Study of an adaptive bolus chasing CT angiography

Abstract

To improve imaging quality and to reduce contrast dose and radiation exposure, an adaptive bolus chasing CT angiography was proposed so that the bolus peak position and the imaging aperture can be synchronized. The performance of the proposed adaptive bolus chasing CT angiography was experimentally evaluated based on the actual bolus dynamics. The experimental results show that the controlled table position and the bolus peak position were highly consistent. The results clearly demonstrate that the proposed adaptive bolus chasing CT angiography that synchronizes the bolus peak position with the imaging aperture by a simple adaptive system is computationally and clinically feasible. Similar techniques may also be applied to conventional angiography to improve imaging quality and to reduce contrast dose and/or radiation exposure.

1. Introduction

Atherosclerosis of the aorta, iliac and lower extremity arteries is common, and its effects on vessel lumen diameter may be hemodynamically significant. Approximately 2% of patients over age 45 have symptoms related to aorto-iliac or lower extremity atherosclerotic disease including pain, tissue loss and tissue infarction [11]. Aortic aneurysmal disease is another prevalent condition. Aneurismal rupture can be potentially fatal secondary to exsanguination [12]. Traditionally, evaluation of atherosclerotic and aneurysmal aortoiliac or lower extremity diseases are performed with conventional catheter-based angiography (CA). In the recent years, CT angiography (Fig. 1) has become a popular alternative [1-5] due to its non-invasive nature and faster scanning using multi-row detector CT scanners.

Fig. 1.

Computer simulated adaptive bolus chasing angiography. The purpose is to move the patient table according to the predicted bolus propagation so that the bolus peak position is synchronized with the imaging aperture.

In CT angiography, contrast material is often administrated to allow a narrow temporal window to obtain optimal visualization of vessels, lesions and tumors [17-19]. The quality of scans depends on the ability to synchronize patient table position with the relatively narrow aperture of the imaging system during propagation of a contrast bolus after intravenous injection. Currently performance of CT angiography routinely relies on scan initiation at a specific level after a pre-chosen threshold of intravascular enhancement is reached but once initiated, no further synchronization of table motion with contrast propagation is achieved. Linear table velocity may become problematic as time increases.

There is abundant literature on fluid mechanics and its applications including physiology [20,21], pharmacokinetics [22], and biomedical engineering [23,24]. Several clinical studies of CT contrast enhancement have been published in the past decade [18,25,26]. Bolus dynamics are complex and influenced by contrast administration protocol and patient characteristics (age, sex, weight, height, cardiovascular status, renal function, (etc.) [18,27]. Peak bolus velocity is rarely uniform; therefore synchronization of it with a fixed, preset table transport often results in less-desirable vascular enhancement. Lack of synchronization may be more problematic when scanning speed is fast, contrast volume is small, injection rate is high (leading to reduced peak duration), and/or variable vessel lumen diameter. Even for a normal individual, it is common that the contrast bolus velocity is rapid in the torso and relatively slow in the legs. Moreover, if asymmetric peripheral vascular disease exists, there may also be substantial variability in flow velocity between the opposite legs. Obviously, adaptive bolus chasing techniques are relevant to CT angiography because of the impact on image quality, as well as the need to limit contrast dose and radiation exposure.

To overcome these problems, several methods were reported in the literature [26,30-32] including, (1) test bolus timing, (2) ROI threshold triggering, and (3) visual cue triggering. The fundamental limitation of all these methods is that there is no table control component and the table motion is still linear which makes synchronization of the bolus peak position and the imaging aperture unlikely.

Our aim was to develop an adaptive bolus chasing CT angiography technique to improve imaging quality and reduce contrast dose and radiation exposure. Works reported here are the experimental results tested in a clinical environment based on actual bolus position and velocity data (Table 1). A key to bolus chasing CT angiography is the prediction of the future bolus position based on the current and past bolus positions. Clearly, the success of the proposed strategy depends on how accurately the model can predict. It was shown [34] that bolus propagation is governed by a set of very large number of partial differential equations which contains too many patient and circulatory stage-dependent parameters that make the model little use for adaptive bolus chasing CT angiography because real time estimation of such a large number of parameters is impossible. Another commonly used model to describe bolus dynamics is the compartmental model [27-29] that is also of little use for adaptive bolus chasing CT angiography. First, the model is a set of equations involving dozens of unknown parameters that are patient and circulatory stage dependent. Obviously, parameters about the patient vessel radius at each stage of the vascular tree are difficult to have in advance. Also, disease-state related parameters are impossible to be quantified prior to an angiogram. The second reason is that the compartmental model describes contrast enhancement specific to a compartment (organ or vessel) [27-29] instead of predicting the bolus dynamics as a function of time. To the best of our knowledge, there is no published result for contrast enhancement in CT which predicts both spatial and temporal propagation of the contrast bolus. To overcome this difficulty, an adaptive model is proposed that contains a minimum number of parameters and allows for accurate real time estimation. Notice that due to the complexity of the circulation system, the blood velocity surge during systole and may reverse during diastole. It is also possible that velocity of an intra-arterial bolus peak within a given peripheral vascular vessel differs from that of the blood within the same vessel. The goal of our adaptive system is to track the bolus peak position and has the best opportunity to optimize the image quality of CTA.

Table 1.

Summary of bolus data sets. (M/F is for Male/Female, Wt for White, AA for African American, Hs for Hispanic, Un for Unknown Race, St for Stricture, Ath for Atherosclerosis, Em for Embolism, Th for Thrombosis, AAA for Abdominal Aortic Aneurysm, and DM I/II for Diabetes Mellitus I/II)

Disease Category	Patient Information	Body Site	Bolus Speed (cm/s)	Mean Speed (cm/s)	ECG-Gated Delay
Occlusive- Stenosis	M/Wt/69;St	Femoral	14-23	10	5/10
	M/Wt/49; Ath	Illiac	10	3.2	2/10
		Femoral	16	5	2/10
	F/AA/56; Ath, St	Illiac	15	7	5/10
	F/Un/55; Ath	Illiac	10	5	3/10
	M/Wt/77; Ath, St	Illiac	8-10	6	2/10
	F/Wt/80; Ath, St	Femoral	8-10	3	2/10
Occlusive- Blockage	F/Wt/62; Em & Th, St	Illiac	15-20	10	2/10
	M/Wt/63; Em & Th	Illiac	9	5	2/10
Aneurismal	M/Wt/70; AAA, DM II	Ab Aorta	8-13	3	2/10
		Illiac	10	2.7	3/10
	M/Wt/79; AAA	Illiac	21	7	3/10
	F/Wt/80; AAA, Em & Th	Ab Aorta	19	2	2/10
		Illiac	11-16	5	5/10
	M/Wt/49; AAA	Ab Aorta	17-22	4	1/10
		Illiac	9-21	4	2/10
Micro-vascular/ Sub-Angiographic	M/Wt/88; DM II, St, Ath	Illiac	9-11	4	2/10
	M/Wt/77; DM II, Ath	Femoral	12-15	10	9/10
		Popliteal	10	7	1
		Popliteal	21	12	1
	F/Wt/35; DM I, Em&Th	Ab Aorta	31-38	16	1/10
		Femoral	15-17	11	2/10
	F/Hs/55; DM II, Ath	Illiac	13-17	5	2/10
	M/Hs/50; DM I, Em&Th, Ath	Femoral	16	6	3/10
	M/Wt/77; DM II, Ath	Femoral	13-14	6	2/10
		Popliteal	18-21	6	3/10
	M/Wt/52; DM II, Ath	Popliteal	10-14	7	1

2. Materials and methods

Using actual bolus positions and velocities extracted from patient data sets that we collected, the authors performed an experiment to test the feasibility and performance of the proposed adaptive bolus-chasing CT angiography scheme that combines two main components: (1) an adaptive model based on real time estimations and predictions of the bolus peak position, and (2) control of the patient table so that the bolus peak position is synchronized with the imaging aperture.

2.1. Bolus data acquisition and extraction

Bolus velocities change substantially depending on vessel diameters which vary greatly relying on age, sex and arterial tree level. Bolus velocities also change significantly from systole to diastole. To be able to test the proposed adaptive scheme in a clinical environment, we first collected actual bolus position and velocity data from routine diagnostic peripheral angiograms on a Siemens AXIOM-Artis utilizing iodinated contrast agent. This imaging modality provides a single plane (2D) fluoroscopic cine window. The scans are generally performed at 15 frames per second (fps). However, a few data sets were collected at a lower capture rate. Furthermore, the AXIOM-Artis system samples the ECG signal concurrently with the X-ray imaging. The sampling rate for the ECG signal is 400 samples per second. The data sets were saved in the Digital Imaging and Communications in Medicine (DICOM) Format, a medical standard in most modalities for transfer of images, movies, and other diagnostic data. Several third-party programs extracted the individual components of the DICOM data files, specifically the cine film and the ECG signal. The cine scenes were extracted from the DICOM file using the RUBO DICOM Viewer, and the ECG signal data was extracted by converting each DICOM data set into .XML (hypertext similar to HTML) format using “Kuratorium OFFIS ‘dcm2xml”’. We developed an algorithm using National Instruments’ LabVIEW to analyze the extracted data. Each DICOM patient data file was opened with RUBO. Then, the movie information was extracted and saved as a Windows Media Video Clip (AVI). The ECG data (native format is in the hexadecimal base) was manually retrieved from the XML file. The algorithm we developed using LabVIEW extracts every frame in the cine sequence and processes these images for analysis of bolus dynamics. First, the image must be converted from a RGB image (pixel is associated with three intensity values: Red, Green, Blue) into a grayscale image (pixel with only one intensity value). Then, the image is defined to a region of interest without significant interfering features. A process called Digital Subtraction is used to remove any stationary artifacts. For example, if each pixel in the current frame is subtracted from its counterpart pixel in the mask frame, stationary objects in the sequence will be suppressed. This will increase the conspicuity of the moving structures, i.e., the contrast bolus, in each frame of the sequence. Table 1 is the summary of bolus dynamic data sets along with their features of interests (M/F is for Male/Female, Wt for White, AA for African American, Hs for Hispanic, Un for Unknown Race, St for Stricture, Ath for Atherosclerosis, Em for Embolism, Th for Thrombosis, AAA for Abdominal Aortic Aneurysm, and DM I/II for Diabetes Mellitus I/II). A typical bolus peak position as a function of time is shown in Fig. 2 which shows clearly that the bolus velocity surges in systole and is relatively stationary in diastole.

Fig. 2.

Extracted bolus position (cm) versus time (second).

2.2. Patient table control

Another key component of our strategy involves a table control system. The purpose of this control system is to move the patient table according to the predicted bolus peak position so that the bolus peak position and the imaging aperture are synchronized. Most existing CT table control systems use an AC or DC stepping motor servo system to move the table. Mechanically and electronically speaking, the system is very complicated for patient comfort and safety. Because of the stepping motor servo system, however, the mathematical equation that describes the CT table motion is surprisingly simple provided that the motor has enough torque. Let Δt be the sampling interval and p_T (kΔt) be the table position at time kΔt. The table position at time kΔt + Δt is given by

$$p_T(k\Delta t+\Delta t)=u\left(k\Delta t\right)$$

where u is the command signal at our disposal informing the stepping motor how many revolutions it should move that translates into the table distance or position. Now, if the bolus peak position p_b(kΔt + Δt) of time kΔt + Δt were available at time kΔt, we could set

$$u\left(k\Delta t\right)=p_b(k\Delta t+\Delta t)$$

and this implies

$$p_T(k\Delta t+\Delta t)=p_b(k\Delta t+\Delta t)$$

a perfect synchronization between the bolus peak position and the controlled table position. The problem is that the future bolus position p_b(kΔt + Δt) is unknown at time kΔt. Therefore, estimation of the future bolus peak position is a key. Keep in mind however what we are interested in is not a complete bolus dynamics but to be able to predict the next bolus position. Let v(kΔt) denote the velocity of the bolus at time kΔt, the next bolus position is given by

$$p_b(k\Delta t+\Delta t)=p_b\left(k\Delta t\right)+v\left(k\Delta t\right)\Delta t$$

Suppose the sampling interval Δt is small, it is reasonable to assume that the two consecutive bolus velocities are close

$$\mid v(k\Delta t+\Delta t)-v\left(k\Delta t\right)\mid \le \delta$$

for some δ. Let $\widehat{v}\left(k\Delta t\right)$ and ${\widehat{p}}_b\left(k\Delta t\right)$ be the estimates of v(kΔt) and p_b(kΔt) respectively. The above discussion provides a way to estimate the bolus velocity and position

$$\widehat{v}\left(k\Delta t\right)=\widehat{v}(k\Delta t-\Delta t)+\mu \left(p_b\right(k\Delta t)-p_b(k\Delta t-\Delta t)-\widehat{v}(k\Delta t-\Delta t\left)\Delta t\right){\widehat{p}}_b(k\Delta t+\Delta t)=p_b\left(k\Delta t\right)+\widehat{v}\left(k\Delta t\right)\Delta t$$

It can be verified that this estimation algorithm has a nice convergence property, i.e.,

$$\mid {\widehat{p}}_b(k\Delta t+\Delta t)-p_b(k\Delta t+\Delta t)\mid \le \delta ∕\mu$$

and the estimation error is bounded by δ/μ, where μ is an adjustable gain. A large value of gain results in a small prediction error but at the same time could increase the sensitivity to noise.

2.3. Software and hardware in implementation

In the experiment, NI Labview (Version 7) was used. This software is widely commercially available. To test the proposed scheme, a prototype CT was constructed (Fig. 3). This prototype consists of four elements: a Master Flex Pump 7550-30, a movable table controlled by a Vexta alpha stepping motor AS46, a Pulnix-6700 camera and a PC (personal computer). The pump is controlled by the PC that simulates a person’s heart which drives the bolus through plastic tubings. The bolus velocity can be arbitrarily assigned by a computer program. In the experiment, the bolus peak position was implemented exactly the same as the actual bolus peak position extracted from the patient data sets. The stepping motor takes commands from the PC through a serial port. This simulates the patient table. The camera, connected to the PC by a PCI card, provides the real time bolus peak position that simulates the CT imaging device. The image acquisition and processing are implemented using NI IMAQ VISION DEVELOPMENT MODULES.

Fig. 3.

The prototype CT.

3. Results

The collected actual bolus positions and velocities from 4 patients were implemented on the prototype CT scanner with the proposed adaptive techniques and the current constant velocity CTA technology, respectively. For the constant velocity CTA, a constant velocity must be pre-determined before CT. In the experiment, a velocity of 30 mm per second was used which is typical in practice. Figures 4-7 show experimental results for these 4 patients. In all figures, the actual bolus peak positions are described by dash-dotted lines. The top diagrams are the CT table positions in solid lines based on the constant velocity technology. The bottom diagrams are the CT table positions in solid lines using the proposed adaptive scheme. The maximum and mean tracking errors for both the proposed adaptive scheme and the constant velocity method are provided (Table 2).

Fig. 4.

Experiment results based on the data of Patient #1.

Fig. 7.

Experiment results based on the data of Patient #4.

Table 2.

Tracking errors

	Constant velocity CT		Adaptive CT
Tracking errors	Maximum (CM)	Mean (CM)	Maximum (CM)	Mean (CM)
Patient #1	19.63	9.18	1.33	0.07
Patient #2	7.09	4.22	1.89	0.20
Patient #3	4.97	2.95	1.09	0.34
Patient #4	29.17	12.99	1.01	0.08

Two conclusions can be drawn from the simulations:

1. The adaptive model reasonably estimates and predicts the bolus propagation derived from actual bolus characteristics.

2. With the predicted future bolus peak position, the controlled table position and the actual bolus peak position are very close. In other words, synchronization between the bolus peak position and the imaging aperture has been achieved because of the adaptive scheme.

4. Discussions

The results can be improved in several ways.

First, the proposed adaptive model is so simple that allows effective online estimation and prediction. The model is based on a simple idea: if the sampling rate is high enough, the velocity of the bolus is almost the same between two consecutive sampling instants and therefore, it is much more efficient to describe the changes between two sampling instants instead of describing the complete dynamics. This eliminates the problem of estimating a large number of unknown parameters online. The experimental results clearly demonstrate the advantages. At the price of a small increase in model complexity, the performance can be further improved by incorporating extra variables. For instance, with available ECG readings for monitoring circulatory stages, the velocity variability due to systole and diastole can be better estimated and compensated. Also, the heart (specifically the left ventricle) serves as the energy source for blood propulsion. Important pump parameters include heart rate (H R), rhythm, stroke volume (SV), contractility ($\frac{dl}{dt}$), and cardiac pre-load and after-load. There are normal ranges for these variables according to age, height, weight and gender. However, these variables can vary markedly in the presence of cardiac disease. Accurate determination of these variables usually requires invasive monitoring, although some valuable data can also be non-invasively derived from echocardiography. With some prior information on the disease related parameters and ECG readings, it is possible to compensate for variability in cardiac pump parameters in the online adjustment of the model parameters. These kind models are referred to as the extended Hammerstein system (6) (Fig. 8). Use of the Hammerstein model in adaptive bolus chasing angiography will be explored.

Fig. 8.

The extended Hammerstein model for bolus-chasing CT angiography. parallel combination of multiple linear system blocks, only one of which is enabled at a given time. Input variables include bolus injection timing and rate, ECG signal and an index to one of pre-defined longitudinal body regions where the current bolus peak should appear. Output variables are the current bolus peak position and dispersion as well as their velocities.

In addition, in our experiments, the current bolus position and velocity were obtained by a camera. In reality, these will be estimated in real time which, as a result of needed imaging analysis, will result in inherent albeit small inaccuracies. Therefore, the adaptive scheme will have to be modified to compensate these errors.

In conclusion, the ability to accurately track the advancing bolus regardless of the cardiac functional parameters will have an enormous clinical value and our study demonstrated the technical feasibility of an adaptive bolus chasing strategy toward this goal in a realistic clinical environment. Our next step will be to refine the structure and parameters of the extended Hammerstein model to allow accurate and reliable tracking of variable, nonlinear, asymmetric and bilateral bolus propagation processes. The proposed adaptive control techniques will also be tested on a multi-slice CT scanner, with the compared end-points including improvement of the imaging quality and reduction in contrast dose.

Fig. 5.

Experiment results based on the data of Patient #2.

Fig. 6.

Experiment results based on the data of Patient #3.