3 Strategies for Dose Reduction
and Improved Image Quality in MSCT
M. Kachelriess, S. Schaller, W. A. Kalender
M. Kachelriess, PhD
Institute of Medical Physics, Krankenhausstrasse 12, 91054 Erlangen, Germany
W. A. Kalender, PhD
Professor, Institute of Medical Physics, Krankenhausstrasse 12, 91054 Erlangen, Germany
S. Schaller, PhD
Siemens Medical Solutions, Siemensstrasse 1, 91301 Forchheim, Germany
CONTENTS
3.1 Introduction 35
3.2 Dose and Image Quality 36
3.3 Dose Reduction with Multidimensional Adaptive Filtering 36
3.4 Dose Reduction with Automatic Exposure
Control 38
3.5 Dose Reduction by User Education 40 3.6 Tube Voltage and Filtration 41 3.7 Penumbra Effects 42 3.8 Dose Information 43 3.9 Discussion 44
References 45
3.1 Introduction
Since the introduction of the fi rst computed tomog- raphy (CT) scanner in 1972, the importance of CT for the medical community has increased drastically.
Presently, whole-body scans with isotropic sub-mil- limetre resolution are acquired routinely during a single breath-hold.
The decisive step towards three-dimensional isotropic data was the introduction of spiral CT in 1989 by W. A. Kalender. This scan mode is based on a continuous rotation of the gantry while simul- taneously translating the patient along the axis of rotation. The scan trajectory is a spiral and allows for truly 3D data acquisition. The z-interpolation step allows one to select the longitudinal position (z-position) of the reconstructed images arbitrarily and retrospectively. The continuous axial sampling
allows for high-quality 3D displays and led to a renaissance of CT (Kalender 2001). The introduc- tion of multi-slice spiral CT (MSCT) in 1998, which allows simultaneous scanning of currently up to 16 slices, further improved the scanner’s volume coverage, z-resolution and scan speed. For example, typical chest exams are carried out with collimations of 1 ×5 mm in 36 s with single-slice, 4×1 mm in 30 s with 4-slice and 16 ×0.75 mm in 10 s with 16-slice scanners (Fig. 3.1).
Recently, patient dose has become an increasingly more important issue: increasing the information density of a scan, e.g. the spatial resolution, requires to increase the photon density (absorbed photons per volume element) by at least the same factor if the signal-to-noise ratio (SNR) is to be maintained. Cur- rently, less than 5% of all X-ray exams are performed with CT. The cumulative dose of CT, however, is of the order of 40% of all X-ray exams. Dose is espe- cially critical in paediatric CT (Brenner et al. 2001).
Practical hardware- and software-based solutions to reduce patient dose are strongly required.
Fig. 3.1. Three generations of spiral CT scanners. The sizes of
the detector elements indicate the slice thickness of single-
slice, 4-slice and 16-slice scanners for typical exams. These
slice dimensions (or “logical” detector dimensions) may differ
from the physical dimensions since detector slices are often
made up by electronically combining several detector rows
This paper gives an overview of CT dose issues and of strategies to reduce the patient dose or, alter- natively, to improve the image quality for a given dose level. Note that in all attempts to reduce dose, image quality is crucial since the diagnostic outcome must not be jeopardized.
3.2 Dose and Image Quality
Image quality is a function of dose. We can quantify image quality in CT mainly by giving two parameters:
(a) the pixel noise σ that is measured as the standard deviation of the pixel values within a homogeneous image region; and (b) the spatial resolution. Due to the circular symmetry of CT scans one usually dis- tinguishes between the in-plane resolution ∆r and the axial- or z-resolution S eff (effective slice thickness). For example, an image noise of σ = 30 HU, an in-plane reso- lution of ∆r=0.7 mm and a z-resolution of S eff =0.8 mm are typical values for the present standard exams. Please note that the resolution values are not identical to the voxel size of the image data which may be chosen to be arbitrarily small during image reconstruction.
Dose is usually quantifi ed by specifying the organ dose or the effective dose (values in mSv). For our pur- poses, it suffi ces to regard the effective tube current time product (I·t) eff , i.e. the “effective mAs value”, as being the characteristic dose parameter since this parameter is proportional to the physical dose values. It is defi ned as the product of the tube current with the time under which a given z-position (slice) is exposed to X-rays;
therefore, the effective mAs depends on the scan’s pitch value p and on the rotation time t rot as follows:
Recall that the pitch value is defi ned as
where d is the table increment per rotation, S is the collimated (nominal) slice thickness and M is the number of simultaneously scanned slices. Its inverse p –1 gives the number of rotations that cover a given z-position. Consider a scan with a table increment of 6 mm and a collimation of 16 ×0.75 mm. This situa- tion corresponds to a pitch of 0.5 and, consequently, each z-position is exposed to X-rays during two full p = M . S
d
rotations (twofold overlapping scan). Further assume a rotation time of 0.4 s and a tube current of 200 mA.
Then, according to the defi nition above, the effective mAs value is 160 mAs. It is this effective mAs value, and not the tube current itself, that is directly related to the image noise σ. The present scanners make use of this fact by allowing the user to specify the effec- tive mAs value. Internally, the scanner computes the correct tube current from the specifi ed effective mAs value and the selected pitch. Consequently, dose and image noise and the effective mAs value do not depend on the spiral pitch, a fact that has often been misunderstood.
Quantitatively, the proportionality
(1)
holds. The constant of proportionality depends in a complicated way on the patient size and density, on the tube voltage and on the prefi ltration.
Prior to the scan, the user specifi es the three parameters effective mAs value, effective slice thick- ness and the in-plane resolution (via the name of the reconstruction kernel) at the scanner console. He is thereby aiming at his preferred image quality. The image noise σ is the only dependent parameter since it cannot be controlled directly.
Equation has important consequences: increasing the spatial resolution requires to dramatically increase the effective mAs value and thereby the dose given that the image noise σ is to be held constant. For example, doubling spatial resolution (i.e. reducing S eff and ∆r to 50%) requires to increase patient dose by a factor of 16 if the pixel noise shall remain at its original level. Since such dose increases are hardly acceptable, scan proto- cols must carefully balance between spatial resolution and image noise; therefore, many applications are designed to either aim at good low-contrast resolution (low image noise) or to produce high spatial resolution.
The fi rst is the case in tumour detection tasks where lower spatial resolution is acceptable. In contrast, lung exams and, especially, exams of the inner ear, require to scan with the highest spatial resolution available.
3.3 Dose Reduction with Multidimensional Adaptive Filtering
To allow for even higher image quality in the future, new software- and hardware-based methods that σ 2 ∝
(I . t) eff . S eff . ∆r 3 1
(I . t) eff = I 0 . t rot
p .
make better use of the applied dose have been fur- nished.
This includes so-called adaptive fi ltering approaches that seek to reduce image noise either by directly manipulating the reconstructed images or by perform- ing dedicated raw-data preprocessing prior to image reconstruction (Eklundh and Rosenfeld 1981;
Lauro et al. 1990; Keselbrener et al. 1992; Hsieh 1994, 1998) . Directly manipulating images turns out to be less effi cient and more critical, even if the fi lters are applied locally (adaptively), since respective algo- rithms tend to smooth parts of the image; therefore, raw-data-based approaches, as they are often provided by the manufacturers, are preferred.
A highly effective method is multi-dimensional adaptive fi ltering (MAF) which is a raw-data-based algorithm using two detector dimensions plus the view dimension (rotation direction) to locally smooth data.
The MAF allows to reduce image noise in cross sections of high eccentricity such as the pelvis, thorax or shoul- der without affecting spatial resolution. In contrast to the said image-based smoothing approaches (includ-
ing the selection of smooth reconstruction kernels), the raw-data-based MAF operates only on a tiny frac- tion of the acquired raw data. The MAF selects those attenuation values for fi ltering that suffer from photon starvation due to high attenuation. Typically, 5% of the data are modifi ed; 95% of the acquired data remain unchanged (Kachelriess and Kalender 2000;
Kachelriess et al. 2001b).
The MAF can be used to remove the well-known streak-like noise artefacts that are oriented along the direction of highest attenuation (typically the lateral direction). Low-contrast objects are often completely hidden by this kind of structured and correlated noise. Increasing the tube current to avoid these artefacts would be the wrong strategy since global improvements are not required. The effect of apply- ing MAF is illustrated in Fig. 3.2. The patient exhibits severe streak artefacts in the shoulder region due to photon starvation. They are largely removed by MAF. Even more, the difference images show no patient structure. This proves that MAF does not simply smooth the image. The MAF instead man-
Fig. 3.2. Noise in the original data (top row) is reduced by up to 50% with multi-dimensional adaptive fi ltering (MAF; middle
row). Since streaks are oriented laterally, the transaxial slice and the coronal MPR show great improvements, whereas there are
no modifi cations in the anteroposterior direction and, consequently, the sagittal reformation is hardly modifi ed. The subtraction
images (bottom row) prove that MAF does remove noise but not structure details
ages to reduce noise and to remove the correlated noise without affecting spatial resolution; thereby, the visibility of low-contrast objects can be greatly improved (Baum et al. 2000).
3.4 Dose Reduction with Automatic Exposure Control
Similar to conventional radiography, an automatism to obtain a defi ned image quality is needed for CT. The present CT scanners allow to specify the spatial resolu- tion, but it is not possible to specify the image noise.
One must instead set the effective mAs product which infl uences image noise only indirectly. The outcome is strongly dependent on patient size, density and shape.
The dependence on patient size and its implica- tions on dose values are best illustrated using a numerical example: a 20 cm water-equivalent object (representing an adult head) attenuates the X-rays by a factor of approximately 50 and only 2% of the pho- tons reach the detector. For a small object of half size (representing a child head), the attenuation factor decreases to 7 and it becomes likely that too many quanta reach the detector. Due to this exponential behaviour, it is hard to manually extrapolate scan parameters from the standard adult patient to small children. This fact is one of the major causes for using too high doses in paediatric CT.
A fi rst step towards AEC has been the introduc- tion of the dose modulation technique. It modulates the tube current during each rotation as a function of patient anatomy to compensate for the different attenuation properties; typically, for anteroposterior projections the tube current is decreased and for lat- eral projections the tube current is increased (Gies et al. 1999; Kalender et al. 1999b).
It can be shown that the optimal way to modulate the tube current as a function of the rotation angle α is proportional to the square root of the attenuation p( α) measured at that angle (Gies et al. 1999; Harpen 1999) :
To achieve optimal modulation, attenuation values must be known. The most effective (in terms of dose utilization) implementations use attenuation values that have been measured 180° in advance to control the tube at the current projection on-line. Figure 3.3 illustrates this principle by giving typical quantum number values. Typical mAs-reduction values that can be achieved with this on-line dose modulation technique are given in Table 3.1, taken from Kalen- der et al. (1999c) and Greess et al. (2000).
These mAs-reduction values have been achieved without increasing image noise. It further turns out that the noise structure of the images acquired with dose modulation is less pronounced and less corre- lated than the noise structure of the standard acquisi- tion, thereby indicating improved image quality. Even more, Monte Carlo calculations of the organ dose and effective dose with and without dose modulation have shown that the mAs-reduction values under- estimate the physical dose reduction (Kalender et I 0 (α) ∝ e 1 2 p(α) .
Fig. 3.3. Scans with constant tube current (red) suffer from extremely different attenuation properties within eccentric objects. Here, photon starvation occurs in the lateral direction while more than suffi ciently many photons are measured anteroposteriorly. Tube current modula- tion (green) greatly reduces these dis- crepancies by modulating the tube cur- rent as a function of anatomy. In total, it reduces the mAs product (here by 30%) and, consequently, patient dose without increasing image noise
Table 3.1. Typical mAs-reduction values that can be achieved with on-line dose modulation technique
Region mAs reduction (%)
Head 18
Shoulder 53
Thorax 22
Abdomen 15
Pelvis 25
Extremity 39
al. 1999c; Schmidt and Kalender 2001, 2002): the values given in the table are only a conservative esti- mate of the true dose reduction achieved with tube current modulation.
Besides the on-line dose modulation techniques, other implementations make use of topogram (or scout view) data from an anteroposterior and a lateral view (Kopka et al. 1995). Sinusoidal curves are fi tted to the topogram data to predict the angle-dependent attenuation for all α; however, besides the need for two topograms, the mAs-reduction values reported are signifi cantly lower than with the on-line method (Kalender et al. 1999c).
All dose-modulation methods have in common that there are only local (rotation-wise) adaptations of the current. Variations of the mean attenuation properties in the longitudinal direction are not taken into account as the tube current’s mean or maximum value is retained after each rotation.
The remaining step towards AEC is to combine the tube current modulation with a tube current control.
The tube current control slowly varies the current’s mean value according to the patient anatomy and thus is able to compensate for varying image noise along the z-direction. The AEC comprises a method to “foresee” the image noise as a function of anatomy.
Then, the tube current can be controlled to achieve a given target noise. Again, there are topogram-based implementations and more sophisticated raw-data- based implementations (Kachelriess et al. 2001a).
In the future, scanners will make use of AEC to assist the user in adapting the tube current to patient anat- omy or even to replace the tube current parameter in favour of directly specifying image noise and hence image quality.
Figure 3.4 shows the tube current curve and the resulting image noise as a function of the anatomi- cal level during a spiral 16-slice scan from neck to thorax. The red and blue graphs correspond to the conventional case: the mean level of the tube current is held constant. The noise curves (red is hidden by blue) show excellent image quality in the neck and insuffi cient image quality in the shoulder. Without AEC, image quality can only be increased by increas- ing the tube current. With AEC, in contrast, it is pos- sible to accept more noise in the neck and to reduce noise in the thorax at the same time. Assume that the user specifi es a constant noise level corresponding to the green horizontal curve. Then, AEC controls and modulates the tube current to achieve the desired level (green oscillating curve). The running mean (green slowly varying curve) gives the local mAs- reduction value as a function of the z-position.
The phantom experiment of Fig. 3.5 demonstrates the performance of AEC. The step phantom consists of three elliptical cylinders of decreasing diameter.
The standard acquisition (constant tube current) suf- fers from varying image quality along the z-axis: high noise in the large cylinder; acceptable noise in the medium cylinder; and unnecessary low noise in the small cylinder. The three axial images further suffer from streak artefacts due to correlated noise. The AEC scan, in contrast, is as specifi ed: constant noise for the complete scan range and reduced noise correlations in the axial images. Besides constant noise, one may also specify other noise profi les, if desired.
Fig. 3.4. Tube current (top) and resulting image noise (bottom) for a scan from neck to shoulder. Red: constant tube current (conventional scan). Blue: tube current modulation. Green: tube current control and modulation (AEC). The thick lines in the top image are the local averages of the modulated graphs and illustrate the dose savings as a function of the z-position. The red graph is hidden by the blue graph in the bottom image
Fig. 3.5. This phantom experiment demonstrates that AEC
achieves constant image noise for all z-positions and that
noise structure can be signifi cantly reduced
Even more potential for dose reduction lies in the combination of AEC and MAF (Kachelriess et al.
2002) . Both methods are complementary: AEC glob- ally infl uences complete projections and MAF is able to locally adjust the noise statistics of single detec- tor readouts. The synergetic effects are illustrated in Fig. 3.6. The AEC reduces the tube current at the lung level where the attenuation is quite low. The increased noise is acceptable. In the shoulder, however, the tube current is globally increased by AEC. Here, MAF fur- ther reduces noise in the most central lateral rays.
An interesting specialization to the tube current modulation and tube current control technique has been recently introduced to reduce dose of retrospec- tively gated cardiac CT exams (Kachelriess and Kalender 1998) : The so-called ECG pulsing tech- nique reduces the tube current for those ECG phases that are unlikely to contribute to the reconstructed images (Jakobs et al. 2002).
3.5 Dose Reduction by User Education
When installed at a location, the CT system is usually adjusted to standard factory settings to give good image
quality. This does not automatically mean the lowest possible radiation dose. Learning how to customize and optimize the scan parameters (which constitute the scan protocols) with regard to dose and image quality is therefore a promising strategy to reduce dose.
However, customizing scan protocols to optimally meet all criteria is a diffi cult task. This is due to the complex dependence of image noise on the patient cross section (including size, weight, etc.). Finding the optimal mAs value, for example, is an iterative process that requires profound experience and, in principle, the possibility to repeat scans with different param- eters until the preferred image quality is achieved.
Dose considerations, however, forbid to repeat scans on a given subject. Phantom or cadaver scans may help, but the results are often too far from reality.
Recently, industry has started to provide tools that retrospectively simulate arbitrary scan parameters (Leidecker et al. 2002). Basically, these tools consist of simulating arbitrary tube currents or, equivalently, mAs products. The resolution can be varied retro- spectively by selecting different reconstruction ker- nels and different effective slice thicknesses.
Two classes of simulation methods are available.
One approach adds random noise to reconstructed CT images. The standard deviation of the random values is a function of the voxel grey value, the source
Fig. 3.6. The combination of AEC with MAF currently is the optimum dose-reduction strategy: due to its tube current modula-
tion and control, AEC increases noise in regions of low attenuation (arrowheads) and decreases noise and noise correlation in
regions of high attenuation and high eccentricity. In contrast to tube current modulation techniques that infl uence the photon
statistics of complete projections, MAF can modify single-detector element readings and further reduce the noise in the shoulder
and pelvis region (arrows). The difference in image indicates that no structure has been smoothed by the procedure
mAs and the desired destination mAs; however, this simple method does not correctly consider the X- ray physics which demand to add noise to the X-ray photon attenuation instead of adding noise to image pixels. Consequently, it cannot reproduce noise struc- ture and artefacts characteristic to CT images.
The second class of algorithms correctly simulates the physics behind CT imaging. To do so, it requires CT raw data and the complete image reconstruction pipe- line. Noise is added to the projection values to account for the desired destination mAs product. Figure 3.7 shows respective results using a raw-data-based tube- current simulator. As expected, image noise increases with decreased tube current and the noise structure generated is physically correct.
3.6 Tube Voltage and Filtration
As we have seen, there is much potential in optimiz- ing the dose usage by optimizing the tube current (or, equivalently, the effective mAs value). What remains to be considered are the effects of the tube voltage U and of the prefi ltration on image quality and on patient dose. It is well known that a high tube voltage yields a high-energy X-ray spec- trum and that the X-ray attenuation coeffi cients µ decrease with increased energy. In other words:
fewer photons are absorbed in the patient. The dose itself scales with both the number and the mean energy of absorbed photons. The number of pho- tons available is proportional to the tube current but is a complicated function of U.
Although these relationships are already complex enough, we have not yet included image quality into these considerations. Since the attenuation coeffi - cients vary with the tube voltage, the reconstructed grey values will also do so ( µ is a function of U);
therefore, it does not suffi ce to regard image noise σ alone as the indicator of image quality. Equation (1)
must instead be extended to put the noise σ and the signal µ into relation:
Here, D is the object diameter and the numerator accounts for the total object attenuation. Low values of ( σ/µ) are desired. Its inverse µ/σ is commonly known as the signal-to-noise ratio (whenever signal differences ∆µ are of interest the SNR is given as
∆µ/σ). Image quality increases with increasing SNR.
The complex dependence between the SNR and the tube voltage cannot be handled analytically, however.
Monte Carlo simulations or phantom experiments are required to analyse the effect of tube voltage on image quality and dose.
For example, a recent study evaluated iodine contrast as a function of tube voltage to improve the scan protocol design for paediatric CT (Schaller et al. 2001). Iodine contrast ∆µ = µ I – µ H
2O and image noise were measured in a thorax phantom equipped with an iodine insert for 80, 120 and 140 kV. Dose was estimated using the CT dose index (CTDI) measured in a 16 cm CTDI phantom. The CTDI is defi ned as the integral over the axial dose profi le normalized to the nominal collimation:
The study resulted in the following values (Table 3.2):
image quality, given by the SNR, must be normalized with the dose value: Quality=SNR 2 /Dose serves as an adequate image quality measure. The Table 3.2 clearly shows the strong dependency of image quality on the tube voltage: the quality of the 80 kV scan differs from that of the 140 kV scan by a factor 2.9. This means that the SNR at 80 kV is 70% better than at 140 kV when the same dose is applied. Alternatively, one can obtain the same image quality at 80 kV with only 34% of the dose.
Fig. 3.7. The dose tutor helps to decrease the tube current from its original level to the desired low-dose protocol (left: original 160 mAs, right: simulated 40 mAs). Noise structure is correctly reproduced. The user can experience with realistic examples which tube current is acceptable
(σ/µ) 2 ∝
µ 2 (I . t) eff . S eff . ∆r 3 e µD
CTDI 100mm = dz D (z) M . S
1
50mm
–50mm冕