| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Developments |
1 From the Departments of Radiology (Y.H.K., P.H.A., C.M.S.) and Medicine (E.R.M.), University of Pennsylvania Medical Center, 1 Silverstein, 3400 Spruce St, Philadelphia, PA 19104. Received August 1, 2002; revision requested October 7; revision received November 8; accepted December 10. Address correspondence to C.M.S. (e-mail: sehgal@oasis.rad.upenn.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and Methods Results Discussion REFERENCES |
|---|
© RSNA, 2003
Index terms: Arteries, extremities, 912.1298 • Arteries, US, 912.1298 • Blood, flow dynamics • Computers • Phantoms
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results Discussion REFERENCES |
|---|
Although the use of angiography and phase-contrast magnetic resonance imaging for measuring flow-mediated vascular dilatation has been proposed (11,12), ultrasonographic (US) imaging continues to be the most commonly used method (6–10,13–19). Real-time imaging, low cost, and ease of use without the need for image-enhancing agents are some of the reasons for this choice.
Currently, US is used exclusively for imaging the arteries in a longitudinal plane. The imaging transducer is oriented along the length of the artery, and the change in diameter in response to flow-related dilatation is measured. The primary reason for choosing a longitudinal view instead of a cross-sectional view is that the former provides clearer definition of the border between the lumen and the arterial wall. This approach has been very successful and has been adopted by many laboratories. However, the method is limited in its accuracy and sensitivity. For example, in healthy volunteers the maximum change in brachial artery diameter is about 10%–20% in response to flow-mediated vasodilatation (8,19). For an artery of 5 mm in diameter, this corresponds to a change of 0.5–1.0 mm, which is equivalent to a few pixels (on the order of 10) with normal image magnification.
Because the change occurs during 1–2 minutes after flow mediation, a sensitivity of greater than 2–3 pixels is needed to detect changes. In patients with compromised endothelial function, the diameter change is likely to be smaller than 10%–20%, requiring an even higher level of sensitivity. A slight change in image plane during scanning with a handheld transducer or movement of the artery during pressure cuff deflation can easily mask these changes and often makes them difficult to detect with a high level of confidence.
It is our hypothesis that the sensitivity of US measurements can be improved considerably by performing cross-sectional imaging during flow-induced vasodilatation. A commonly cited limitation of cross-sectional imaging is that it does not provide a clear definition of the arterial wall. Although this continues to be of some concern, recent advances in imaging technology involving the use of high-frequency probes and the compounding of images to reduce image speckle can provide images with sufficient spatial resolution for making vasodilatation measurements.
The goal of this study was to evaluate the use of cross-sectional US imaging for measuring flow-mediated vasodilatation. To this end, we propose a user-guided automated approach for measuring change in vessel area during vasodilation. The approach was tested on vascular flow phantoms. Finally, the feasibility of the approach was demonstrated with images of the brachial artery.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results Discussion REFERENCES |
|---|
Experiment 2
The second set of experiments consisted of imaging rubber tubing with pulsatile flow. The goal was to demonstrate that the proposed method can detect changes in cross-sectional area that occur during pulsatile flow. The pulsatility was generated by compressing the tubing distal to the transducer at an approximate rate of 1 Hz. The images were recorded on a videotape and digitized at 30 frames per second. The cross-sectional boundaries of the tubing in the first 100 sequential frames were traced manually by C.M.S. and with a computer by using UGABD. The area measurements were obtained by counting the pixels enclosed in manually or UGABD-traced boundaries. UGABD and manual measurements were compared statistically by using correlation techniques.
Experiment 3
The third set of experiments was conducted to demonstrate the feasibility of using UGABD in the brachial artery of a human subject to detect flow-mediated dilatation. The approval of the Institutional Review Board of the Office of Regulatory Affairs of the University of Pennsylvania, as well as informed consent from the volunteer, was obtained. Occlusion was created by inflating a sphygmomanometric cuff on the arm. After maintaining an inflation pressure of 150–200 mm Hg for 5 minutes, the pressure was released quickly to induce a brief high-flow state and vasodilatation of the brachial artery. The brachial artery was imaged continuously in the cross-sectional plane for 5 minutes after the release of the pressure cuff.
The images were recorded on videotape and digitized at 30 frames per second by using an analog-to-digital media converter (Model DVMC-DA2; Sony, Tokyo, Japan). The first 100 frames of digitized images from each phase of the experiment (ie, 1 minute before inflation of the pressure cuff and 1, 2, 3, 4, and 5 minutes after pressure cuff deflation) were chosen for analysis of changes in arterial area as a function of time during pressure release. To simulate longitudinal imaging, the changes in diameter of the brachial artery along the long axis were measured in the same data. The images obtained 1 minute after pressure release were used to compare computer (UGABD)– versus human (C.M.S.)–traced boundary studies.
All US imaging was performed with a 12-5-MHz broadband linear-array transducer and an ATL 5000 scanner (Philips ATL, Bothell, Wash).
UGABD Algorithm
Accurate segmentation of images is critical in determining the area of blood vessels. The major difficulty arises in identifying the edges between the lumen and the vessel wall. Often, the "knowledge of human experts" is applied in searching for the edges of the object of interest. This process is carried out either with manual tracing or by using rule-based criteria to classify pixels within the region of interest. The method adopted in our study incorporates a rule-based approach for measuring the area of the cross section.
The algorithm is based on the hypothesis that if the images are acquired at a high frame rate, the coordinates of the arterial wall in an image will be in close proximity to the coordinates of the vessel wall in the previous frame. That is, the coordinates of a lumen–arterial wall boundary in any given frame can be used to search for the arterial boundary in the next consecutive frame. This process, if repeated in sequence from one frame to the next, can be used to detect boundaries in a series of consecutively acquired images. The process in which a user defines a boundary in one image and this information is "propagated" from one frame to the next for automated boundary detection is referred to as UGABD. This algorithm is described quantitatively below.
The initial boundary, IB(t), of a blood vessel in an image acquired at time t is defined by N points with pixel coordinates Xu and Yu as follows:
|
The IB(t) on an image can be either user defined or obtained from the analysis of a previous image acquired at time t - 1. IB(t) is used as a guide for determining the final arterial boundary. For boundary detection, the highest contrast possible in gray scale is desirable. In US images, a blood vessel appears as a hypoechoic region surrounded by echogenic arterial walls. The radial directions originating from the center of the blood vessel provide the highest contrast at the lumen–arterial wall edges and were therefore the basis of our analysis. Radial profiles in brightness, as shown in Figure 1, are extracted from the images along equispaced lines from the center of gravity of IB(t). The Cartesian coordinates XR and YR of each pixel on the extracted radial line were determined with the Bresenham algorithm (20). If Gt(XR,YR)
represents the gray-level values of the rasterized pixels (XR,YR) of the radial line at an angle , the brightness profile P(t) along is calculated as
|
and R = 1, 2,..., Rmax.
|
(t) was then convolved with a one-dimensional arithmetic filter to suppress high-frequency noise. This was followed by convolution of P(t) with a one-dimensional edge filter to enhance the edge feature. The sizes of the arithmetic filter and edge filter were chosen arbitrarily to be 10% of Rmax of the P(t) profile. The edge filter used was a step filter consisting of -1 and 1. Lumen–arterial wall edge was determined by detecting the zero crossing of the first derivative of the edge-filter–convolved P(t). The search region for boundary detection was restricted to ±10% of the IB(t). The rationale for this restriction was that, because the images were acquired at a rapid frame rate of 30 frames per second, the boundaries were not thought to deviate substantially from one frame to the next. If more than one peak was detected in the search region, the peak with larger amplitude was chosen to define the border. The edges detected for all the angles were used to construct the final boundary, FB(t), which was described by a new set of edge coordinates, Xf and Yf, as follows:
|
The arterial area A(t) for the image at time t was measured by counting all the pixels enclosed within the detected boundary FB(t) (21). To determine absolute area, we multiplied the number of pixels by pixel size. Pixel size was determined by measuring the number of pixels between two markings on an image separated by a known distance.
In the final step, FB(t) was low-pass filtered to reduce noise and propagated to the next image, t + 1—that is, FB(t) was defined as IB(t + 1). The steps described in Equations 1–3 and the area measurements were repeated with consecutive images to generate arterial area–versus-time curves.
Statistical Analysis
Pearson correlation (R2) between areas measured with manual tracing and those measured with computer-detected (ie, UGABD) boundaries was determined by using commercial software (Excel; Microsoft, Redmond, Wash).
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results Discussion REFERENCES |
|---|
To assess the sensitivity of UGABD for detecting pulsatility during motion similar to that observed in arteries, elastic rubber tubing was imaged in both steady and pulsatile flow conditions. Figure 2 shows the cross-sectional area measurements made with images of the rubber tubing phantoms. As anticipated, the cross-sectional area did not change (ie, there was less than 0.5% peak-to-peak variation) in steady flow conditions but underwent a cyclic change (showing approximately 6% peak-to-peak variation) in pulsatile flow conditions. The UGABD algorithm performed robustly and detected boundaries in all of the images analyzed (approximately 1,000 frames). A uniform cyclic change in cross-sectional area was observed in pulsatile flow at approximately 1 Hz or 30 frames per second (Fig 2). The mean diameter during pulsatile flow was 109.5 mm2, versus 107.6 mm2 during steady flow.
|
|
|
|
|
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results Discussion REFERENCES |
|---|
The area measurement performed with the rigid tubing provided a direct comparison between UGABD and actual measurements. A difference of less than 5% between the two measurements was observed. This difference was probably due to systematic errors in the two measurements. The actual area of the phantom was calculated on the basis of the diameter of the tubing by assuming circular geometry. A small error in the diameter measurement or an error due to rounding of numbers would have been increased by the power of two in area measurements. In addition, the deviation of the cross section from the circular shape assumed in the area calculations also may have contributed to the difference in the measurements. Errors may have arisen during imaging itself because the image plane may not have been exactly orthogonal to the long axis of the tubing. Finally, because area measurements required definition of the pixel dimensions, a small error in calibration would have been amplified in the calculation of area.
For the proposed technique to be successful, it should be sensitive enough to detect changes in cross-sectional area during pulsatile flow. The experiments with rubber tubing evaluated this aspect of the technique. The mean area of the rubber tubing during pulsatile flow was 109.5 mm2, as compared with 107.6 mm2 during steady flow conditions. The difference of 1.4% is small and is due either to systematic errors in the measurements or to slight inflation of the tubing during pulsatile flow. During steady-state flow, the areas measured were unchanged. In contrast, when the flow was pulsatile, area measurements showed a clear cyclic pattern, with peak-to-peak variation of 6%. Both of these results are consistent with expectations and demonstrate that the UGABD algorithm is robust and has the sensitivity to detect the rapid changes in area during pulsatile vasomotion.
UGABD-detected Boundaries versus Manual Tracings
The advantage of the UGABD algorithm versus manual tracing is that the former requires very little input from the user. Only one region of interest is needed for the algorithm to extract the remaining regions of interest. The question that follows is how well the UGABD algorithm estimates cross-sectional area compared with manual tracing. On images of phantoms obtained during pulsatile flow, the mean area measured with UGABD tracing was 3.2% ± 1.0 larger than the mean area measured with manual tracing. A similar comparison of pulsatile-flow images in the brachial artery showed a larger difference of 8.3% ± 4.4.
Interestingly, whereas for phantoms, the manually detected area measurements were consistently lower than the UGABD-detected area measurements throughout the entire data set, for brachial arteries, the trend was just the opposite, and manually detected area measurements were larger in all cases. This indicates that the UGABD algorithm and human boundary detection mechanisms have different but consistent biases. The cause for this difference is not known and may be related to the nature of the boundaries in the two cases. In the phantom study, the smaller discrepancy in detected area between the human observer and the UGABD algorithm can be attributed to clearer boundary definition in the US images. Both the human observer and the UGABD algorithm "agreed" on the locations of the actual boundaries. For brachial artery images, the larger discrepancy is likely due to the less clear, albeit sufficient, arterial wall boundary definition in the US images. Despite the differences, when phantom and brachial artery data were pooled together and the UGABD and manual tracings were compared directly against each other, an excellent correlation of R2 = 1.00 was observed, with unit slope. These results clearly demonstrate that although there may be differences between the two methods on a case by case basis, their results correlate highly on average.
Cross-sectional Imaging and Flow-mediated Dilatation
Monitoring of flow-mediated dilatation with US is increasingly being performed for assessing endothelial function. The automated boundary detection method reported herein can be directly applied for this purpose. According to our measurements, vessel cross section increased from the baseline value of approximately 8,000 pixels to approximately 11,700 pixels at 2 minutes after deflation of a pressure cuff. By 4 minutes, the cross-sectional area leveled off to approximately 9,900 pixels. This pattern of change in brachial artery dimensions is similar to earlier observations at imaging along the long axis (3,7–10,14,18,19).
The primary reason for choosing longitudinal imaging is that it produces good boundary definition. Although this is an important factor, longitudinal imaging has several disadvantages. Even in healthy persons, the increase in brachial artery diameter in response to vasodilatation is on the order of 10%–20%, or approximately 10 pixels. Such a small change in the number of pixels is problematic and in many cases does not adequately compensate for artifacts in measurements caused by the misalignment of the transducer or by the movement of the artery in and out of the imaging plane during the course of imaging. Another artifact that is commonly ignored in longitudinal imaging arises from the deviation of the shape of brachial artery from circular geometry due to compression of the overlying tissue by the transducer. Figure 3 illustrates this effect. The transducer compresses the artery along the vertical axis and elongates it laterally. Variations in pressure from the transducer can cause different degrees of compression and lateral dilatation. Because longitudinal imaging provides one-dimensional measurement, it cannot account for vasodilatation that occurs in orthogonal directions. As a result, scatter in the measurements obtained with the longitudinal imaging technique can mask some of the changes associated with flow-mediated dilatation.
A drawback of the cross-sectional mode of imaging is the lack of sufficient arterial wall definition. Further improvements in image quality are desirable, but the results of the present study show that cross-sectional US images in the current form show enough detail to enable boundary detection. Cross-sectional imaging captures arterial vasodilatation in all radial directions. The area measurements integrate changes in all directions that produce large changes in the number of pixels during flow-mediated dilation. In this study, an increase of 3,738 pixels from baseline to peak vasodilatation was observed. In the same data set, the change in diameter along the depth in the image (corresponding to one-dimensional measurement in longitudinal imaging) was only 23 pixels. Therefore, at the same US spatial resolution, vasoactive effects are amplified 150 times at cross-sectional imaging as compared with longitudinal imaging. Thus, it is reasonable to anticipate that cross-sectional imaging should provide higher sensitivity and better reliability in measuring flow-mediated dilatation.
In summary, we propose a new technique for automated boundary detection of the brachial artery in US images. The proposed method involves user-guided tracing and operates on radial profiles. Cross-sectional US images obtained with a state-of-the-art scanner have sufficient detail to enable boundary detection. The automated area measurements derived from the images were within 5% of the actual measurements. The proposed technique is easy to use and reveals larger change than longitudinal imaging during flow-mediated dilation. In the future, it may be feasible to measure flow-mediated dilatation in patients with higher sensitivity and reliability by using cross-sectional US.
|
FOOTNOTES |
|---|
Author contributions: Guarantors of integrity of entire study, Y.H.K., C.M.S.; study concepts, C.M.S.; study design, Y.H.K., C.M.S.; literature research, Y.H.K., E.R.M.; clinical studies, E.R.M., P.H.A.; experimental studies, Y.H.K., P.H.A., C.M.S.; data acquisition and analysis/interpretation, Y.H.K., C.M.S.; statistical analysis, Y.H.K.; manuscript preparation, Y.H.K., C.M.S.; manuscript definition of intellectual content and editing, all authors; manuscript revision/review and final version approval, Y.H.K., P.H.A., C.M.S.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results Discussion REFERENCES |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
