A New Automated Computerized Analyzing System Simplifies Readings and Reduces the Variability in Ultrasound Measurement of Intima-Media Thickness
Background and Purpose A computerized analyzing system with manual tracing of echo interfaces for measurement of intima-media thickness and lumen diameter in carotid and femoral arteries was previously developed by our research group and has been used for many years in several laboratories. However, manual measurements are not only time consuming, but the results from these readings are also dependent on training and subjective judgement. A further problem is the observed drift in measurements over time. A new computerized technique for automatic detection of echo interfaces was therefore developed. The aim of this study was to evaluate the new automated computerized analyzing system.
Methods The new system is based on dynamic programming and includes optional interactive modification by the human operator. Local measurements of vessel echo intensity, intensity gradient, and boundary continuity are extracted by image analysis techniques and included as weighed terms in a cost function. The dynamic programming procedure is used for determining the optimal location of the vessel interfaces in a way that the cost function is minimized.
Results With the new automated computerized analyzing system the measurement results were less dependent on the reader’s experience, and the variability between readers was less compared with the old manual analyzing system. The measurements were also less time consuming.
Conclusions The new automated analyzing system will not only greatly increase the speed of measurements but also reduce the variability between readers. It should also reduce the variability between different laboratories if the same analyzing program is used. Furthermore, the new system will probably prevent the problem with drift in measurements over time.
B-mode ultrasound allows direct noninvasive measurement of the intima-media complex in carotid and femoral arteries. The noninvasive ultrasound method is increasingly used in clinical research relating to the pathophysiology of atherosclerosis and also for the evaluation of preventive measures in randomized clinical trials.1 2 3 4 5 6 For measurement of IMT, most research groups use some kind of computerized analyzing system. These systems are usually based on manual tracing of the different echo interfaces in the ultrasound image.7 8 9 However, manual measurements are not only time consuming, but the results from these readings are also dependent on training and subjective judgement. A further problem is the observed drift in measurements over time.10 11
In previous studies of reproducibility the CVs for interobserver and intraobserver variability in measurement of mean IMT of the common carotid artery were 10.2% and 10.6%, respectively.7 10 This includes both recording and analysis of images with the manual analyzing system. Rereading variability, when the same images were measured twice but 12 months apart, was 3.8%. This means that approximately one third of the variability in the ultrasound method is due to variability in measurements. In accordance with other research groups, the data from rereading of IMT also showed that a small but significant drift may occur when readings are done a long time apart.10 11
To reduce the variability in IMT measurements, an automated analyzing system was designed and implemented in a PC/Microsoft Windows environment. The objective was to develop a boundary detection technique that not only is as accurate as possible but also robust, simple to handle, and less time consuming compared with manual tracing measurements. A dynamic programming procedure is used for the automatic detection of echo interfaces, which combines multiple measurements of echo intensity, intensity gradient, and boundary continuity. The system also includes optional interactive modification by the human operator. Analyzing systems based on automatic detection have previously been presented.12 13 14 However, in these systems the automatic reading is based on single measurements of echo intensity or intensity gradient (edge strength) along the vessel boundary, not taking boundary continuity into account.
The aim of this study was to evaluate the new automated analyzing system for measurement of IMT and LD in predefined sections of the carotid and femoral arteries.
Development and Validation Procedure
The development of the automated computerized analyzing system was part of a PhD thesis work at Chalmers University of Technology15 and has been in progress for approximately 3 years. All ultrasound images used for the development of the technique were recorded and analyzed at the Wallenberg Laboratory (see below). The research and programming of the automated analyzing system and also the present evaluation of the program were performed in a UNIX computer system. The program was thereafter reimplemented into a PC system with Microsoft Windows.
The first step in the programming of the new automated analyzing system was to measure a number of ultrasound images of good quality from the common carotid artery (27 images) with the manual analyzing system. The measurement contours from these measurements were saved and served as reference for the automatic outlining. Manual contours from a further set of 42 images of different qualities were added for refinement and validation of the automatic procedure. Another 60 images from the common carotid artery were used in an evaluation of the first version of the program. The manual contours from these images were thereafter added to the first 27+42 images to further improve the automatic outlining. This procedure was repeated a second and a third time with continuous refinement of the program. Manual contours from measurements in the carotid artery bulb and in the common femoral artery were also used to adjust the program to be applicable to these measurement areas as well. A total of approximately 500 images were used for the refinement and validation of the automatic detection system.
The final program version was thereafter tested in the present evaluation study. Images (not earlier used) from a consecutive group of patients and control subjects (n=50) were analyzed in both the old manual tracing analyzing system and the new automated analyzing system. Furthermore, the same images were independently analyzed by three technologists with different experience in reading of ultrasound images.
Examination was performed with an ultrasound scanner (Acuson 128) with a 7-MHz linear transducer and a transducer aperture of 38 mm. The electrocardiographic signal (lead II) was simultaneously recorded to synchronize the image capture to the top of the R wave to minimize variability during the cardiac cycle.
The carotid arteries were scanned at the level of the bifurcation, and the femoral arteries were examined distal to the inguinal ligament, at the site where the artery divides into the superficial femoral artery and the profound femoral artery. Images for IMT measurements were recorded from the carotid bulb, the common carotid artery, and the common femoral artery, respectively. At the position of the thickest part of the far wall (visually judged), a frozen longitudinal image was captured and recorded on videotape. The procedure was repeated three times to achieve three separate images for analysis (see below). A short sequence of real-time images was also recorded on videotape to assist in the interpretation of the frozen images.
Manual Measurement of IMT and LD
The ultrasound images from the videotape were first analyzed in the computerized analyzing system based on manual tracing of the echo interfaces described by our group in an earlier report.7 IMT was defined as the distance from the leading edge of the lumen-intima interface to the leading edge of the media-adventitia interface of the far wall. The measurement of IMT in the carotid artery was made in two separate segments: along a 10-mm-long segment in the common carotid artery and also along a 10-mm-long segment in the carotid artery bulb. Approximately 10 boundary points were marked along each echo interface by use of a digitizer table and a mouse. Between these marked points the echo interfaces were interpolated by the computer, so that 100 boundary points were analyzed for each 10-mm section. The computer program calculated the average thickness along the 10-mm-long section (IMTmean) and also the maximum thickness of the analyzed section (IMTmax). Mean and minimum LD of the common carotid artery (LDmean and LD min) were defined by the distance between the leading edges of the intima-lumen interfaces of the near and lumen-intima of the far wall.16 Measurements in the common femoral artery were made in a way similar to that for the carotid artery but along a 15-mm-long section proximal to the bifurcation into the superficial femoral artery and the profound femoral artery.4
Automated Measurement of IMT and LD
The images were also measured in the new analyzing system based on automatic detection of the echo structures in the ultrasound image but with the option to make manual corrections by the operator. In the new analyzing system, the digitized image and the program functions are shown on the same PC monitor (Fig 1⇓). An extra monitor is also directly connected to the tape recorder for visual evaluation of the real-time images on the videotape. This is done to assist in a correct interpretation of the interfaces in the frozen ultrasound images.
The starting point of the measurement area is set by the operator, and a 10- or 15-mm box is automatically drawn. The different echo interfaces are initially outlined fully automatically. If obvious errors are detected by the reader, it is possible to modify the measurement by marking a correct echo in the ultrasound image. In this case only one or two manually marked points are often needed to guide the automatic outlining to the correct interface in the whole segment. The analyzing system also has an option for manual marking along the whole echo interfaces in the segment. The program gives the average and also the maximum thickness of the intima-media complex (IMTmean and IMTmax) as well as LDmean and LDmin, as in the manual analyzing system. It is possible to save both the measurement results and the contours of the measurement lines. This option makes it possible to later reload the ultrasound image and also reload the measurement contours.
Automatic Detection of Echo Interfaces by Use of a Dynamic Programming Procedure
The automatic searching starts with an estimation of the approximate positions of the echo interfaces and finishes with the refined interface positions. The criterion for the optimized search is based on a cost function that contains three terms, each of which has a unique weight factor. The three terms reflect typical image features that can be associated with the interface between different anatomic structures. For a detailed description of general dynamic programming procedures, see Reference 1717 .
Cost Function and Determination of Weight Factors
An essential step in the development of the new automated analyzing system was to determine the terms to be included in the cost function and their weight factors (Fig 2⇓). These weight factors direct the relative importance of the terms: echo intensity, intensity gradient, and boundary continuity. For example, a low value of the weight factor for boundary continuity will allow for a more irregular boundary, whereas a high value forces the system to detect a more straight boundary. This has important implications for the handling of echo dropouts and irregularities along the vessel interface. Therefore, it is important that the weight factors are tuned to the specific image characteristics of each segment and each interface. The tuning is performed in a trimming procedure and should thereafter not be modified. If other ultrasonic instrumentations give other characteristics, the weight factors might need some adjustments.
Dynamic Programming Procedure
The estimated values of the three boundary features (echo intensity, intensity gradient, and boundary continuity) are included as weighed terms in the cost function so that each image point is associated with a specific cost that in turn correlates with the likelihood of that point being located at the echo interface. The boundary detection algorithm inspects all points (pixels) in the image (Fig 2⇑), considering all possible lines that may form the echo interface, and gives favor to that which minimizes the cost function. The number of all possible ways of forming a boundary in this manner is enormous, and therefore an exhaustive search cannot be applied. However, the optimization procedure referred to as dynamic programming can solve this problem with dramatically fewer operations.17
All statistics were analyzed with the use of SPSS for Windows 6.1. Means and SDs for differences between the readers were calculated. Interobserver error (s) was then calculated according to the formula s=SD/√2. The CV describes the difference as a percentage of the pooled mean value (x̄) and was calculated according to the formula For comparison between the two analyzing methods, the Mann-Whitney U test was used. The relationships were illustrated with Pearson’s correlation coefficient (r).
Comparison Between the Manual and Automated Analyzing Systems
The relationships between the two analyzing methods were highly significant (Table 1⇓). However, the automated analyzing system gave higher mean values of carotid IMT than the manual system (Table 1⇓). See “Discussion” and Fig 3⇓.
Variability Between Readers
The same images were measured by three laboratory technologists with different experience regarding analysis of ultrasound images. When two experienced technologists used the manual analyzing system, there was a significant difference in mean values of common carotid IMT. When the automated analyzing system was used, the mean values of IMT and LD were similar (Table 2⇓).
The CVs for IMT measurements between a technologist with long experience and a technologist without previous experience of analysis of ultrasound images were 1.2%, 3.8%, and 6.9% for common carotid IMTmean, carotid bulb IMTmean, and common femoral IMTmean, respectively (Fig 4⇓).
Manual corrections of the automatic outlining of the lumen-intima or the media-adventitia interfaces of the far wall were performed in 17% of all images from the common carotid artery. Most corrections were minor in nature (12%). Manual corrections of the interfaces in the far wall of the carotid artery bulb were performed in 58% (minor changes in 21%) of all images and in 67% (minor changes in 13%) of all images from the common femoral artery.
Comparison of Time for Analysis
Analysis of one subject included measurement of 3 images from the common carotid artery, 3 images from the carotid bulb, and 3 images from the common femoral artery. A complete analysis also included the choice of frozen images from the videotape and image digitization and also the evaluation of the real-time images on the tape. This whole procedure takes approximately 45 minutes with the manual analyzing system compared to 15 to 18 minutes with the new automated analyzing system. The measurement procedure for 3 images from common carotid artery and 3 images from the carotid artery bulb takes approximately 4 minutes.
The development of a new automated analyzing system improves and facilitates the ultrasound measurement of IMT and LD in carotid and femoral arteries. The analyzing program is built in a PC/Windows environment with the program functions and the digitized ultrasound image on the same monitor and with the different steps in the program mouse controlled. This makes the new analyzing station more compact than the earlier system, and it also has ergonomic advantages compared with the old manual analyzing system. The automatic outlining of the echo structures and the easy handling of the program functions make the measurements much faster compared with manual tracing of the echo interfaces.
A thorough evaluation of the new program has been performed. This evaluation showed that with the new automated analyzing system the measurement results were less dependent on the readers experience, and the variability between readers was also less compared with the old manual analyzing system. There was still a small difference in common carotid IMT between the readers (0.007 mm; Table 2⇑), which was due to the possibility of doing manual corrections. Therefore, when performing the analyses the readers should be directed to only correct obvious errors and false gaps in the echo interface defined by help of the real-time images on videotape.
Few corrections were made in the measurement of common carotid IMT. However, many of the images from the carotid artery bulb and the common femoral artery needed corrections. The main cause was the occurrence of plaques in these areas. One reason for a false detection could be echoes with high intensity within the plaque. To lead the automatic outlining to the correct echo was mostly easy and could sometimes be made by marking one or two points of the correct interface. Other corrections consisted of marking of gaps in the echo interface. Small gaps in the echo interface were mostly outlined correctly since the automatic procedure also includes a boundary continuity term. This also means that artifacts within the blood stream with low continuity are not falsely outlined.
For one of the readers the automated analyzing system gave higher values of IMT compared with the manual analyzing system. A possible explanation for this is illustrated in Fig 3⇑. The automatic procedure always marks the interface at the point of the maximal gradient, that is, the maximal change in echo intensity (or approximately at the point of the maximal gradient since the program also takes into account the other terms included in the cost function). When measurements are performed manually, the point of the maximal gradient is mostly marked, but sometimes the threshold for visibility of the echo interface (for the human eye) is above this point, especially in the weaker lumen-intima echo. In those cases the operator tends to mark more closely to the top of the intensity curve for the lumen-intima interface. This gives a thinner IMT compared with automatic measurements (Fig 3⇑).
A comparison was made between a technologist with long experience and a technologist without previous experience of analysis of ultrasound images. The CV for measurement of common carotid IMT was less than in previous studies of rereading with the manual analyzing system by the same experienced technologist.10 The CVs for measurement of carotid bulb IMT and common femoral IMT were similar compared with previous rereading studies.10
Automated analyzing systems that use the intensity gradient method (edge strength) for boundary detection have the advantage of applying well-defined measurement points since the algorithm always selects the points associated with the maximal gradient,12 13 14 but they also have some major disadvantages. The most prominent one seems to be the lack of a boundary continuity constraint. This makes the algorithm less robust because it frequently results in irregular boundaries due to echo dropouts and scattering phenomena. If the ultrasound image has smooth and clearly visible vessel interfaces along the whole segment, which is unusual, no major differences will be noticed between the boundaries detected by the intensity gradient method and the dynamic programming method. The reason for this is that both algorithms include a mechanism for searching points with high gradient, or edge strength, and the fact that the dynamic programming algorithm also includes a continuity term has no meaning in cases of images with smooth and clearly visible interfaces. But in the case of a partly irregular surface of the intima-media complex with echo dropouts, the dynamic programming algorithm will mostly bridge the gaps in the way a trained reader would have traced the interface. This is because the automatic program was trimmed, or tuned, to allow for the same amount of flexibility as shown by an experienced technologist. However, in cases in which the gap is wide or where there are several possible routes that the boundary detection could follow, even the dynamic programming algorithm may fail.
In conclusion, automatic measurement of IMT and LD is less time consuming. It also reduces the variability between readers and will probably also reduce the variability between different laboratories if the same analyzing program is used. Furthermore, the new automated analyzing system will probably prevent the problem with drift in measurements over time in a longitudinal study.
Selected Abbreviations and Acronyms
|CV||=||coefficient of variation|
This study was supported by grants from the Swedish Heart-Lung Foundation, the Swedish Medical Research Council (project B96-27X-10880-03A), and the Astra Hässle Cardiovascular Research Laboratories, Mölndal, Sweden. The authors acknowledge the excellent technical assistance of Caroline Schmidt and Anna Frödén. The computer laboratory of the Medical Faculty (MEDNET), Göteborg University, is also acknowledged for providing computing facilities.
- Received March 6, 1997.
- Revision received June 20, 1997.
- Accepted July 3, 1997.
- Copyright © 1997 by American Heart Association
Blankenhorn DH, Selzer RH, Crawford DW, Barth JD, Liu C-r, Liu C-h, Mack WJ, Alaupovic P. Beneficial effects of colestipol-niacin therapy on the common carotid artery: two- and four-year reduction of intima-media thickness measured by ultrasound. Circulation. 1993;88:20-28.
Salonen JT, Salonen R. Ultrasound B-mode imaging in observational studies of atherosclerotic progression. Circulation. 1993;87(suppl II):II-56-II-65.
Wikstrand J, Wendelhag I. Quantitative ultrasonography of carotid and femoral arteries. In: Lanzer P, Rösch J, eds. Vascular Diagnostics. Berlin, Germany: Springer-Verlag; 1994:129-138.
Furberg CD, Adams HP, Applegate WB, Byington RP, Espeland MA, Hartwell T, Hunninghake DB, Lefkowitz DS, Probstfield J, Riley WA, Young B, for the ACAPS Research Group. Effect of lovastatin on early carotid atherosclerosis and cardiovascular events. Circulation. 1994;90:1679-1687.
Bots ML, Hofman A, de Bruyn AM, de Jong PTVM, Grobbee DE. Isolated systolic hypertension and vessel wall thickness of the carotid artery: the Rotterdam Elderly Study. Arterioscler Thromb. 1993;13:64-69.
Wendelhag I, Wiklund O, Wikstrand J. On quantifying plaque size and intima-media thickness in carotid and femoral arteries: comments on results from a prospective ultrasound study in patients with familial hypercholesterolemia. Arterioscler Thromb Vasc Biol. 1996;16:843-850.
Touboul P-J, Prati P, Scarabin P-Y, Adrai V, Thibout E, Ducimetière P. Use of monitoring software to improve the measurement of carotid wall thickness by B-mode imaging. J Hypertens. 1992;10(suppl 5):S37-S41.
Gariepy J, Massonneau M, Levenson J, Heudes D, Simon A, and the Groupe de Prévention Cardio-vasculaire en Médecine du Travail. Evidence for in vivo carotid and femoral wall thickening in human hypertension. Hypertension. 1993;22:111-118.
Gustavsson T, Liang Q, Wendelhag I, Wikstrand J. A dynamic programming procedure for automated ultrasonic measurement of the carotid artery. IEEE Comput Cardiol. 1994:297-300.
Bellman RE, Dreyfus S. Applied Dynamic Programming. Princeton, NJ: Princeton University Press; 1962.