods section). For completeness and to assist comparison with other studies, we did, however, compute percent diameter stenosis (%DS), also, as well as two other indices, minimum absolute width of segment (Mi~WS) and edge irregularity index (EII), views has tbe advantage of reflecting the degree of focal stenosis, which has a pronounced effect on flow resistance due to the strong Poiseuille dependence of that property on vascular diameter. Edge irregularity index is a convenient indicator of the degree of diffuse atherosclerosis, and allows early atherosclerosis to contribute to the results. STARS was confined to these indices, but since we retained the vessel boundary data for the entire study, further indices (e.g. atheroma volume) could be computed. Not unexpectedly, in view of the common underling data, there was strong correlation between the four indices reported. We found, however, that the absolute measures were better correlated with LDL cholesterol than was %DS. Densitometric techniques (e.g. Spears et at”) were considered to be still an insufficiently reliable approach to obtaining relative or absolute arterial cross-sectional area measurements from film. This view is supported by conclusions drawn by Reiber et UP’ from an extensive review of developments carried out by various investigators. Three dimensional calibration’*.13 has been successfully used in studies with small numbers of patients14*15, but we found it too inconvenient fur a study involving scores of patients. In STARS we reported angiographic changes on both a per-patient and per-segment basis. Whilst the former has greater clinical appeal, it overlooks the fact that within a given patient, segments may behave differently over time’“. Accordingly we found a very low intra-class correlation for changes in an~og~phic indices within patients. Per-segment results also increase the power of the statistical analysis and this may explain why significant differences between the diet and diet plus cholestyramine groups were only found with these data4. Examining changes in orographic indices as continuous variables instead of selecting arbitrary ‘cut offs’ (that may reduce the power of the analysis) is preferred. There are numerous potential sources of error in QCA**“; we minimized these as follows. Tight criteria were imposed in selecting segments for analysis. Frames were consistently selected at the end-diastolic cardiac phase, where the frame-toframe changes occur at their minimum rate. Measurements from three adjacent frames at this
1
2
3
4
LDL Cholesterol
Fire
8
Association
between
in-trial
LDL-choiesferoI
level and
(a) dMAW’S,
phase were averaged. An automated edge detection algorithm of considerable sophistication and robustness was employed. This involved the firstbut not the second-derivat.ive of the image data, with consequent advantages in noise-avoidance and simplicity of correcting systematic errors arising from blurring by the imaging system. Selection of the single frame showing the instantaneous most severe visualization of stenosis would result in stenosis values of apparent maximum severity, but these values would be prone to be overestimates because of the noise level in individual images. We have adopted the more systematic approach of using stenosis values from S-frame averaging. Great care was exercised in ensuring the actual catheter dimensions were associated with that part of- the catheter tip analysed. Pincushion distortion was also corrected for. Practical advantages arise from separating the analysis phases of imagedigitization and of imageprocessing by digitally storing images from an entire intervention trial: frames can be more criticalIy matched when an analyst recalls them onto the monitor; averaging measurements from adjacent film frames becomes simple; future access to the film could be uI~available~ i~lformation in the film could be lost if the film were damaged: digital information can be copied losslessly. The image archive accumulated during the STARS trial is currently one of the largest collections of high-resolution digitized ( 1024 x 1024 pixel) coronary cin~~ngiographic images in the world. It is intended to add images from other studies in due course. As this database expands
(h) -1MinAWS.
(c) XII,
5
6
7
8
(mmoi/l)
and (d) J%inS
it should allow convenient investigation of future hypotheses related to the variability and classification of coronary morphology and t,o the development, location, extent and severity of atherosclerotic disease. It would be advantageous if internationally accepted standards were to be established for data processing in studies of this kind, as this would facilitate comparison of results. So called regression trials are a relatively new approach for assessing the effect of lipid lowering therapy on the course of coronary atherosclerosis. Various randomized trials (FATS”, SCOR’*, Lifestyle’“, Leiden”“, NHLBI”‘, CIM”2, POSCH=) have been congruent in demonstrating that lipidlowering treatment beneficially affects change in luminal dimensions. Associated change in angina1 symptoms has been recorded in Lifestyle and STARS, and that changes in sequential angiograms may be employed to predict risks of subsequent mortality due to coronary heart disease has also been suggested by data from the POSCH study. Despite substantial progress in other imaging modalities, notably computed tomography, ultrasound and nuclear magnetic resonance, and despite the many limitations referred to above, remains the definitive procedure for c&x assessing changes in atherosclerosis in clinical trials.
We are grateful for support Bristol-Myers Laboratories.
from Unilever Ltd and This work has ben-
363
Qwntitation
of comnaly
cinamgiogram:
J.N.H.
Brunt
et al.
Figure ! 3 (a) An example of a patient whose disease progressed with usual care. These are smal1 portions of frames. Each portion of 256 x 256 pixels was taken from the 1024 x 1024 pixel stored digitized version of the frame. Detected vessel boundaries are shown as continuous li ines. The d otted lines further away from the vessel on each side indicate the ends of the scan rectangles for sampling image grey values {see also Figure 21. The arrow heads indicate a part of the vesset where the change in disease (narrowing of the vessel) is parti~nlarly large. (b) The u idth functions derived from the vessel boundaries
efited also from a project concerned with angiographic blood flow measurement, supported by the British Heart Foundation. Thanks also to the conscientious referees for constructive suggestions. 6.
REFERJINCES 1. Leung WH, quantitation
2. 3.
4.
5.
364
Sanders
WJ, Alderman
EL. A coronary
artery
and data management system for paired Catheterization and Cardiouascular Diagcineangiograms. nosis 1991; 24: 121-34. Reiber JHC, Serruys PW (eds). 7%~ Duelists in Quantitative Coronary A~o~u~~y’ Kluwer Academic, Dordrecht, 1988. Reiber JHC. An overview of coronary quanti~tion techniques as of 1989. In ‘~~ntitati~e Choral A~o~u~hy ’ Reiber JHC, Serruys PW (edsf , Kluwer, Dordrecht, 199 1, 55-132. Watts GF, Lewis B, Brunt JNH, Lewis ES, Coltart DJ, Smith LDR, Mann JI, Swan AV. Effects on coronary artery disease of lipid-lowering diet, or diet plus cholestyramine, in the St. Thomas’ Atherosclerosis Regression Study (STARS) Lancet 1992; 339: 563-9. Austen WG, Edwards JE, Frye RI et al. AHA committee report. A reporting system on patients evaluated for cor-
7. 8.
9. 10.
onary artery disease Circulation 19’75, 51: 5-40 (Note: these pages are numbered separately from the conventional numbering scheme of the journal, and are published at the back of the April issue. Not all libraries have retained them because of their proximity to the advertising section), Montanari, LJ. On the optimal detection of curves in noisy pictures Comm. of the ACM 1971; 14: 335-45. Azzopardi PJ, Pycock D, Taylor CJ, Wareham AC. An experiment in model-based boundary detection Pvoc. 4th A&y Vision Conference (AVC88) 1988, 31-6. Selzer RH, Shircore A, Lee PL, Hemphill L and Blankenhorn DH. A second look at quantitative coronary angiogof the art in raphy: some unexpected problems. In ‘State ~ua~titut~~e Canal A~~o~a~h~’ Reiber JHC, Serruys PW feds) Martinus N~hoff, Dordrecht, 1986, 125-43. Doi K and Rossmann K. Effect of focal spot distribution on blood vessel imaging in ma~i~catio~ radiography. Radiology 1975; 114: 435-41. Spears JR, Sandor T, Als AV et al. Computerized image analysis
for
quantitative
measurement
of vessel
diameter
from cineangiograms. Circulation 1983; 68: 453-61. 11. Keiber JHC, van der Zwet PMJ, von Land CD, Koning G, van Meurs B, Buis B, van Voorthuisen AdE: Quantitative coronary arteriography: equipment and technical requirements. In Advances in Quantitative Coronary Arteri-
ogru@y Reiber JHC, Serruys PW (eds): Kluwer, Dordrecht, 1992; 75-l 11. 12. Mackay SA, Potel MJ, Rubin JM. Graphics methods for tracking three-dimensional heart wall motion. Computers and Biomedical Research 1982; 15: 455-73. 13. Colchester ACF, Hawkes DJ, Brunt JNH, du Boulay GH, Wallis A. Pulsatile blood flow measurements with the aid of .3-D reconstruction from dynamic angiographic recordings. In Infmmation Processing in Medical Imaging-9. Bacharach SI, (ed.) Martinus Nijhoff, Dordrecht, 1986: 247-65. 14. Brunt.JNH, du Boulay GH, Wallis A, Wicks DAG. Fluid flow measurement in vessels from digital subtraction angiography video recordings. Neuroradiolog? 1991; 33
(Suppl):
456-i-8.
15. Brunt JNH, du Boulay GH, Wallis A. The measurement of blood flow waveforms from X-ray angiography. Part 2: Application to video recordings of digital subtraction angiography: Proc I&n Mech Engrs: Journal of Eqineting in Medicine 1992; 206, Part H: 87-91. 16. de Feyter PJ, Serruys PW, Davies MJ, Richardson P. Lubsen J, Oliver MF. Quantitative coronary angiography to measure progression and regression of coronaq atherosclerosis. Circulation 1991; 84: 412-23. 17. Brown G, Albersg, Fisher LD et nl. Regression of coronary artery disease as a result of intensive lipid-lowering therapy in men with high levels of apolipoprotein B. N I:‘ngl, J Med 1990; 323: 1289-98.
18. Kane JP, Malloy MJ, Ports TA et al. Regression of coronary atherosclerosis during treatment of familial hypercholesterolemia with combined drug regimens.,JAMA 1990: 264: 3007-l”. 19. Ornish D, Brown SE, Scherwitz LW. d ~1. Can lifestyle changes reverse coronary heart disease? Lnnc~t 1990; 336: 129-33. 20. Arntzenius AC, Kromhout D, Barth JD el al. Diet, lipoproteins, and the progression of coronary atherosclerosis: NPW Engl J Med 1985; 312: 805-l 1. 21. Brensike JF, Levy RI, Kelsey SF et nl. Effects of therapy with cholestyramine on progression of coronary arteriosclerosis: results of the NHLBI type II coronary intervention study. Circulation 1984; 69: 313-24. 22. Blankenhorn DH. Nessim SA. Johnson RI,, Sanmarco ME, Azen SP, Cashin-Hemphill L. Beneficial effects of combined colestipol-niacin therapy on coronary atherosclerosis and coronary venous bypass grafts. JAMA 1987; 257: 3233-40. 23. Buchwald H, Marco RI,, Matts JP, PI al. Effect of partial ileal bypass surgery on mortality and morbidity from coronary heart disease in patients with hypercholesterolemia-Report of the Program on the Surgical Control of the Hyperlipidemias (POSCH). N Engl ,J Mpd 1990; 323: 946-55.
365
Med. Eng. ERWORTH EINEMANN
Qu~~~~ve coronary c~e~~o~aphy study of atherosclerosis J.N.H. Brunt*, D.J. Coltart’
G.F. Watts+, B. Lewis+, L.D.R.
Smith’
for the
and
“Dept of Medical Biophysics, Stopford Building, Manchester Universi~, Manchester Ml3 9PT, UK; +Depts of Endocrinology & Chemical Pathology %ardiology, UMDS, St Thomas’ Hospital, London SE1 7EH, UK Received
November
1993, accepted
February
and
1994
ABSTRACT An~o~ap~y is the ~~n~ti~e ~ocedurefor c~ar~ct~ing the extent and course ojcoron~~ artery diseme. We describe the ~thodo~~ retired to measure, with sternal r%~o~v~ngpow~, ~n~~uph~c chenps in canal arty disease. We utilised recent te~hn~~~‘~~~ d~e~p~ts in image di~.t~~tion, storage and analysis. The measures of change quant@ed both diffuse and focal atherosclerosis. Frames from angiographic tine films were digitized at high resolution (1024 x 1024 pixels, 8 bit grq scale) and archived on optical disk. Four radiographic projections were stored to ensure good visualization qfas many as possible of a set qf tenmajor arterial segments. Edges of segments and catheter were automatically delineated by computer using a dynamic programming algorithm involving a costfunction which contained terms based on edge strength and on continuity. For every digitized radiographic projection, delineation, was repeated in three adjacent frames, to improve precision. Edge points for each corona9 segment were stored on disk. From these we computed the mean width along the sent (pixelsj. Scaling to obtain the Mean Absolute Width of the SeFt ~.~~, mm) was achieved using catheter dirn~s~o~s known porn rni~o~~, systematic error due to imaging system line-spread function being corrected using data from computer simulations and phantom studies. Correction for geometric image intensifier distortion was also applied. We used the methodology in a randomized, controlled trial of the effect of lipid-l owering therapy, the St Thomas ’ Atherosclerosis Regression Study. The jundamental measure of change of disease in each segment was the change in MAWS (AMAWS}. Using in-vitro and in-vivo studies we established that the overall resolving power for one segment AMAWS was 0. IO mm at 2 mm width and 0.14 mm at 4 mm width. Subsidia? end+oints were the change (A) in minimum absolute width of segment (~~~nA~~, edge ~~e~~a~ty index (EZZ) and #rcent d~~~t~ stenosis (%DS). A %DS {the &onv~t~o~~L angiographic measure ojc~ona~ disease) was si~~~~~ntly correlated with change in all indices, closest chelation being seen with AEZI (r = 0.94, p
diagnostic
image pro-
1995, Vol. 17, 356-365, July
INTRODUCTION In assessing the natural history of coronary atherosclerosis, and in trials to determine treatment effects, the preferred imaging technique is serial coronary angiography. X-ray projection images of the con~ast~ont~ning coronary arteries are obtained from various directions to ensure good visualization of all the major vessels and of any lesions that may be present. The X-ray image transduction chain incorporates an image Correspondence to: DrJ.N.H. Brunt, Department of Medical Physics, CIat&erbridge Centre for Oncology, Bebington, Wirral, Merseyside, L63 4JY, UK. rPresent address: University Department of Medicine, University of Western Austrafia, Box X2213 GPO Perth, WA 6001, Australia.
intensifier to permit images to be recorded at a rate of 25 orA more per-second while avoiding excessive radiation dose to the patient. The standard recording medium is 35 mm tine film’, though digital recording may replace this in the future. Visual interpretation of coronary angiograms is informative but has limitations: (1) in&a- and inter-observer variability, (2) only relative calibres (e.g. percent stenosis) can be estimated, (3) it cannot form the basis from which to calculate flow resistance and hence assess the physiological significance of the disease. These limitations stimulated the development of computerized techniques for quantitative coronary angiography (QCA). Quantitation is now used at many centres
world~vide, and the similarities and differences among the leading centres have been documented in the reviews by Reibe?. Quantitative analysis most commonly involves measuring changes in arterial segments over a period of time by. comparing results obtained from two films. In this paper we describe our methodology for QCA. It has been used in the St Thomas’ Atherosclerosis Regression study (STARS)” to measure, with optimal resolving power, angiographic changes in coronary artery disease. We utilized recent technological developments in image digitizati(~n, storage and analysis. The changes measured in STARS quantified both diffuse and focal atherosclerosis. METHODS Analysis apparatus
Corona? images on 35 mm film were selected for ~igitizatlon usmg a Tagarno dual-channel projectlon system and subsequently projected into an electronic camera with digital output suitable for processing by computer. Figure 1 shows block cliagrams of the apparatus used. I;iEm ~~~~~(.~~~~~~a~~~ ~~t~~~,~~~. hatching frames from film-pairs were selected using the dual-channel projector, with facilities for taking Polaroid
a) Cine angiographic
b) Frame
selection,
Imaging
using
dual-channel
projection
projection screen
processed films
C) Image digiti~tion
pictures digitizing
of the selected frames.
frames,
to be used when
Film projector for dig-itizaticrn. Because the illumination intensity produced by a normal angiographic projector would be too high to project directly onto the solid state sensor of the digitizing camera, the condenser optics were designed to provide uniform but low intensity illumination. Compared with standard optics, there were three changes. First, the usual lamp of several hundred watts power was replaced with a 10 W quartz halogen lamp. Secondly, no condenser mirror was used. Thirdly, the condenser lens was moved approximately 0.1 m away from the light source: this reduced, by an order of magnitude, the solid angle it subtended while improvitlg the uniformit~7 ~~f~ill~~rni~ation. Thus overall ill~irnin~~ti~~n intensitv at the film was approximatetv 0.5% of normal. If$@&ing ~‘(zr~era. A Videk Megaplus camera (Lastman Kodak Company, Rochester, USA) with a charge coupled device solid state sensor array of 1320 x 1035 pixel resolution and with a-bit digital output, was used. The camera was digitally linked to the analysis computer. Optics were chosen so that each 1 mm square of tine iihn corresponded to a 55 x 55 pixel portion of digitized image. In most tine frames analysed, data storage was limited to the central 1024 x 1024 pixels of the sensor array (i.e. 1 megabyte, MB, per image). This corresponded to a square of film of 18.6 mm side length, and therefore encompassed virtually all the exposed emulsion of the complete (&neformat) frame. The heart was imaged onto the film with a magnification of approximately 0.17; hence each pixel corresponded to a square of side typically 0.11 mm at the heart. ~~~)r~~~Ltirlg environ merit md storage of image c&n. All digitized images, typically twenty to thirty per patient, were stored at fLll1 resolution using both write-once-read-multiple (Optimem) and rewritable (Sony) optical disks, the former providing archiving and the latter being used for day-to-day access. These devices were networked to a Sun workstation (running Pascal programs), with an image manipulation co-processor which could hold multiple images or portions of images. This was convenient for region selection and matching in adjacent film frames and between serial films. When comparing matched angiographic projections from two films, 6 MB of memory stored digitized versions of three adja.cent frames from each film. Analysis protocol
Image analysis computer
Figure s&rtion
1
Block diagrams of (a) tine-film acquisition, and (t‘) image (ii~~ti~~ti~n and anatysis systems
magnetic and optical disks
(b) frame
~a~~~~&~~~~~~~ ~ab~~~~ ~~o~oGo~~s.To rninimise variation between first and second an~ograms a standard approach was employed: Judkins percutaneous femoral artery technique; ‘Cordis’ 8F polyurethane catheters; “Urografin’ contrast medium; no vasodilators. Four views of the right and 7 of the left coronary artery were recorded on ‘Agfa Scopix RPl’ 35 mm tine ftlm, along with
357
Quantitation
of coronary
[email protected]:
J.N.H.
Brunt
et al.
images of a grid (1 cm spacings) attached to the image intensifier to allow correction of pincushion distortion. Radiographic angulations, table height, and the sequence of injections (which predominantly influenced the elapsed time from first injection) were recorded and reproduced as exactly as possible, by the same cardiologist, at repeat angiography. segment selection. We investigated ten major arterial segments: the proximal, middle and distal segments of the right coronary artery, the left main stem, the proximal and distal circumflex artery, the proximal, middle and distal segments of the left anterior descending artery, and the obtuse marginal artery, using the anatomical definitions of the American Heart Association5. Paired, coded films were compared on the dualchannel projector and paired frames (i.e. images matched between serial studies) of the same radiographic views (typically four in each film) were photographed. The frames selected for analysis were taken at enddiastole (as assessed from visual comparison of frames before and after the selected frame). They had to clearly demonstrate particular segments in the same frame as the catheter tip. There had to be no nearly parallel overlying vessels, and the vessels had to be uniformly filled with contrast medium. Polaroid photographs were coded, and marked to identify the segments for image analysis. Coding was done by an investigator who was aware of the film sequence but blinded to the therapeutic group. Films were then transferred to the location of the digitizing projector/camera. Digitization of the coded frames and analysis of the resulting digital images (guided by photographs) were done by an investigator blinded to the sequence and therapeutic group of the angiograms and photographs. Arterial
Segment boundaly delineation and width function. The anatomically defined length of each
analysable segment was input to the analysis computer by pointing to a few points along its centreline. The completely automatic boundary delineation process (Figure 2) then took place along the whole length, as follows. A set of scan rectangles (of length typically 60 pixels and of several pixels Scan rectangles
for sampling
width) perpendicular to the vessel centre line sampled profiles of grey values in the image. The scan rectangles were spaced at intervals of four pixels - approximately 0.45 mm - along the centreline. To find each side of the vessel a computer algorithm searched for the eight most significant correct-direction (grey value rising or falling) edges (i.e. peaks or troughs in the first derivative function) along each entire profile. The edge points from the set of profiles were used as candidate boundary points in a dynamic programming algorithm6,‘. This found the lowest cost path using a cost function which contained terms based on edge strength and on continuity. A width function was derived from the distances of closest approach of the two boundaries and a three-point smoothing function was applied to this function. In measuring widths, we followed a suggestion of Selzer et al* to improve precision by averaging the results derived from three sequential tine frames (the frames themselves were not averaged). Since there were, deliberately, no facilities for manual ‘correction’ of boundaries, if boundary detection was obviously artefactual, the corresponding width function was omitted from the average. Measurements obtained from vessel boundaries: angiographic indices. The measurements obtained
from the width function are described below. All were ‘derived after correcting for pincushion distortion and (for absolute measures) applying a calibration factor. Comparison of change in angiographic indices was carried out by linear regression analysis. (a) Mean absolute width of segment (MAWS). Width measurements expressed in pixels were obtained along an entire arterial segment as described (see Figure 3) and the mean value of width along the segment was computed. The calibration factor (9.00 pixels per mm was a typical value) was used to convert the pixel values to absolute units (mm). Differential magnification between the catheter and the segment (due to X-ray beam divergence) was not corrected for, but the variation of its possible magnitude between similarly angulated views in successive examinations of the same patient was estimated in quoting the overall resolving power of the system.
image grey values width
interpolated at minimum
width
I
position of min. width distance -eeee-
Vessel centreline
-
Vessel
boundaries
to be delineated
Figure 2 Diagrammatic representation ing boundaries of coronary vessels
358
of methodology
for delineat-
along vessel
figure 3 Schematic diagram showing the ‘width function arterial segment. It was from the set of width measurements tuting this function that the changes in the angiographic MAWS, MinAWS, %DS and EII were computed
for an constiindices
(b) Minimum absolute width of sepment (MinAWS). The minimum width was c;mputed as the minimum value among the width values spaced along the segment (i.e. re-using the data required to calculate MAWS). (cl Percentage diameter stenosis (%DS). Percentage diameter stenosis was derived from the ratio of the minimum width along a segment to a computed width representing an approximation to the width that the vessel would have had at that point in the absence of disease. It was obtained by iteratively fitting (by least squares) a straight line to the plot of width values against distance along the vessel. When performing the fit, the weights applied to values corresponding to the part of the segment containing the minimum width were a function of the degree of stenosis computed at each iteration and were progressively reduced during the three iterations, so that the line was ultimately fitted essentially to the widths of the less diseased parts of the segment only. (d) Edge irregularity index (EII). The edge irregularity index used was the ratio obtained by dividing the standard deviation of the (abovementioned) least squares fit by the value of MAWS for the segment. Accounting unvisualised
for data ji-om different views segments. When a segment
and j&r
was visualized and calibrated in more than one view, its MinAWS (and MAWS) values were computed by averaging across the views. The %DS value used was that from the view producing the most severe value. Segments Lln~~sualized because of total occlusion proximal to the measurement site could not undergo edge-detection, but overall assessment would have been seriously incomplete without accounting for them. Therefore we defined a ‘functional width’ for any such segment: if blood did not reach a more distal segment because of an occlusion in a more proximal segment, the functional width of the more distal segment was zero, even though we did not know whether this unvisualized segment was itself occluded. Segments perfused and visualized as a result of coilateral flow were measured in the usual way. Calibration
and correction
of systematic
errors
~a~i~atio~ te&hn~que. The process of calibration converts measurements of vessel widths expressed in pixels to measurements expressed in absolute units (mm). We detected the edges of the catheter using the same automated delineation algorithm used for the vessels, and achieved calibration by comparison with a table of catheter tip diameters, obtained by micrometry. Line stead function. The use of detection of extrema in the first derivative within the edge detector leads to systematic error in width measurement, because the line spread function (LSP) of the imaging system shifts the positions of these maxima. To determine the correction for
this systematic error, we undertook a computational simulation of vessels with diameters 0.53mm, 0,79mm, 1.33mm, 1.97mm and 3.22 mm. The blurring effect of LSFs (of gaussian shape and with full widths at half maxima ranging from 0.6 mm to 1.6 mm) was assessed by convolving them with the simulated vessel profiles. The shifts in the positions of the first derivative extrema in the blurred vessel profiles were measured, and compared with measurements from Perspex phantoms (see Results). Geometric distotiion correction. A grid of wires forming squares of 10 mm side length was placed against the image intensifier input screen and was filmed. For a representative image of the grid, the correction coefficients required to change the image of each square (a geometrically distorted quadrilateral) to a true square were computed by automated image analysis, and were stored. The correction coefficients were subsequently applied to the vessel edge point co-ordinates in all images. This process was greatly facilitated by the fact that for both grid and vessel images, digitization was of virtually complete film frames (to 1024” pixels), and not of portions of frames (for example, appr~~ximately a quarter of a frame to 512’). The latter approach would have entailed correctly recording and applying offsets.
Validation resolving
procedures: power
accuracy,
precision
and
Accuracy was assessed using in z&o studies. A Perspex phantom was constructed, consisting of tubes of diameter 3.22 mm containing stenoses of diameter 1.97 mm, 1.33 mm, 0.79 mm and 0.53 mm. All dimensions of the tubes were accurately known, a travelling microscope being used to check their calibres. The true mean diameters of defined lengths of the tubes were computed. The tubes were filled with contrast medium and filmed immersed in a 25 cm dimension water enclosure. Measurements of MAWS for the defined lengths of tube were compared with the computed true mean diameters. The imprecision of measurement was determined from in Gvo data. Repeated analysis of arterial segments was done as follows, to assess variability. We deliberately digitized different frames {counter discrepancies of 1 and 2)) and repeated centreline definition and automatic detection in each. An analogous procedure was done with the catheters used in calibration. Uncertainties associated with the manufacturing tolerance of the catheters, and with changes in the degree of cardiac expansion between different end-diastolic phases (changing the catheter/segment differential magni~cation) were aIso q~lantified, since these are major contributors to the long-term variability. The overall resolving power (defined as the minimum measurable variat-ion in mean absolute width of segment, MAWS), incorporating all these sources of variability (see Results), was computed.
359
Clhical application iu the St. Thomas’ atherosclerosis regression study STARS was designed to test the effects of a lipidlowering diet or diet plus cholestyramine on the course of angiographically defined coronary artery disease in middle aged men with hypercholesterolaemia. Ninety patients were randomized to receive usual care, diet or diet plus chotestyramine with angiograms performed at baseline and after 3 years. A maximum of 10 pairs of coronary segments” were analysed using the techniques described above. Further details of the study are given elsewhere 4. Here we report on correlations between changes in the angiographic indices and on the correlations between these indices and the in-trial plasma concentrations of low~ensi~ lipoprotein (LDL) cholesterol. RESULTS Effect of line spread measurement
function
on width
The results of the computational simulation of the effect of line spread function on measured width are shown in Figure 4. The systematic error (change in distance between points of maximum slope) can be seen to vary from an overestimation of narrow arteries to an underestimation of broad ones. For comparison the four vertical lines in the figure correspond to the uncertainty values on experimental me~uremen~ from Perspex phantoms (see ‘Validation Procedures’). A correction function was derived from the phantom measurements and applied to all vessel width measurements. Validation resolving
of accuracy, power
precision
and overall
~~~~e 5 shows a composite image from an X-ray film of four Perspex phantom vessels. As explained in more detail in the caption, images from three adjacent film frames are included, as SYSTEMATIC
ERROR
IN WIDTti
Appararus Appnrarur Apparatus
function function function
MEASUREMENT widlh 0.6 m m width 0.8 m m width 1.0 m m
-. Vessel width (mm)
Figure4 The systematic error in width measurement (change in distance between points of maximum slope) can be seen to vary from an overestimation of narrow arteries to an underestimation of broad ones. For comparison the four verticai lines in the figure correspond to the uncertainty values on experimental measurement from Perspex phantoms (see ‘Validation Procedures’)
360
are graphs of widths from the individual frames. A further graph shows the average measurements from all three frames, and the associated reduction in noise. From the measurements of four tubes in three film frames, the accuracy of MAWS was calculated as the mean of unsigned differences between (LSF-corrected) MAWS and true mean diameters. The accuracy (and standard deviation, SD) was 0.03 (0.02) mm. By repeated analysis (see above) of arterial segments, the variability was found to be 0.07 (0.05) mm. The overall resolving power (defined above) is illustrated in Figure 6. An additive contribution to this was made by vessel measurement reproducibility and multiplicative contributions came from man~acturing tolerance, catheter catheter measurement reproducibility and cardiac expansion uncertainty. The catheter manufacturer’s information stated that the tolerance of the catheter tip diameter was *l%. Reproducibility of measuring the same catheter in different film frames was +I%. The contribution of cardiac expansion uncertainty (due to a variation in degree of expansion between cardiac cycles/examinations altering the position of the vessel, relative to the catheter, along the X-ray axis by an amount estimated to be 8 mm) was &l-3%. The description ‘multiplicative’ is applied to these contributions because percentage uncertainties in them scale up absolute uncertainties in vessel width measurements depending on them. Resolving power was computed as the quadrature combination of the individual contributions. Resolving power deteriorates by approximately a factor of 2 between 0.5 and 5 mm width, and the values are 0.10 and 0.14 mm at 2 mm and 4 mm respectively. A mean value of MAWS across a population of patients is approximately 3 mm (mean MAWS in STARS supports this), and an average resolving power (1 SD) for single segment AMAWS is 0.12 mm. We now consider resolving power per patient, estimated by dividing the resolving power for a single segment by the square root of the number of segments analysed in the patient. The worst resolving power that might be encountered would occur if only one, relatively wide (4 mm), segment could be measured. Resolving power would then be 0.14/1°.5, i.e. 0.14mm. The best resolving power would occur if all ten segments were towards the narrow end of the range of widths and were all measurable. Resolving power would then be 0.10/10°.5, i.e. 0.03 mm. We use the mean of the worst case(0.14 mm) and best case (0.03 mm) as the resolving power (1SD = 0.085 mm) for an average patient. This calculation is employed in preference to the following less cautious approach which yields apparently better resolving power. The average resolving power for single segment AMAWS would be divided by the mean number (approximately 6) of measurable segments per patient: resolving power per patient would then be 0.12/6’.“, i.e. 0.05 mm. Ideally, the variation of resolving power with width would always be stated, but for comparison with other work it appears helpful also to quote
(b)
/
IO 20
Vessel length / mm
Vessel length / mm
I
I
30
/
40
50’ 6r
Vessel length / mm
Vessel Iength / mm
Figure 5 (a) A composite image from an X-rav film of a Prrsprx phantonr. I he phantom consisted oi‘ four tubes each rcprrsrrrting a stenosis of different severity. In this Figure the images of the foul- tubes have been orient,tted so that the lube axes are horizontal atrd approximateh contiguous: The three horizorltallporicntated composite images were obtained tram three adjacent film frames. (h) The width measurements from these three individual frames arc’ shown in the left-most three graphs. The hori~otrral scale of thr graphs r’igbt-most ~mph shrnvr the ,ncragr measttrrments is compressed by a factor of approximately four cornpatwith that of the images. ‘hof noise reductictn hy averatfin~ is rrariilv apparettt from all three frames. The effectiveness RESOLVING
A, X
POWER
cardiac expansion uncdnainty catheter measurement reproducibility
progression and regression for patients and for individual segments was based on the resolving power, using a 2SD criterion; overall and segmental progression was thus defined as a change exceeding -0.17 mm and -0.24 mm, respectively, and regression as change exceeding +0.17 mm and +0.24 mm, respectively.
Angiographic findings in STARS: response to lipid lowering treatment and correlations with in-trial LDL cholesterol Vessel width (mm)
Figure 6 Overall resolving power of MAWS: the minimum mrasurable variation in mean arterial width. An additive contribution to this was made by vessel measurement reproducibility and multiplicative contributions c.amr from catheter manufacturing tolerance, catheter measurement nproducihility and cardiac expansion uncertainty
a single, ‘average’ figure for a segment of average width (approximately 3 mm). The ZSD figure for such a segment is 0.24 mm. For similar reasons we also provide a 2SD value for the resolving power per patient (0.17 mm). Change in MAWS was the principal end-point in the STARS trial, and an indication of overall change in coronary atherosclerosis for each patient was given as an overall MAWS change, averaged from all changes obtained from the measurable segments. The definition of disease
In STARS, 489 paired coronary artery segments were available for analysis, with a mean (SD) of 6.4 (2.0) segments per patient. The mean withinpatient changes in the four indices of luminal dimensions all demonstrated benefit with lipid lowering intervention. Per-segment results are detailed elsewhere”. I;igure 7 shows the associations between A%DS (the conventional angiographic measure of coronary disease) and AiMAwS (Fig 7a), AMi~WS (Fig 7b) and AEII (Fig 7c), the closest correlation being seen with AEII. In-trial plasma LDL cholesterol was significantly correlated with changes in all angiographic indices (Fig 8)) but closer correlations were found with AMAWS and AMinAWS (Figs Sa, 8b) than with AEII and A%DS EII (Figs 8c, 8d). Example images (with computer tracings) from a patient whose disease progressed with usual care are shown in Fig 9.
361
@ant&ion
of coronary
cineangiograms:
J.N.H.
Brunt
et al
-2.0~
60 AMhA%% = - 0.01 - ff.02dWS
1.5
pcO.&I?
1.0 0.5 0.0 4.5 -1 .o .I .5 .2.0 -SO
-40
40
-40
-20
0
20
0
A%wsmeter
20
40
64
20
40
60
sienosis
Association between change (A) in % diameter stenosis (%DS) and (a) A mean absolute width of segment (A~WS), (b) A minimum absolute width of segment (AMinAWS), and (c) A edge irregularity index (AEII)
DISCUSSION We have described the methodology required to measure, with optimal resolving power, angiographic changes in coronary artery disease. We utilised recent technological developments in storage and analysis. The image di~~zadon~ changes measured in STARS quantified both dif fuse and focal atherosclerosis. Coronary angiography has a number of limitations in this context. It is invasive, The angiogram is a sequence of Z-dimensional ‘shadowgrams’ of opacified vessels and as such does not visualise the arterial wall. It may underestimate the extent of disease (especially early atherosclerotic changes). Several processes may contribute to the changes seen: atherosclerosis, arterial spasm, thrombosis, post-stenotic dilatation. The visualization does not distinguish between these. Never-
362
theless its use in several trials has proved highly informative and is congruent with the incidence of clinical coronary events. The information contained in the automatically-detected, stored vessel boundaries was expressed in various ways* The primary angiographic index of atherosclerosis we chose was an absolute measurement, mean absolute width of segment (MAWS). This has the important advantage over the commonly used relative measurement percent diameter stenosis that it does not require reference section(s) of vessel of assumed normality (see methods section). For completeness and to assist comparison with other studies, we did, however, compute percent diameter stenosis (%DS), also, as well as two other indices, minimum absolute width of segment (Mi~WS) and edge irregularity index (EII), views has tbe advantage of reflecting the degree of focal stenosis, which has a pronounced effect on flow resistance due to the strong Poiseuille dependence of that property on vascular diameter. Edge irregularity index is a convenient indicator of the degree of diffuse atherosclerosis, and allows early atherosclerosis to contribute to the results. STARS was confined to these indices, but since we retained the vessel boundary data for the entire study, further indices (e.g. atheroma volume) could be computed. Not unexpectedly, in view of the common underling data, there was strong correlation between the four indices reported. We found, however, that the absolute measures were better correlated with LDL cholesterol than was %DS. Densitometric techniques (e.g. Spears et at”) were considered to be still an insufficiently reliable approach to obtaining relative or absolute arterial cross-sectional area measurements from film. This view is supported by conclusions drawn by Reiber et UP’ from an extensive review of developments carried out by various investigators. Three dimensional calibration’*.13 has been successfully used in studies with small numbers of patients14*15, but we found it too inconvenient fur a study involving scores of patients. In STARS we reported angiographic changes on both a per-patient and per-segment basis. Whilst the former has greater clinical appeal, it overlooks the fact that within a given patient, segments may behave differently over time’“. Accordingly we found a very low intra-class correlation for changes in an~og~phic indices within patients. Per-segment results also increase the power of the statistical analysis and this may explain why significant differences between the diet and diet plus cholestyramine groups were only found with these data4. Examining changes in orographic indices as continuous variables instead of selecting arbitrary ‘cut offs’ (that may reduce the power of the analysis) is preferred. There are numerous potential sources of error in QCA**“; we minimized these as follows. Tight criteria were imposed in selecting segments for analysis. Frames were consistently selected at the end-diastolic cardiac phase, where the frame-toframe changes occur at their minimum rate. Measurements from three adjacent frames at this
1
2
3
4
LDL Cholesterol
Fire
8
Association
between
in-trial
LDL-choiesferoI
level and
(a) dMAW’S,
phase were averaged. An automated edge detection algorithm of considerable sophistication and robustness was employed. This involved the firstbut not the second-derivat.ive of the image data, with consequent advantages in noise-avoidance and simplicity of correcting systematic errors arising from blurring by the imaging system. Selection of the single frame showing the instantaneous most severe visualization of stenosis would result in stenosis values of apparent maximum severity, but these values would be prone to be overestimates because of the noise level in individual images. We have adopted the more systematic approach of using stenosis values from S-frame averaging. Great care was exercised in ensuring the actual catheter dimensions were associated with that part of- the catheter tip analysed. Pincushion distortion was also corrected for. Practical advantages arise from separating the analysis phases of imagedigitization and of imageprocessing by digitally storing images from an entire intervention trial: frames can be more criticalIy matched when an analyst recalls them onto the monitor; averaging measurements from adjacent film frames becomes simple; future access to the film could be uI~available~ i~lformation in the film could be lost if the film were damaged: digital information can be copied losslessly. The image archive accumulated during the STARS trial is currently one of the largest collections of high-resolution digitized ( 1024 x 1024 pixel) coronary cin~~ngiographic images in the world. It is intended to add images from other studies in due course. As this database expands
(h) -1MinAWS.
(c) XII,
5
6
7
8
(mmoi/l)
and (d) J%inS
it should allow convenient investigation of future hypotheses related to the variability and classification of coronary morphology and t,o the development, location, extent and severity of atherosclerotic disease. It would be advantageous if internationally accepted standards were to be established for data processing in studies of this kind, as this would facilitate comparison of results. So called regression trials are a relatively new approach for assessing the effect of lipid lowering therapy on the course of coronary atherosclerosis. Various randomized trials (FATS”, SCOR’*, Lifestyle’“, Leiden”“, NHLBI”‘, CIM”2, POSCH=) have been congruent in demonstrating that lipidlowering treatment beneficially affects change in luminal dimensions. Associated change in angina1 symptoms has been recorded in Lifestyle and STARS, and that changes in sequential angiograms may be employed to predict risks of subsequent mortality due to coronary heart disease has also been suggested by data from the POSCH study. Despite substantial progress in other imaging modalities, notably computed tomography, ultrasound and nuclear magnetic resonance, and despite the many limitations referred to above, remains the definitive procedure for c&x assessing changes in atherosclerosis in clinical trials.
We are grateful for support Bristol-Myers Laboratories.
from Unilever Ltd and This work has ben-
363
Qwntitation
of comnaly
cinamgiogram:
J.N.H.
Brunt
et al.
Figure ! 3 (a) An example of a patient whose disease progressed with usual care. These are smal1 portions of frames. Each portion of 256 x 256 pixels was taken from the 1024 x 1024 pixel stored digitized version of the frame. Detected vessel boundaries are shown as continuous li ines. The d otted lines further away from the vessel on each side indicate the ends of the scan rectangles for sampling image grey values {see also Figure 21. The arrow heads indicate a part of the vesset where the change in disease (narrowing of the vessel) is parti~nlarly large. (b) The u idth functions derived from the vessel boundaries
efited also from a project concerned with angiographic blood flow measurement, supported by the British Heart Foundation. Thanks also to the conscientious referees for constructive suggestions. 6.
REFERJINCES 1. Leung WH, quantitation
2. 3.
4.
5.
364
Sanders
WJ, Alderman
EL. A coronary
artery
and data management system for paired Catheterization and Cardiouascular Diagcineangiograms. nosis 1991; 24: 121-34. Reiber JHC, Serruys PW (eds). 7%~ Duelists in Quantitative Coronary A~o~u~~y’ Kluwer Academic, Dordrecht, 1988. Reiber JHC. An overview of coronary quanti~tion techniques as of 1989. In ‘~~ntitati~e Choral A~o~u~hy ’ Reiber JHC, Serruys PW (edsf , Kluwer, Dordrecht, 199 1, 55-132. Watts GF, Lewis B, Brunt JNH, Lewis ES, Coltart DJ, Smith LDR, Mann JI, Swan AV. Effects on coronary artery disease of lipid-lowering diet, or diet plus cholestyramine, in the St. Thomas’ Atherosclerosis Regression Study (STARS) Lancet 1992; 339: 563-9. Austen WG, Edwards JE, Frye RI et al. AHA committee report. A reporting system on patients evaluated for cor-
7. 8.
9. 10.
onary artery disease Circulation 19’75, 51: 5-40 (Note: these pages are numbered separately from the conventional numbering scheme of the journal, and are published at the back of the April issue. Not all libraries have retained them because of their proximity to the advertising section), Montanari, LJ. On the optimal detection of curves in noisy pictures Comm. of the ACM 1971; 14: 335-45. Azzopardi PJ, Pycock D, Taylor CJ, Wareham AC. An experiment in model-based boundary detection Pvoc. 4th A&y Vision Conference (AVC88) 1988, 31-6. Selzer RH, Shircore A, Lee PL, Hemphill L and Blankenhorn DH. A second look at quantitative coronary angiogof the art in raphy: some unexpected problems. In ‘State ~ua~titut~~e Canal A~~o~a~h~’ Reiber JHC, Serruys PW feds) Martinus N~hoff, Dordrecht, 1986, 125-43. Doi K and Rossmann K. Effect of focal spot distribution on blood vessel imaging in ma~i~catio~ radiography. Radiology 1975; 114: 435-41. Spears JR, Sandor T, Als AV et al. Computerized image analysis
for
quantitative
measurement
of vessel
diameter
from cineangiograms. Circulation 1983; 68: 453-61. 11. Keiber JHC, van der Zwet PMJ, von Land CD, Koning G, van Meurs B, Buis B, van Voorthuisen AdE: Quantitative coronary arteriography: equipment and technical requirements. In Advances in Quantitative Coronary Arteri-
ogru@y Reiber JHC, Serruys PW (eds): Kluwer, Dordrecht, 1992; 75-l 11. 12. Mackay SA, Potel MJ, Rubin JM. Graphics methods for tracking three-dimensional heart wall motion. Computers and Biomedical Research 1982; 15: 455-73. 13. Colchester ACF, Hawkes DJ, Brunt JNH, du Boulay GH, Wallis A. Pulsatile blood flow measurements with the aid of .3-D reconstruction from dynamic angiographic recordings. In Infmmation Processing in Medical Imaging-9. Bacharach SI, (ed.) Martinus Nijhoff, Dordrecht, 1986: 247-65. 14. Brunt.JNH, du Boulay GH, Wallis A, Wicks DAG. Fluid flow measurement in vessels from digital subtraction angiography video recordings. Neuroradiolog? 1991; 33
(Suppl):
456-i-8.
15. Brunt JNH, du Boulay GH, Wallis A. The measurement of blood flow waveforms from X-ray angiography. Part 2: Application to video recordings of digital subtraction angiography: Proc I&n Mech Engrs: Journal of Eqineting in Medicine 1992; 206, Part H: 87-91. 16. de Feyter PJ, Serruys PW, Davies MJ, Richardson P. Lubsen J, Oliver MF. Quantitative coronary angiography to measure progression and regression of coronaq atherosclerosis. Circulation 1991; 84: 412-23. 17. Brown G, Albersg, Fisher LD et nl. Regression of coronary artery disease as a result of intensive lipid-lowering therapy in men with high levels of apolipoprotein B. N I:‘ngl, J Med 1990; 323: 1289-98.
18. Kane JP, Malloy MJ, Ports TA et al. Regression of coronary atherosclerosis during treatment of familial hypercholesterolemia with combined drug regimens.,JAMA 1990: 264: 3007-l”. 19. Ornish D, Brown SE, Scherwitz LW. d ~1. Can lifestyle changes reverse coronary heart disease? Lnnc~t 1990; 336: 129-33. 20. Arntzenius AC, Kromhout D, Barth JD el al. Diet, lipoproteins, and the progression of coronary atherosclerosis: NPW Engl J Med 1985; 312: 805-l 1. 21. Brensike JF, Levy RI, Kelsey SF et nl. Effects of therapy with cholestyramine on progression of coronary arteriosclerosis: results of the NHLBI type II coronary intervention study. Circulation 1984; 69: 313-24. 22. Blankenhorn DH. Nessim SA. Johnson RI,, Sanmarco ME, Azen SP, Cashin-Hemphill L. Beneficial effects of combined colestipol-niacin therapy on coronary atherosclerosis and coronary venous bypass grafts. JAMA 1987; 257: 3233-40. 23. Buchwald H, Marco RI,, Matts JP, PI al. Effect of partial ileal bypass surgery on mortality and morbidity from coronary heart disease in patients with hypercholesterolemia-Report of the Program on the Surgical Control of the Hyperlipidemias (POSCH). N Engl ,J Mpd 1990; 323: 946-55.
365