See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/51397092

softMip: A Novel Projection Algorithm for Ultra-Low-Dose Computed

Tomography

Article in Journal of Computer Assisted Tomography · May 2008

DOI: 10.1097/RCT.0b013e31812e4b37 · Source: PubMed

CITATIONS READS

3 765

3 authors:

Henning Meyer Ralf Juran

Google Switzerland Charité Universitätsmedizin Berlin
37 PUBLICATIONS 255 CITATIONS 20 PUBLICATIONS 393 CITATIONS

SEE PROFILE SEE PROFILE

Patrik Rogalla
University Health Network
347 PUBLICATIONS 5,629 CITATIONS

SEE PROFILE

All content following this page was uploaded by Patrik Rogalla on 21 June 2019.

The user has requested enhancement of the downloaded file.

 ORIGINAL ARTICLE

softMip: A Novel Projection Algorithm for

Ultra-Low-Dose Computed Tomography
Henning Meyer, MD, Ralf Juran, Dipl. Phys, and Patrik Rogalla, MD

from that of conventional radiography.1 However, the 95%

Abstract: Two projection algorithms are currently available for
dose reduction is achieved at the expense of markedly
viewing computed tomography (CT) data sets: average projection
higher image noise. Multidetector CT enables very fast
(AVG) and maximum intensity projection (MIP). Although AVG
scanning with high resolution. A large amount of image
images feature good suppression of image noise but reduced edge
data is generated, for example, 500 to 600 images in an
sharpness, MIP images are characterized by good edge sharpness but
abdominal scan acquired during a single breath-hold. The
also amplify image noise. Ultra-low-dose (ULD) CT has very low
information content of these CT data sets is much higher
radiation exposure but has high image noise. Maximum intensity
than that of radiograph or ultrasound. However, the flood of
projection images of ULDCT data sets amplify image noise and are
images and the higher noise require a suitable postproces-
therefore unsuitable for image interpretation in the routine clinical
sing technique for image interpretation. Established techni-
setting. We developed a synthesis of both algorithms that tries to
ques for postprocessing CT source data sets are average
unite the respective advantages. The resulting softMip algorithm was
projection (AVG) and maximum intensity projection (MIP),
implemented in C++ and installed on a workstation. Depending on
which generate a few succinct images from a large set of
the settings used, softMip images can represent any graduation
thin-slice images (Fig. 1). The AVG algorithm effectively
between MIP and AVG. The new softMip algorithm was evaluated
suppresses noise in the thick slice generated from averaging
and compared with MIP and AVG in terms of image noise and edge
of a number of thin slices. However, small details are often
sharpness in a series of phantom experiments performed on 7
lost, and the radiologist interpreting AVG images must be
different CT scanners. Furthermore, image quality of the transition
aware of partial volume artifacts due to thick image slices.2
from AVG to MIP by means of softMip was compared with the
The well-known MIP algorithm was proposed for post-
image quality of simply blending AVG and MIP. Images generated
processing magnetic resonance angiography data sets.3,4 It
with softMip showed less image noise than MIP images (P G 0.0005)
selects the highest value along a stack of voxels as a
and higher edge sharpness than AVG images (P G 0.0005). The
resulting pixel. An asset of MIP is that it highlights
softMip transition from AVG to MIP had a better ratio of edge
radiopaque structures such as opacified vessels or calcifica-
sharpness and image noise than blending (P G 0.0005). Our results
tions. Therefore, an important area of application of MIP is
suggest that softMip is a very promising projection procedure for
CT angiography (CTA).5Y7 However, both algorithms are
postprocessing cross-sectional image data, especially ULDCT data
unsuitable for processing ULDCT dataVAVG because
sets.
small structures are lost in the thick slices generated, and
Key Words: ultra-low-dose computed tomography, image noise, MIP because it further amplifies the already high image
edge sharpness, maximum intensity projection, average projection, noise of ULDCT.
image filter
Aim of the Study
(J Comput Assist Tomogr 2008;32:480Y484) The aim of the study presented here was to develop a
projection algorithm for MDCT combining the advantages of
AVG and MIP, namely an improved image quality resulting
from suppression of noise and an improved identification of
U ltra-low-dose (ULD) computed tomography (CT) is a
protocol for multidetector CT (MDCT) with a markedly
reduced radiation dose compared with standard protocols.
pathology. The algorithm was then objectively evaluated in a
phantom experiment using edge sharpness and image noise as
In the clinical setting, ULDCT has established itself criteria.
especially for the detection of high-contrast objects. The
radiation exposure of ULDCT is not relevantly different MATERIALS AND METHODS
softMip
From the Institut für Radiologie, Charité-Universitätsmedizin Berlin,
Similar to MIP and AVG, in softMip, a pixel is
Germany. calculated from the Hounsfield unit (HU) values along a
Received for publication March 27, 2007; accepted May 25, 2007. corresponding viewing ray. These values constitute the
Reprints: Henning Meyer, Institut für Radiologie, Charité-Universitätsmedi- profile P of pixel p. Length l of P depends on the slice
zin Berlin, Charitéplatz 1, 10117 Berlin, Germany (e-mail: Henning. thickness s and projection geometryVin parallel projection
Meyer@charite.de).
Henning Meyer received a grant from Philips Medical Systems, Best, The l = s and in perspective projection l 9 s. What is crucial for
Netherlands. softMip is to incrementally order P, resulting in Ps with
Copyright * 2008 by Lippincott Williams & Wilkins Ps(0) = min (P) and Ps(l) = max (P). Pixel p is then calculated

480 J Comput Assist Tomogr & Volume 32, Number 3, May/June 2008

Copyright @ 2008 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
J Comput Assist Tomogr & Volume 32, Number 3, May/June 2008 softMip Y a Novel Projection Algorithm

FIGURE 1. Maximum intensity projection and AVG

algorithms.
FIGURE 3. Test phantom.
as the normalized weighted integral of the ordered profile
(Fig. 2): cylinder in which are arranged several blocks of Teflon,
1 l
x Plexiglas, and polyethylene, generating both high contrasts
p¼ l X0 f w IPs ðxÞdx between water and Teflon (0 HU vs ~990 HU), moderate
X0 f w ðxÞdx l contrasts between water and Plexiglas (~120 HU), and
negative contrasts with polyethylene (~j90 HU).
where fx(x) with 0 e x e 1 is the weighting function of Scanning experiments were performed on different CT
softMip, which strongly determines the result of softMip fw. scanners and an electron beam CT scanner8 to obtain results
For example, that are independent of the CT scanner. The scanners used are
  listed in Table 1. All experiments were performed using a
1; if x ¼ 1
f MIP
w ðxÞ ¼ standardized low-dose CT protocol (1-mm slice thickness, 3
0; else; effective mAs, 120 kV, 1.43 pitch) unless scanner limitations
will generate an MIP and fwAVG (x) = 1 an AVG. A softMip required adjustment of the parameters. From each CT source
image with noise suppression and MIP features is data set, coronal and sagittal projections with a slice thickness
generated by fw (x) = x4. The softMip algorithm was of 20 mm were generated as AVG, MIP, and softMip.
implemented using Visual Studio.NET (Microsoft Corpora-
tion, Redmond, Wash [http://msdn.microsoft.com/vstudio]) GradientFinder
and C++. The projection images generated with MIP, softMip,
and AVG were used to analyze the transitions between water
and the different blocks. Along these transitions, edge
Phantom Experiment sharpness and image noise were determined. We defined
A test phantom for constancy measurement of the edge sharpness as the gradient of an edge in the image from
Siemens Somatom DR Series (Siemens Medical Solutions, the 20% minimum to the 80% maximum because the increase
Siemens AG, Munich, Germany) was used for the phantom is nearly linear in this range. The gradient was determined by
experiments (Fig. 3). It consists of a water-filled Plexiglas calculating the profile of the respective projection as the mean

FIGURE 2. softMip algorithm.

* 2008 Lippincott Williams & Wilkins 481

Copyright @ 2008 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Meyer et al J Comput Assist Tomogr & Volume 32, Number 3, May/June 2008

TABLE 1. Computed Tomography Scanners Used for the Phantom Measurements

Scanner Type Vendor
Aquillion 4 4-slice MDCT Toshiba Medical Systems, Nasu, Japan
Aquillion 16 16-slice MDCT Toshiba Medical Systems
Aquillion 64 64-slice MDCT Toshiba Medical Systems
Tomoscan AV/EU Single-slice CT Philips Medizin Systeme GmbH, Hamburg, Germany
PQ5000 Single-slice CT Philips Medizin Systeme GmbH
Brilliance 16 16-slice MDCT Philips Medizin Systeme GmbH
Somatom Sensation 16 16-slice MDCT Siemens Medical Solutions, Erlangen, Germany
EBT C-150XP Electron beam CT GE Imatron, GE Medical Systems, San Francisco, Calif

of all lines of the image. Before SDs were calculated, the Statistics
profile was subtracted from all values within the transition. In Statistical analyses were performed using OpenOffice
this way, it is possible to determine noise at the edge. Calc 1.1.2 (OpenOffice.org [http://www.openoffice.org]) and
To ensure precise and objective measurement of SPSS 12.0 (SPSS Software GmbH, Munich, Germany [http://
these parameters, an analysis tool (GradientFinder) was www.spss.com/de/]). Statistical significance was assumed at
programmed in Ruby (http://www.ruby-lang.org). This pro- a probability of error of P G 0.05. The values given in the
gram uses a set of corresponding AVG, MIP, and softMIP diagrams are medians. The error bars correspond to the 25%
images to perform the measurements after all relevant and 75% quartiles.
transitions have been labeled.

MIP-AVG Blending RESULTS

To compare softMip images with a simple mixture of In preliminary tests, good results were achieved using
AVG and MIP, series of images can be generated in which an stepwise linear functions as weighting functions.
AVG image is blended into an MIP image of the same slice. The weighting function
This is done on the one hand by creating a smooth transition  x 
f sof tMip
ðxÞ ¼ 2 ; if x
0:5
from AVG to MIP by superimposing an MIP image on an w 1:5x  0:5; else;
AVG image of increasing intensity. Such a smooth transition
can also be created using softMip to superimpose the 2 was subjectively selected for a more detailed evaluation of
weighting functions, fwMIP and fwAVG. softMip.
In both series, edge sharpness and image noise can be We examined a total of 375 edges; minimum, median,
determined for each edge as described above. If the edge and maximum number per CT were 43, 46.5, and 47, respec-
sharpness and noise of the AVG image are set to 0 and those tively. Figure 4 presents the edge sharpness, and Figure 5
of the MIP image to 1, and the measured values are then presents the noise values determined in the phantom
normalized to these 2 values, the resulting values can be experiments. The results show that, for all CT scanners
represented in the form of a noise/edge sharpness diagram. used, noise is lowest in the AVG images (7.4) with slightly
The area under the curve (AUC) of the diagram can be higher noise values for softMip (9.0) and the highest noise in
interpreted as a measure of how well noise is suppressed MIP images (22.9). The distribution is similar for edge
and edges are enhanced in the respective blending series. sharpness: it is lowest for AVGs (4.5) and highest for MIP

FIGURE 4. Edge sharpness of the 3 projections. FIGURE 5. Noise of the 3 projections.

482 * 2008 Lippincott Williams & Wilkins

Copyright @ 2008 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
J Comput Assist Tomogr & Volume 32, Number 3, May/June 2008 softMip Y a Novel Projection Algorithm

FIGURE 8. Ultra-low-dose CT urography showing coronal MIP

(A) of the complete volume and softMip (B). Zoomed images
FIGURE 6. Blending curves for softMip and MIP-AVG transitions. of the dilated pylon are shown. Notice the lower image noise
and nice depiction of the contorted ureter in B.
(13.2), with the values for softMip in between (6.1). All
differences are statistically significant (P G 0.0005). while increasing contrast. Thus, softMip images enable better
To compare softMip blending with MIP-AVG evaluation of pathology. softMip is not a rigid projection
blending, a set of weighting functions was selected to algorithm because the features of the images generated
transform fwAVG via fwsoftMip to fwMIP: strongly depend on the weighting function fw selected. As a
  result, a large set of different projection images can be
vIf sof
w
tMip
ðxÞ þ ð1jvÞIf AVG
w ðxÞ; if 0 e v G1
f w ðx; vÞ ¼ generated, ranging from minMip, through AVG, to MIP and
ð2jvÞIf sof
w
tMip
ðxÞ þ ðvj1ÞIf MIP
w ðxÞ; if 1 e v e 2 including a variety of intermediate steps between these
This function moves from AVG v = 0 via softMip v = 1 projections. It is likely that the noise-sharpness ratio can be
to MIP v = 2. Figure 6 shows the edge sharpnessYnoise optimized further by using different weighting functions.
diagram of the median values of all CT scanners for both Moreover, the optimal ratio must also be adjusted for
types of blending. A large AUC indicates a transition with different clinical applications.
low image noise and high edge sharpness. A comparison of The MIP algorithm was originally developed for the
the AUCs is given in Figure 7. The median AUC is 0.41 for evaluation of magnetic resonance angiography3,4 but has
MIP-AVG blending and 0.56 for softMip blending. shown to be useful in searching for pulmonary nodules and
The differences between the individual scanners and for evaluation of the bronchial system as well.9 An important
the total were tested for statistical significances using Wilcoxon domain of MIP is CTA because the selective depiction of
test. The significance level was P G 0.0005 for all cases. dense voxels allows good evaluation of contrast-enhanced
Finally, Figure 8 shows a clinical application of vessels.6 Under specific conditions, intraluminal thrombi
softMip on a ULDCT urography. seen on axial thin slices may not be present on MIP images if
the vessel is depicted laterally.5,7 Because softMip, unlike
DISCUSSION MIP, also generates images from average values of the slices,
The results of this study show that postprocessing of such intraluminal thrombi may not be obscured on softMip
low-dose CT data sets with softMip can decrease image noise images.
Future Work
A first clinical study using softMip for processing
abdominal ULDCT data sets for the diagnosis of urinary
tract stones shows promising results and will be published
soon.
Computed tomography angiography might also benefit
from the improved image quality of softMip images
compared with MIP images. At the same time, the fact that
softMip uses average values might overcome the masking of
structures on MIP images by dense structures in front or
behind. This issue must be addressed in future studies.
Finally, further development of softMip comprises the
systematic investigation of different weighting functions in
different applications. It is also conceivable to perform
automatic computer-based optimization of the weighting
FIGURE 7. Areas under the blending curves. function by means of phantom scans, resulting in a

* 2008 Lippincott Williams & Wilkins 483

Copyright @ 2008 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
 Meyer et al J Comput Assist Tomogr & Volume 32, Number 3, May/June 2008

preselectable compromise between edge sharpness and 5. Prokop M, Shin HO, Schanz A, et al. Use of maximum intensity
image noise. projections in CT angiography: a basic review. Radiographics.
1997;17:433Y451.
6. Napel S, Marks MP, Rubin GD, et al. CT angiography with spiral CT and
REFERENCES maximum intensity projection. Radiology. 1992;185:607Y610.
1. Rogalla P, Klüner C, Taupitz M. Ultra-low-dose CT to search for stones in 7. Marten K, Funke M, Obenauer S, et al. The contribution of different
kidneys and collecting system. Aktuelle Urol. 2004;35:307Y309. postprocessing methods for multislice spiral CT in acute pulmonary
2. Zou Y, Sidky EY, Pan X. Partial volume and aliasing artefacts in helical embolism. Rofo. 2003;175:635Y639.
cone-beam CT. Phys Med Biol. 2004;49:2365Y2375. 8. Becker CR, Schätzl M, Schoepf UJ, et al. Technical bases and acquisition
3. Laub GA, Kaiser WA. MR angiography with gradient motion refocusing. conditions of electron-beam computed tomography. Radiologe.
J Comput Assist Tomogr. 1988;12:377Y382. 1998;38:987Y992.
4. Keller PJ, Drayer BP, Fram EK, et al. MR angiography with 9. Napel SL, Bergin C, Paranjpe D, et al. Maximum and minimum intensity
two-dimensional acquisition and three-dimensional display. Work in projection of spiral CT data for simultaneous 3D imaging of the
progress. Radiology. 1989;173:527Y532. pulmonary vasculature and airways. Radiology. 1992;185:126.

ERRATUM
An error appeared in an article by Pollock et al, in the March/April 2008 issue of Journal of Computer Assisted
Tomography (J Comput Assist Tomogr. 2008;32(2):291Y293). A footnote on page 291 states BThere are no conflicts of
interest, and no financial support was provided. This case has not been otherwise published.[

The correct footnote is as follows:

There are no conflicts of interest. Research support was provided by R01EB004673. This case has not been otherwise
published.
This error has been corrected in the online version of the article available at www.jcat.org
The publisher regrets the error and any inconvenience it has caused.
1. Pollock JM, Deibler AR, West TG, et al. Arterial spin-labeled magnetic resonance imaging in hyperperfused seizure
focus: A case report. J Comput Assist Tomogr. 2008;32(2):291Y292.

484 * 2008 Lippincott Williams & Wilkins

View publication stats

Copyright @ 2008 Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.