Key words brain - CT - image manipulation/reconstruction
Introduction
Computed tomography (CT) of the brain is one of the most frequently performed radiological
examinations in hospitals with emergency rooms or neurological wards. During the history
of CT so far, image quality of brain CT scans has always been a topic of research.
The anatomy of the neurocranium gives rise to particular challenges when performing
CT [1 ]
[2 ]
[3 ]: firstly, highest soft-tissue resolution is required to display the normal differentiation
of gray matter and white matter, which are very close in X-ray attenuation. This is
of special interest in stroke imaging, since slight cortical or basal ganglia hypoattenuation
(isodensity to the adjacent white matter) and so-called sulcal effacement are well
known to be early CT signs of definite brain tissue damage in ischemic stroke [4 ]. Secondly, there is always a diagnostic uncertainty regarding the posterior fossa
because of skull base-related beam hardening artifacts, which can only partly be overcome
by using spiral acquisition techniques, sometimes at the expense of contrast resolution
[1 ]
[5 ]
[6 ]
[7 ].
All major vendors of CT systems have recently introduced iterative reconstruction
(IR) algorithms as a product. Since image noise is reduced compared to the traditional
filtered back projection (FBP) reconstruction method, IR can be used in the clinical
setting with two different major aims: firstly, to achieve reduction of patient dose
exposure by reducing the exposure settings; secondly, to achieve a possible improvement
in image quality by reduction of image noise and beam hardening artifacts [8 ]
[9 ]
[10 ]
[11 ]
[12 ]
[13 ]
[14 ]
[15 ]
[16 ]
[17 ]
[18 ]. Other possible fields of applying IR include e. g. reduction of metal artifacts
or scatter compensation, which may become even more important in the future [19 ]
[20 ].
Some recent studies have sought to demonstrate a potential effect of IR techniques
in brain CT [8 ]
[17 ]
[21 ]. The intention of this study is to present an intraindividual comparison of different
settings of the 4th generation IR technique iDose™ (Philips Healthcare, Best, The Netherlands) in combination
with different filter kernels without having to perform additional scans in the clinical
setting. Conventional filtered back projection reconstructions serve as the standard
of reference.
Compared to earlier generations of IR tools and artifact reduction algorithms, iDose™
reveals a noise power spectrum, which is very close to that of FBP [22 ]. Therefore reconstructions applying iDose™ can be expected to have an image appearance
that is familiar to clinicians. An additional parameter is the iDose™ level, which
ranges from 1 – 7 and is used to define the strength of the iterative reconstruction
technique in reducing image quantum mottle noise (range: 11 – 55 % noise reduction
relative to a corresponding FBP reconstruction) [22 ]. The level can be defined independently from the radiation dose at which an acquisition
is performed.
Materials and Methods
The local institutional review board approved this study (WF-003/12). The requirement
for written informed consent was waived.
Data acquisition and reconstruction
Raw data of brain CT acquisitions of 31 consecutive patients (13 male, 18 female;
median age 63 years; age range: 21 – 97 years) were collected and used for this study.
All exams were performed in the department’s standard sequential mode on a Brilliance
iCT 256™ scanner (Philips Healthcare, Best, The Netherlands). A summary of the acquisition
parameters is listed in [Table 1 ]. All exams were performed during the clinical routine and therefore always initially
reconstructed with FBP according to the standard parameters of the department; followed
by regular radiological reporting.
Table 1
Summary of key CT acquisition parameters. The slice collimation during sequential
acquisition was 16 × 0.625 mm. The reconstructed slice thickness was 2.5 mm infratentorial
and 5 mm supratentorial.
Tab. 1 Darstellung der wesentlichen CT-Akquisitionsparameter. Die Schichtkollimation bei
der inkrementellen Akquisition war 16 × 0,625 mm; rekonstruierte Schichtdicken waren
2,5 mm infratentoriell und 5 mm supratentoriell.
kV
mAs
mean scan length (mm) ± SD
mean CTDIvol (mGy) ± SD
mean DLP (mGy × cm) ± SD
mean effective dose (mSv) ± SD
infratentorial
120
333
50.7 ± 4.4
61.5 ± 0
311.5 ± 27.2
0.65 ± 0.06
supratentorial
120
310
97.4 ± 7.3
57.2 ± 0
557.2 ± 41.7
1.17 ± 0.09
Raw image data for study evaluation were transferred to a prototype reconstruction
processor featuring iDose4 ™ software (Philips Healthcare, Best, The Netherlands) in order to perform the new
iterative reconstructions using the following settings: iDose™ level 0 (= filtered
back projection), 1, 3 and 4, respectively. With this selection the percentage change
in image noise from one level to the next is almost the same (about 12 % on average).
Additionally for each setting 3 different brain filter kernels (UA = smooth, UB = standard,
UC = sharp) were applied, which resulted in 12 different stacks of CT images for each
patient. The slice thickness was 2.5 mm for the infratentorial space and 5 mm supratentorially,
according to the department’s standard. The department’s default reconstruction settings
were FBP with standard filter kernel UB.
Subjective image evaluation
Two experienced radiologists (J. B. and M. R., > 6 years experience) and three less
experienced radiologists (A. L., H. S. and D. H. < 2 years experience in clinical
radiology or no experience in reading brain CT) performed scoring of the image data.
The rationale for having different levels of experience was to find out if confirmed
habits might influence the perception of the IR reconstructions and hence the subjective
scores.
Scoring was performed by applying a precisely defined 4-point scale with regard to
artifact load, gray-white matter differentiation and overall image impression (1 = worst,
4 = best), in analogy to former work on the same topic [9 ]
[12 ]
[23 ]. The particular criteria for image evaluation as used in the current study are summarized
in [Table 2 ]. Image viewing was performed on a PACS-Workstation (PACS IW, GE Healthcare, Milwaukee,
MI). For scoring, the images of every reconstructed dataset were displayed anonymously
in the department’s standard window settings (infratentorial: WW = 90, WL = 30; supratentorial:
WW = 60, WL = 35) and under reporting conditions. The raters were allowed to scroll
through the complete dataset, supra- and infratentorial sections have not been evaluated
separately.
Table 2
Definition of subjective image evaluation criteria.
Tab. 2 Definition der subjektiven Kriterien der Bildbewertung.
score
rating
Criteria
1
non-diagnostic
excessive image noise and hardening artifacts
delineation of gray and white matter mostly impossible
2
suboptimal quality, barely diagnostic
substantial image noise and hardening artifacts
delineation of gray and white matter feasible
3
average diagnostic quality
some image noise and hardening artifacts
ordinary delineation of gray and white matter
4
excellent diagnostic quality
very little image noise, no hardening artifacts
easy delineation of gray and white matter
Objective image evaluation
For the evaluation of the absolute contrast of gray and white matter, 5 pairs of small
rectangular regions of interest (ROI, size: 1.2 × 1.2 mm = 3 × 3 px) in adjacent cortical
gray matter and adjacent white matter were defined in each dataset (see fig. 3 for
ROI example) using maximum zooming and copied likewise to every reconstruction in
order to measure mean CT values (Hounsfield Units, HU). Contrast C is defined as the
difference of CT values in the small ROIs referring to gray and white matter. For
the evaluation of image noise, one pair of larger rectangular ROIs (size: 6.1 × 6.1 mm = 14 × 14 px)
in almost homogeneous areas of cortical gray matter and white matter was defined and
copied likewise to every reconstruction in each dataset in order to measure the objective
image noise. Image noise N is represented by the average of the standard deviation
(SD) of HU in both larger ROIs. The contrast-to-noise ratio (CNR) for each dataset
was calculated from the contrast measured in the small ROIs and image noise measured
in the larger ROIs. Means for C, N and CNR over all data sets were calculated. The
data are presented in [Table 3, ]
[Fig. 3 ].
Table 3
Results of the ROI-based contrast and noise measurements, see also [Fig. 3 ] for graphical illustration of the results and location of the ROIs.
Tab. 3 Ergebnisse der ROI-Messung von Kontrast und Bildrauschen, siehe auch graphische Darstellung
der Ergebnisse und Lage der ROIs in [Abb. 3 ].
filter kernel
iDose level
UA
(brain smooth)
UB
(brain standard)
UC
(brain sharp)
gray-white contrast (HU) ± SD
0
11.0 ± 0.6
11.9 ± 0.5
14.4 ± 0.7
1
9.9 ± 0.6
10.9 ± 0.5
13.3 ± 0.7
3
8.9 ± 0.5
9.9 ± 0.4
12.0 ± 0.6
4
8.2 ± 0.5
9.2 ± 0.5
11.3 ± 0.5
image noise (HU) ± SD
0
1.9 ± 0.1
2.3 ± 0.1
3.4 ± 0.1
1
1.8 ± 0.1
2.1 ± 0.1
3.0 ± 0.1
3
1.4 ± 0.1
1.7 ± 0.1
2.6 ± 0.1
4
1.3 ± 0.1
1.6 ± 0.1
2.3 ± 0.1
contrast-to-noise ratio ± SD
0
5.9 ± 0.3
5.3 ± 0.2
4.3 ± 0.1
1
5.7 ± 0.4
5.1 ± 0.3
4.4 ± 0.1
3
6.3 ± 0.5
5.7 ± 0.5
4.7 ± 0.1
4
6.6 ± 0.7
5.9 ± 0.5
4.8 ± 0.1
Statistical analysis
Statistical analysis of the subjective scorings was carried out by applying a random
intercept model. All calculations were performed in SPSS 20.0 (SPSS Inc., Chicago,
IL). Patients and raters were defined as random intercept, raters were nested within
their experience. The following parameters were defined as fixed effects: sex, IR
level setting, filter kernel, interaction between IR level setting and filter kernel.
The intraclass-correlation coefficient (ICC) was calculated to describe the inter-rater
reliability.
Results
The consecutive collection of datasets represents a typical mixture of findings. In
14 subjects (45 %) no pathological findings occurred. White matter lesions as signs
of microangiopathy were frequently observed (10 subjects, 32 %, thereof 6 female),
3 subjects had residuals of territorial stroke (no acute stroke patient in the collective),
3 intracranial bleedings occurred, 1 subject presented with intracranial metastases.
Illustrating images of typical supratentorial reconstructions in a male subject with
small lacunar infarction at the level of the right basal ganglia are presented in
[Fig. 1a ]; additionally the subjective scores are annotated. It has to be noted that the differences
appear visually more evident when directly compared on a reading screen of the PACS
workstation. Additional illustrating images show a typical infratentorial slice with
constant skull base artifacts ([Fig. 1b ]) and a supratentorial slice with hemorrhagic metastases ([Fig. 1c ]).
Fig. 1 Presented in a is a representative set of reconstructions of a supratentorial slice in a 69-year-old
male patient with a small lacunar infarction on the level of the right basal ganglia
(WL/WW = 35/60). The IR level increases from left to right, iDose 0 equals filtered
back projection (FBP). Numbers in indexes represent the mean subjective score of the
respective reconstruction for this particular case that can vary significantly from
case to case. b shows a similar collection of reconstructions of an infratentorial slice in a 55-year-old
female patient (WL/WW = 30/90) illustrating the unchanged skull base artifacts regardless
of the iDose level. c shows a collection of reconstructions of a supratentorial slice in a 59-year-old
male patient suffering from multiple hemorrhagic cerebral metastases (WL/WW = 35/60)
illustrating the influence of iDose level and filter kernel on image noise. Lesion
conspicuity does not benefit from higher iDose levels.
Abb. 1 a zeigt repräsentative Rekonstruktionen einer supratentoriellen Schicht bei einem 69-jährigen
Patienten mit kleinem lakunärem Infarktresiduum in den Basalganglien rechtsseitig
(WL/WW = 35/60). Die IR-Stufe steigt von links nach rechts an, iDose 0 entspricht
gefilterter Rückprojektion alleine. Die Ziffern in den Indices zeigen den mittleren
subjektiven Score der jeweiligen Rekonstruktion, welcher von Fall zu Fall allerdings
stark variiert. b zeigt eine vergleichbare Anordnung an Rekonstruktionen einer infratentoriellen Schicht
bei einer 55-jährigen Patientin (WL/WW = 30/90). Die Artefakte an der Schädelbasis
sind unverändert, unabhängig von der iDose-Stufe. c zeigt eine Anordnung von Rekonstruktionen einer supratentoriellen Schicht bei einem
59-lährigen Patienten mit multiplen hemorrhagischen Metastasen (WL/WW = 35/60). Der
Einfluss von iDose Stufe und Filterkernel auf das Bildrauschen ist deutlich erkennbar,
die Abgrenzbarkeit der Läsionen profitiert nicht von höheren iDose-Stufen.
Subjective image evaluation
Inter-rater reliability was high with an ICC = 0.97 (adjusted for the effect of different
levels of professionalism). No significant influence of the different levels of professionalism
was observed (p = 0.203).
The analysis of the scores revealed significant (p < 0.05) differences in the mean
scorings of nearly every combination filter kernel and iDose level. All adjusted mean
scores (± 95 % confidence intervals, CI) are graphically displayed in [Fig. 2 ]. With an increasing iDose level, a substantial increase of the scores of the reconstructions
when applying the sharp filter kernel (UC) was present with eventually significantly
best scores at iDose level 4 (p < 0.001). An opposite trend was observed regarding
the reconstructions when applying the smooth filter kernel (UA), which scored nearly
equally well with the intermediate kernel (UB) at iDose level 0 (p = 0.75) and declined
in scores with increasing iDose levels. The intermediate kernel (UB) scored significantly
best at iDose levels 0 and 1, and still slightly higher than UC at iDose level 3 (not
statistically significant, p = 0.142).
Fig. 2 The graph displays the adjusted subjective scores (error bars represent the 95 %
confidence interval) of the different combinations of iDose level and filter kernel.
A continuous increase in score is only observed when applying the sharp kernel UC,
resulting in the significantly best score at iDose level 4 (versus UB, p < 0.001).
Abb. 2 Der Graph zeigt die adjustierten subjektiven Scores (die Fehlerbalken zeigen das
95 % Konfidenzintervall) der unterschiedlichen Kombinationen von iDose Stufe und Filterkernel.
Ein kontinuierlicher Anstieg des Scores zeigt sich nur bei Anwendung des harten Filterkernels
UC, signifikant beste Werte werden bei iDose Stufe 4 erreicht (versus UB, p < 0,001).
The overall adjusted mean score of datasets of female subjects was 3.05 (95 % CI:
2.77 – 3.32) compared to the mean score of male subjects with 2.86 (95 % CI: 2.58 – 3.13).
Although the confidence intervals overlap, the difference reaches statistical significance
(p = 0.023).
The bone-related beam hardening artifacts at the level of the skull base have not
been influenced by the different iDose level settings (see [Fig. 1b ]).
Objective image evaluation
The detailed results of the objective evaluation are presented in [Table 3 ] and graphically displayed in [Fig. 3a–c ]; additionally an illustration of the ROI placement is presented in [Fig. 3 d ].
Gray-white contrast was best when applying the sharp filter kernel (UC) and worst
when applying the smooth filter kernel (UA) regardless of the iDose level. With an
increasing iDose level, a uniform decrease of the contrast level by about 2.5 HU from
iDose level 0 to iDose level 4 was measured for every filter kernel applied.
Image noise was highest when applying the sharp filter kernel (UC) and lowest when
applying the smooth filter kernel (UA) regardless of the iDose level. With an increasing
iDose level, a substantial decrease of the image noise was measured, which was most
pronounced when applying the sharp kernel (absolute noise decrease from iDose 0 to
iDose 4: UC: 1.1 HU; UB: 0.7 HU; UA: 0.6 HU).
At every filter kernel the resulting CNR increased moderately with an increasing IR
level. The best CNR as well as the strongest increase over the range of iDose levels
were observed when applying the smooth filter kernel (UA: 0.7 points increase versus
0.6 at UB and 0.5 at UC).
Discussion
Discussing IR techniques in CT, dose reduction is the most important issue with respect
to different body regions [8 ]
[10 ]
[11 ]
[12 ]
[13 ]
[16 ]
[18 ]
[24 ]
[25 ]. Compared to the closely adjacent ocular lens, the brain is not that sensitive to
radiation exposure. However, all named organs would seriously benefit from dose reduction
and the associated reduction of scattered radiation in brain CT.
Image quality of brain CT might be of even greater interest, particularly regarding
the discrimination of gray and white matter and the well-known skull base-related
artifacts, which have been a research topic for decades [1 ]
[2 ]
[5 ]
[7 ]. Many reports of the application of different IR techniques in thoracic, abdominal
and cardiovascular CT imaging have been published and have raised hopes concerning
a possible improvement of image quality also in brain CT [10 ]
[12 ]
[13 ]
[23 ]
[26 ]
[27 ]
[28 ].
Therefore, this study addressed the intraindividual effects that can be possibly achieved
when applying a certain IR technique (iDose) and the other main component of CT reconstruction
(i. e., the filter kernel) on constant raw data input.
The key result of both the subjective as well as the objective evaluation is that
there is a major interdependence between the two key variables of image reconstruction
(contrast and noise), which is obviously similarly recognized by inexperienced raters
not familiar with reading brain CT. In subjective scoring this results in significant
trends towards a better image impression at sharper filter kernels with an increasing
IR level and subsequent decrease of subjective quality at softer kernels with an increasing
IR level.
However, the standard of reference, which is FBP without IR, scored equally well in
the subjective evaluation compared to the best-scoring combination at higher IR levels,
when a standard kernel was used in FBP. The good results of the sharp kernel at higher
IR levels are not obviously supported by the objective evaluation since the CNR is
higher and the image noise is lower for smooth and standard kernels with an increasing
IR level compared to the sharp kernel, regardless of the IR level applied (see [Fig. 3b, c ]).
Fig. 3 The gray-white contrast was measured as the difference of mean CT value in 5 small
adjacent ROIs in white and cortical gray matter (1.2 × 1.2 mm = 3 × 3 px) as indicated
in d (closed arrows). The graph a presents the means over all cases that show a similar decline by more than 2 HU with
an increasing IR level up to iDose 4; at every IR setting the highest measured contrast
is achieved when applying the sharp filter kernel (UC). Image noise was measured in
larger ROIs in white and cortical gray matter (6.1 × 6.1 mm = 14 × 14 px) as indicated
in d (open arrows). Image noise decreases with an increasing IR level, and the strongest
effect is observed at the sharp filter kernel (UC), as shown in graph b . The contrast-to-noise ratio (CNR, graph c ) as calculated from the contrast measured in the small ROIs and image noise measured
in the larger ROIs shows a moderate increase with an increasing IR level, which is
similar at every filter kernel. The best CNR at all IR levels is observed when applying
the smooth filter kernel (UA).
Abb. 3 Der Grau-Weiß-Kontrast wurde als Differenz aus mittlerem CT-Wert in 5 kleinen benachbarten
ROIs in weißer und grauer Substanz gemessen (1,2 × 1,2 mm = 3 × 3 px), wie in d gezeigt (geschlossene Pfeile). Der Graph a zeigt die Mittelwerte über alle Fälle, welche eine gleichartige Abnahme über insgesamt
mehr als 2 HU mit Anstieg der IR-Stufe auf maximal iDose 4 aufweisen. Der höchste
objektive Grau-Weiß-Kontrast besteht immer bei Anwendung des harten Filter-Kernels
UC. Das Bildrauschen wurde in größeren ROIs in weißer und grauer Substanz gemessen
(6,1 × 6,1 mm = 14 × 14 px), wie in d gezeigt (offene Pfeile). Das Bildrauschen nimmt mit zunehmender IR-Stufe ab (Graph
b ), der stärkste Effekt zeigt sich hier bei Anwendung des harten Filterkernels UC.
Das Kontrast-zu-Rausch-Verhältnis (Contrast-to-noise ratio, CNR; Graph c ) ergibt sich als Quotient aus den ROI-Daten zu Kontrast und Bildrauschen. Hier zeigt
sich bei allen Kombinationen ein moderater Anstieg mit ansteigender IR-Stufe, höchste
Werte werden bei Anwendung des weichen Filterkernels erreicht.
On the other hand, due to the distinct modulation transfer functions (MTF) of the
filter kernels, the gray-white contrast is substantially better when applying a sharp
kernel compared to smooth and standard kernels. Additionally, these absolute CT values
actually defining the contrast between gray and white matter converge with an increasing
IR level, regardless of the filter kernel applied. In other words: The contrast is
somewhat reduced as the IR level increases, but keeps best when applying a sharp kernel.
Here we may find some objective support for the result of the subjective scoring.
The apparent mismatch between objective evaluation in terms of CNR and subjective
evaluation demonstrates that CNR may fail as an indicator of objective image quality
in the case of CCT. Instead, image noise and contrast should be regarded independently.
If the image contrast is low as with filter UA, the subjective image quality is reduced
at higher IR levels despite reduced noise, as image contrast becomes even lower. If
the image contrast is high as with filter UC, the subjective image quality is improved
at higher IR levels, as reduced noise is perceived more positively than the reduction
in image contrast.
From the clinical point of view, and this is what is represented by the subjective
evaluation, which did not particularly discriminate between image noise and contrast,
the combination sharp kernel/iDose 4 scored very well. However, the image impression
is very similar to a FBP image with a smooth or standard filter kernel, which scored
equally high (see also [Fig. 1 ], [2 ]). Therefore – although there was no significant difference between the experienced
and the inexperienced raters –, we might have been observing an effect of a department’s
preference or a group’s familiarization with a certain kind of image characteristics
preferring moderate contrast combined with low noise, which would also be in line
with the lack of variability between the raters. Other departments preferring higher
contrast combined with moderate noise might come to different conclusions.
Which iDose feature can now possibly be translated into an improvement for diagnostic
CT of the brain? Practically, when applying iDose in brain CT, our department would
not go to levels beyond iDose level 3 since the loss in contrast may become too high.
However, diagnostic quality is a very subjective matter and therefore may be substantially
improved by iterative reconstruction when the radiologist concerned normally prefers
images with increased contrast, associated with increased noise (like filter UC).
A radiologist normally preferring low noise at moderate contrast (as with filter UA)
may not find a benefit in iDose or any other current IR technique.
One can imagine that next generation radiologists will be getting more and more familiar
with a certain difference in CT image impression. However, the visual IR effects have
to be evaluated in larger cohorts with particular pathological lesions and preferentially
in multi-center studies comprising institutes with different preferences concerning
contrast and noise.
Due to the retrospective character of the study and since we did not perform repeated
scans in the same patients, we were not able to look for possible effects due to variation
of dose settings, like has been reported for IR techniques of different vendors [17 ]
[21 ]
[29 ]. However, in the studies by Korn et al. and by Rapalino et al., helical CT acquisitions
have been performed, which are known to carry some advantages regarding skull base-related
artifacts, which were not improved by iDose in our collective. The helical acquisition
mode in brain CT has not been evaluated very intensively, but the few existing reports
refer to some loss in contrast resolution in comparable axial multiplanar reformations
(MPR) compared to sequential CT slices [1 ]
[6 ]
[7 ]. Therefore, having the decline in contrast resolution with increasing iDose level
in mind, helical brain CT might not benefit from IR.
An interesting future application may arise with the upcoming possibilities and growing
use of computer-aided diagnostic (CAD) tools working with tissue segmentation, e. g.
to detect local swelling in acute stroke [30 ]
[31 ]
[32 ]
[33 ]. A speculative but nevertheless imaginable future situation could be a brain CT
scan reconstructed in two ways: first, with vastly minimized noise resulting from
highest IR level, for CAD; second, with subjectively preferred visual settings, for
the human reader.
The most important limitation of this study may be due to the local preconditions
regarding personal preferences in perception of noise and contrast in brain CT as
well as the relatively narrow default window settings, which have also been applied
in this study.
Conclusion
Different combinations of iDose level and filter kernel substantially influence the
subjective and objective image quality of brain CT scans. The largest improvement
using IR might result for radiologists normally preferring high contrast at the expense
of increased noise. In such a setting IR could become an additional instrument of
controlling particular image characteristics. Our study does not allow giving recommendations
regarding the use of IR as a general dose reduction instrument in brain CT.