A preliminary study on adaptive field-of-view tracking in peripheral digital subtraction angiography

Journal of X-Ray Science and Technology 11 (2003) 149 159 149 IOS Press A preliminary study on adaptive field-of-view tracking in peripheral digital subtraction angiography James R. Bennett a,b,er-weibai c, John I. Halloran d and Ge Wang a,e a CT/Micro-CT Laboratory, Department of Radiology, University of Iowa, 200 Hawkins Drive, Iowa City, IA 52242, USA b E-mail:James-R-Bennett@uiowa.edu c Electrical and Computer Engineering, University of Iowa, 4316 Seamans Center, Iowa City, IA 52242, USA E-mail: Er-Wei-Bai@uiowa.edu d Department of Interventional Radiology, Cedar Valley Medical Specialists, 602 Ansborough Ave, Waterloo, IA 50604, USA e E-mail: Ge-Wang@uiowa.edu Abstract. About two million peripheral angiographies are performed annually in the United States, hence a reduction in exposure would yield significant healthcare benefits. The synchronization of bolus traveling, the table motion, and the fluoroscopic imaging chain can be highly effective for dose reduction in Digital Subtraction Angiography (DSA) by minimizing the field-of-view according to the vascular anatomy of the region traveled by the bolus. The goal of this paper is to demonstrate the feasibility of adjusting the field-of-view while tracking the contrast bolus, thus reducing the dosage of both the bolus and the radiation. The speed of the bolus is respectively estimated in the systole and diastole stages. An EKG-gated Hammerstein model is used to predict the bolus chasing speed. A real-time algorithm is designed to extract the bolus dynamics, and define the field of view transversely and longitudinally. A limb stabilization technique is also presented to suppress any significant image misalignment. Our simulation results show that the proposed techniques are promising for clinical applications. Keywords: Digital subtraction angiography, real-time imaging 1. Introduction The goal of Digital Subtraction Angiography (DSA) is an optimized image showing only opacified vessels. This is accomplished by taking an initial mask image of the region of interest without any radiographic contrast applied. A subsequent image of the same region is acquired after contrast has been injected into the vessels. The mask is subtracted from this image yielding an image demonstrating only contrast filled vessels. This technique is used extensively in leg angiography. The field of view of the imaging device is smaller than the leg, therefore, it is necessary to image this limb in multiple steps or by a technique referred to as bolus chasing. Bolus chasing is a method of following the contrast bolus by smooth, non-segmented motion of the either the examination table or the camera. In stepping angiography the mask images are taken in regions or stations. These stations are aligned below each 0895-3996/03/$8.00 2003 IOS Press. All rights reserved

150 J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral other along the longitudinal axis of the leg, typically starting at the pelvis and moving down to the foot. After the preparations for the contrast injection are complete, the fluoroscopy machine returns to the exact position of the first mask image. When contrast is injected, the fluoroscopy machine simultaneously energizes at a specified frame rate. When the leading edge of the contrast bolus approaches the bottom of the first frame, the clinician presses a button. This advances the frame to the next station, located immediately below its previous position and slightly overlapping it. The sequence repeats itself until the frame is at the last mask position. Each of the mask frames are subsequently subtracted from the contrast images, yielding a map of the contrast agent through the leg. Many studies have been performed on bolus tracking in CT and MR angiography [1,3,6,8 16]. However, bolus tracking in DSA appears to be limited to autostepping technology that has been developed by some manufacturers of fluoroscopy equipment [17]. In this paper, we propose a series of techniques focusing on optimization of fluoroscopic equipment design. The main approach of these techniques is adapting frame location and field of view to the contrast bolus characteristics. Section 2 is the main section of this paper in which we describe the methodology of our reduction techniques. Subsection 2.2 describes a technique to reduce the number of frames taken in the first station of the DSA procedure. Subsection 2.3 depicts a frame size reduction method and the accompanying controls involved. Section 3 explains a bolus velocity modeling technique. Section 4 is a brief description of preliminary results from the bolus tracking technique. Finally, Section 5 is a brief discussion and conclusion. 2. Methods and materials 2.1. Data acquisition The raw data used in this paper was obtained by digitally scanning films from conventional clinical leg DSA procedures performed on a Siemans Multistar. These films contain all of the subtracted frames from the contrast run, as well as the mask images. The contrast frames sequentially are displayed to reproduce real-time bolus propagation. Each contrast frame is accompanied by the frame rate and specific location of its station. This information allowed the authors to very accurately reproduce the timing and position of the bolus movement. The proposed techniques will then be applied to the data. Furthermore, the data will be used to compare the exposure reduction that might be obtained by utilizing the proposed techniques. 2.2. Exposure reduction in the first frame The clinical data collected indicates that often there are images acquired at the first station (of the stepping angiogram) that could be eliminated without loss of pertinent data. Three ways by which these non-diagnostic images may occur include the following. First, images may be acquired prior to the contrast bolus arrival in the imaging frame. This may occur if the distal tip of the catheter is remote to the imaging frame. Second, a slight delay occurs after the initiation of the contrast injection and the arrival of the contrast bolus to the distal tip of the catheter (frame 1, Fig. 1). There is a delay option on most contrast injectors and most manufacturers suggest utilizing at least a one second delay to avoid missing the bolus edge. If a delay is used, this can be another potential source of superfluous images. Third, extra images may result from slow propagation of the contrast through the frame (Fig. 1). This

J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral 151 Fig. 1. Frame series of bolus propagation in the first station. can occur in diseased vessels and in patients with poor cardiac output. Many of these frames may be eliminated if one could predict the velocity of the bolus. The proposed frame reduction technique utilizes two initial images to estimate the bolus velocity. Using this velocity, it is possible to approximate the time it takes the bolus to travel from the top to the bottom of the imaging frame allowing for optimal opacification of the vessels in the frame. There are difficulties in estimating the bolus velocity. Pulsatility of blood flow in the large arteries is the most prominent. As shown in Figs 2 and 3, the fluid velocity surges during systole and is minimal or absent during

152 J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral Fig. 2. Distance propagated during diastole in the first station. diastole. To compensate for this variable velocity, EKG readings from the patient may be integrated with this prediction algorithm. The algorithm will use these readings to select appropriate timing for the two initial exposures. Most accurate and reproducible exposures likely would be obtained during times of more uniform velocity i.e. diastole. These exposures would be obtained during consecutive periods of diastole. By acquiring the distance propageted by the leading edge of the bolus during systole using a bolus tracking algorithm it will be possible to predict the number of systolic intervals required to propagate the bolus to the bottom of the frame. Combining this information with the heart rate, the algorithm could accurately predict the time required for the bolus to reach the bottom of the frame. The fluoroscopy machine will delay the third and final exposure until the bolus is predicted to arrive at the bottom of the frame, yielding optimal vasculature opacification, and most importantly reducing radiation exposure to the patient. 2.3. Frame size reduction Reduction of the imaging frame size is another way to reduce patient radiation exposure. The size of the frame recording the bolus peak needs only be slightly larger than the bolus itself. Currently, this frame has the same height and width of the mask frame, represented in Fig. 4c with the bold outline. When following the contrast bolus peak with this frame size, excess area is radiated for a second time because current clinical practice uses the same frame size and positions of the mask run. Two mechanisms are required for tracking the bolus movement through the leg: transverse and longitudinal control units. 2.3.1. Transverse frame control Automatic tracking of the horizontal position of the bolus in real time presents a substantial difficulty. This is the result of differentiating between groups of small vessels and one large vessel. The clinician has detailed knowledge of the vascular anatomy. Using this, he/she can estimate the approximate position

J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral 153 Fig. 3. Distance propagated during systole in the first station. of the vessels of interest in relation to the well defined skeletal anatomy. The mask images could be compiled to form one complete image of the skeletal anatomy. The clinician would draw a line on the compiled mask image approximating the position of the vessels of interest as shown in Fig. 4a. The horizontal position of this line will be used to center the frame tracking the bolus. This may eliminate the need to automatically acquire the horizontal position of the main vessels. Because there is some degree of anatomical or surgically produced variation from patient to patient, the clinician still may need to adjust in real time the actual frame width while tracking the bolus according to his/her visual assessment of the actual position of the vessels. A hypothetical representation of the reduced frame sizes and positions is shown in Fig. 4b. 2.3.2. Longitudinal frame control Further exposure reduction is possible by creating an algorithm to minimize the overlap area between successive frames by monitoring, processing and predicting the peak bolus velocity. After the map is drawn, the clinician will select the initial contrast frame size and location on the bolus map. Once all preparations for the contrast run are complete, the clinician starts the contrast injection as well as the controlling software (further work may include automating this startup). When the software starts, it energizes the fluoroscopy machine at a set rate. The bolus location algorithm will process the image from every frame recorded by horizontally scanning each row of pixels. Each horizontal scan will summate the number of contrast pixels. The horizontal row that has a drop in contrast density below a threshold value will be used as the vertical position of the leading edge of the bolus. This algorithm is similar to the one proposed by Wu et al. The difference is that Wu et al., use rectangles instead of lines to estimate the position of the bolus [17]. Furthermore, the algorithm, working in real-time, will take the peak bolus positions and produce an optimal frame position on the predefined map and frame rate. 2.3.3. Limb stabilization There are a number of external factors that affect the precision of the acquired images. The most prominent of which is patient movement. When there is movement, the contrast and mask frames do

154 J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral Fig. 4. a) Mask Image with map drawn b) Light boxes represent possible frame reduction size c) Dark rectangles represent current frame size. not align correctly, producing inconsistencies in the subtracted image. These motion-induced errors adversely affect the quality of the subtracted image, the extent of which depends on the magnitude and complexity of the motion. The resultant movement can be attributed to the time delay between the mask and contrast runs and the physical effect of the contrast injection. The elements controlling the frame position are independent of any movement of the limb(s), therefore it is not possible to physically compensate for movement. Unfortunately, it is very difficult for the patient to remain motionless throughout the procedure, and in most cases it is necessary to correct for movement. By shifting the contrast and mask images in reference to each other using landmarks as references, it is possible to realign artifacts in the two images for subtraction. This process, called pixel shifting drastically improves the subtracted image in cases of patient motion. The drawback of this process is that when a patient moves his/her limb, artifacts in the image do not shift linearly or solely in two dimensions. This causes a varying amount of distortion because the compensation mechanism for movement is linear and two dimensional. A reasonable solution for this problem may be to simply fixate the limb being imaged to a stationary object. Use of a pneumatic clamping device to immobilize the knee and ankle would theoretically eliminate leg motion. Without limb movement, pixel shifting may be unnecessary or needed minimally. This has two very important impacts: distortion from image misalignment will be negligible and the sequence of mask images can be compiled to one complete mask image without performing extensive

J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral 155 Static-switch Nonlinearity Linear Block Systolic period V 1 Diastolic period V 2 EKG Readings Fig. 5. Interation of Hammerstein Model. adjustments associated with pixel shifting. These adjustments would be even further complicated using the reduced frame size technique because it would diminish the number of landmarks for alignment. 3. Control theory The key to bolus tracking is to model the bolus propagation. Bolus propagation modeling has been extensively investigated in many clinical studies. Provided that the pulsatility of the blood flow is not strong, the bolus propagation can be very well modeled by a linear system. Let the input be the propagation as a function of time at the injection position, the impulse response of the linear model is the convolution of a delayed normal density function and an exponential density. Unfortunately, this model alone is not very useful for our application. The pulsatility varies during angiography and becomes a key contributing factor in determining the propagation velocity. It was reported that the bolus velocity surges during systole and is comparatively minimal during diastole. Therefore, to completely represent the linear dynamics of the propagation during either the diastolic or systolic period as well as the switching nonlinear behavior, we propose a Hammerstein model [2]. The Hammerstein model is a special kind of nonlinear system that consists of a static nonlinearity followed by a linear dynamic system. Note that our purpose is to estimate the dispersion velocity so that an accurate time can be predicted at which the vessel contrast is highest at the next station and consequently X-ray images can be taken. Clearly, what we need is not a complete dispersion model but a simple and reliable way to estimate the velocity. To this end, we propose a Hammerstein model in which the dispersion velocity is represented by two constants depending on the pulsatility (Fig. 5): { V1,systole, V = V 2, diastole, Where the velocities V 1 and V 2 are estimated from experiments (Figs 2 and 3). From our experiments, V 1 =15cm/s and V 2 =1.5cm/s. This model is fairly simple but is good enough to capture the fluid velocity. A much more detailed model could be developed if more accurate velocity estimates are required. To incorporate the pulsatility information, EKG readings from patients are integrated with prediction algorithm. Once the bolus propagation velocity is estimated, control part can be easily established. Let y[kh] represent the actual position of the imaging device at time kh that is clearly subject to the dynamics of the fluoroscopy machine y[(k +1)h] =α(z 1 )y[kh]+β(z 1 )u[kh]

156 J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral where z is the z-transform operator and u is the control input to the machine. When the bolus propagation velocity becomes available, it is trivial to calculate the desired position for the imaging device y [(k+1)h]. Now, by defining the control signal β(z 1 )u[kh] =y [(k +1)h] α(z 1 )y[kh] the desired position at time (k +1)h is achieved. y[(k +1)h] =y [(k +1)h] 4. Results A preliminary algorithm was designed and applied to demonstrate the feasibility of the bolus velocity prediction algorithm. The algorithm integrates the horizontal pixel contrast values in an image. The profile of these sums indicates the position of the leading edge of the bolus. The bolus velocity can be determined by measuring the displacement of the bolus leading edge per unit of time. The bolus velocity is different during systole and diastole. Therefore, one must calculate both systolic and diastolic velocities in order to achieve higher prediction accuracy. Figure 6 represents a simulation of systolic velocity capture. Two frames were taken before and after systole. The summation algorithm was applied to these images and the resulting data was graphed according to the vertical position of the pixel row. An experimental contrast threshold was set (dotted bold line Fig. 6). A drop of the contrast density below threshold indicated the vertical position of the bolus leading edge (bold arrows Fig. 6). The bolus displacement is determined by measuring the distance between the first subthreshold value in the initial frame and the first subthreshold value in the reference frame (dotted arrow Fig. 6). The timing of the systolic interval was estimated because the device used to acquire these images was not integrated with the patients EKG. The frame rate information, e.g., 3 frames per second allows calculation of the time delay between images. Since one has both the traversed distance and the time delay, the velocity is known, e.g., 15 cm/second. This technique for velocity estimation will be incorporated into the previously mentioned Hammerstein model. Furthermore, the velocity estimation will be used in the exposure reduction method in subsection II.B. The techniques described in this paper, when applied to our experimental and simulated data, provide an accurate and real-time prediction of bolus location. This may allow for a reduced field-of-view in capturing the bolus, which would lead to a reduction in radiation exposure and possibly a reduction in contrast dose. A liberal representation of a simulated field-of-view was shown in Fig. 4b. This simulation resulted in an about 75 percent reduction of radiated area in the frames shown. Depending on several factors related to a specific case, the actual value might be about 50%, being less optimal than our idealized simulation. The reduction in contrast dosage is very hard to quantify without further experimentation and simulation. 5. Discussion and conclusion The current peripheral DSA uses two kinds of methods. First, using an algorithm embedded in the controlling software, some equipment is able to Autostep between stations by tracking the leading edge of the bolus, eliminating the clinician issued stepping commands. This technique helps minimize step-timing errors, which in worst cases may require a re-run of the DSA procedure. The second method

J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral 157 Fig. 6. Horizontally summated contrast density graphs with corresponding frames. The Y-axis represents vertical distance in frame (1 mm spacing). The X-axis represents the summated contrast density. utilizes bolus chasing. In this case, the imaging device is stationary and the patient lies on a floating table. The clinician manually adjusts the table position to keep the bolus peak in view of the imaging device. Both of these strategies are practical solutions, and have room for improvements. The techniques

158 J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral proposed in this paper represent a major refinement of the existing solutions to this problem for significant radiation exposure reduction. However, the techniques may present control problems by exceeding the capabilities of current machinery. With further testing, these problems will become clearer and the necessary changes to equipment design criteria will compensate to meet these control requirements. The method for compensation of variable bolus velocity may be used for CT angiography as well. A few algorithms have been developed for modeling and predicting bolus behavior in CT imaging [1,4,5]. Integration of the Hammerstein model using EKG data into these algorithms may enhance their ability to predict bolus velocity. A well defined bolus velocity allows for further precision in estimating bolus location, which may result in reduction of contrast dose and/or radiation exposure. In conclusion, we have demonstrated that peripheral DSA techniques can be improved by integrating our proposed techniques. Further testing and development of these techniques will be needed for their application in actual clinical procedures. The limb immobilization technique requires little modification of the existing technology. The field-of-view and rate optimization may require significant changes in the hardware and software of the fluoroscopy machine. Further evaluation and refinements of the proposed techniques will be performed in the future. We hope that these concepts will lead to development of effective patient radiation reduction techniques. Acknowledgment We appreciate the advice and suggestions for improvement of this manuscript by Michael Vannier, MD at the University of Iowa. A special thanks to Laurence Bennett, PhD for his support in the development of this manuscript. References [1] D. Fleischmann and K. Hittmair, Mathematical Analysis of Arterial Enhancement and Optimization of Bolus Geometry for CT Angiography Using the Discrete Fourier Transform, J. of Computer Assisted Tomography 23(3) (1999), 474 484. [2] E.W. Bai, Frequency domain identification of Hammerstein models, IEEE Trans on Automatic Control 48(4) (2003), 530 542. [3] F. Cademartiri et al., Parameters Affecting Bolus Geometry in CTA: A Review, J. of Computer Assisted Tomography 26(4) (2002), 598 607. [4] G. Wang and M. Vannier, Bolus-chasing Angiography with Adaptive Real-Time Computed Tomography (U.S. Patent 6,535,821, allowed on 11/26/2002, issued on 3/18/2003). [5] G. Wang et al., Model of intravenous bolus propagation for optimization of contrast enhancement, Proceedings of SPIE 3978 (2000), 436 447. [6] G.D. Rubin et al., Multi-detector row CT angiography of lower extremity arterial inflow and runoff: initial experience, Radiology 221(1) (2001), 7 10. [7] K. Kump et al., Comparison of Algorithms for Combining X-Ray Angiography Images, IEEE Trans. On Med. Imaging 20(8) (2001), 742 750. [8] M. Prince et al., Contrast-enhanced Abdominal MR Angiography: Optimization of Imaging Delay Time by Automating the Detection of Contrast Material Arrival in the Aorta, Radiology 203 (1997), 109 114. [9] M. Boos et al., Contrast-Enhanced Magnetic Resonance Angiography of Peripheral Vessels, Investigative Radiology 33(9) (1998), 538 546. [10] M. Prince et al., Automated Detection of Bolus Arrival and Initiation of Data Acquisition in Fast, Three-dimensional, Gadolinium-enhanced MR Angiography, Radiology 203 (1997), 275 280. [11] M.K. Stehling, J.A. Lawrence et al., CT angiography: expanded clinical applications, American Journal of Roentgenol 163(4) (1994), 947 955. [12] R. Puls et al., Multi-Slice Spiral CT: 3D CT Angiography for evaluating therapeutically relevant stenosis in peripheral arterial occlusive disease, Rontgenpraxis 54(4) (2001), 141 147.

J.R. Bennett et al. / A preliminary study on adaptive field-of-view tracking in peripheral 159 [13] R. Watts et al., Anatomically tailored k-space sampling for bolus chase 3D MR digital subtraction angiography, Radiology 218 (2001), 899 904. [14] V. Ho et al., Optimization of Gadolinium-Enhanced Magnetic Resonance Angiography Using an Automated Bolus- Detection Algorithm (MR SmartPrep), Investigative Radiology 33(9) (1998), 515 523. [15] Y. Wang et al., Bolus chase MR digital subtraction angiography of the entire lower extremity in one minute, Radiology 207 (1998), 263 269. [16] Y. Wang et al., Timing algorithm for bolus chase MR digital subtraction angiography, Magnetic Resonance in Medicine 39 (1998), 691 696. [17] Z. Wu et al., Real-Time Tracking of Contrast Bolus Propagation in X-Ray Peripheral Angiography, IEEE Workshop on Biomedical Image Analysis (1998), 164 172. [18] Mallinckrodt, EPIX Medical And Mallinckrodt Announce Results From Phase II Clinical Trial For MS-325 Magnetic Resonance Contrast Agent, http://www.mallinckrodt.com/news/news releases/98-07-16ms-325.html, 1998.