Key words thorax - CT - iterative image reconstruction - statistical image noise - dose reduction
Introduction
Multidetector computed tomography (MDCT) is fast, readily available and provides high-resolution
cross-sectional imaging which has an increasing impact on diagnostics and therapy.
Therefore, the number of examinations is continuously rising [1 ]. Thus, computed tomography (CT) scanning leads to increasing radiation exposure
which is widely discussed (1 – 3). Consequently, whether modalities such as ultrasound
or magnetic resonance imaging (MRI) can be used alternatively has to be checked carefully.
Sufficient image quality has to be achieved at the lowest possible dose level for
the remaining large number of CT indications. With respect to this, several technical
developments, such as automatic exposure control, automatic approaches to adapt the
tube voltage to patient size and the planned examination type, and others, have already
been successfully implemented in the past [1 ]
[4 ]
[5 ]
[6 ]
[7 ]
[8 ].
Besides other developments, improved reconstruction algorithms aim to reduce image
noise and are, therefore, a major focus of current developments in CT imaging [9 ]
[10 ]
[11 ]
[12 ]. This is promising for either improving image quality or reducing the dose. One
particular example of algorithms in this field is the adaptive statistical iterative
reconstruction technique (ASIR, GE Healthcare, Waukesha, WI), which is capable of
two working modes: a) covering the xy-plane (slice mode), or b) additionally covering
the z-axis (volume mode).
Although almost all vendors offer comparable options for their new scanner systems,
only a limited number of publications have investigated ASIR effects in human chest
imaging [13 ]
[14 ]
[15 ]
[16 ]
[17 ]
[18 ]
[19 ]
[20 ]. Current experiences using a 30 %-50 % ASIR are mainly based on phantom studies
[19 ] and vendor recommendations [13 ]
[16 ]
[17 ]
[21 ]. Many publications in this field are restricted to one or two ASIR blendings [13 ]
[16 ]
[17 ]
[20 ]. Sub-analysis of the particularly promising volume mode is limited to a single case
report [22 ]. Besides this, only one study did not restrict the evaluation to axial images but
also included the diagnostically relevant multiplanar reformations [20 ]. However, neither of these studies in patients used the lung kernel for image reconstructions
or the extended ASIR range, which can be varied between 0 % and 100 %.
The aim of this clinical study (part 2) is on the one hand to compare image quality
in dose-reduced 64-row CT of the chest at different levels of ASIR compared to full-dose
baseline examinations reconstructed solely with filtered back projection (FBP) in
a realistic upgrade scenario, and on the other hand to cross-check these results with
those of a similar study concerning the abdomen (Part 1, [23 ]).
Materials and Methods:
The Declaration of Helsinki was strictly followed, and a waiver of consent was granted
by the Institutional Review Board.
Phantom scans
A phantom (CATPHAN 500, Phantom Laboratory, Salem, NY) was scanned with the LightSpeed
VCT XT (LS-VCT, General Electric, Waukesha, WI, USA) and also with the HD750 machine
(GE) with and without the HD mode, using calculated, fixed parameters (200mA, 120 kV),
so that an identical radiation dose exposure was achieved compared to the follow-up
patient study. Images in the standard kernel and lung kernel were compared visually.
The subjective analyses were based on the quantitative discrimination of line pairs
in high contrast resolution and on the visibility and identifiability of round lesions
as the low contrast resolution targets.
Patients, material and design
Consecutive patients with a baseline CT examination of the chest on the LS-VCT, and
a CT follow-up within one year on the HD750 scanner were included, n = 46. The exclusion
criteria included obvious changes with potential influence of the underlying disease
on image quality, such as relevant alterations in tumor growth, lung infiltrations
or systemic edema.
Relevant technical parameters of both CT scanners were identical, but LS-VCT used
the V-Res detector (GE) whereas HD750 utilized the Gemstone detector (GE), which allows
for an acquisition of 2496 projections per gantry rotation (HD scan mode = high-definition
scan mode) instead of 984 projections compared to LS-VCT. As HD scan mode is a prerequisite
for the volume mode of ASIR, it had to be used for follow ups, but it was not available
for the baseline scans. According to the vendors’ information, ASIR assumes that CT
images contain the real information overlain by a dose-dependent level of noise. ASIR
uses a noise-predicting model to separate this information and identifies and decreases
the noise overlay iteratively. This process starts with FBP. The raw data is recalculated
after each step of noise reduction in the prediction model resulting in an optimized
raw dataset as the basis for the next step. This process is repeated until convergence
of noise reduction and edge preservation effects are reached. ASIR offers two different
modes to achieve these effects. In contrast to slice mode, the volume mode additionally
takes adjacent information along 1.75 cm of the z-axis into account, and thus is more
time-consuming but promises to be more effective. ASIR 2.0 can be merged with FBP
for follow-ups in 10 % steps of ASIR influence on the total reconstruction of the
images (i. e., image reconstruction eventually uses an x % ASIR and 100 %-x % FBP
images).
Scanning protocols
Patients were in a supine, feet first position. Standardized contrast-enhanced chest
scans were performed in the cranio-caudal direction. GE Healthcare uses a concept
known as the noise index (NI). The noise index (NI) relates to the standard deviation
of Hounsfield units in a water phantom of a specific size. A lookup table is used
to map the patient-specific attenuation values measured on the CT projection radiograph
(“scout” image) to tube current values for each gantry rotation according to a proprietary
algorithm. The algorithm is designed to maintain the same image noise level as the
attenuation values change from one rotation to the next. Within this study the NI
[24 ]
[25 ] was increased from 29 (baseline) to 45 for follow-up examinations, following the
manufacturer’s recommendation. Both scanners use the scout views to automatically
modulate the mA per rotation in order to achieve the predefined NI level [16 ]. An overview of other relevant protocol parameters is shown in [Tab. 1 ].
Table 1
CT scan parameters of the baseline examination performed on a LightSpeed VCT XT and
follow-up examination on a Discovery HD750 differing in noise index and number of
projections per gantry rotation (HD scan mode = high-definition scan mode; usable
on Discovery HD750).
Tab. 1 CT Scan-Parameter der Voruntersuchung (durchgeführt mit Lightspeed VCT XT) und der
Nachuntersuchung (durchgeführt mit Discovery HD750). Diese unterschieden sich im Rauschindex
(NI) und der Anzahl der Projektionen pro Röhrenrotation (HD-Scan Modus = high-definition
Modus, nur verfügbar auf dem Discovery HD750).
Lightspeed VCT XT
Discovery HD 750
(baseline)
(follow-up)
noise index
29
45
detector
V-Res
Gemstone
high definition scanning
no (984 projections)
yes (2496 projection)
high definition image reconstruction
no
no
tube voltage (kv)
120
120
auto mA tube current range (mA)
80 – 600
80 – 600
rotation time (s)
0.5
0.5
collimation (mm)
0.625
0.625
detector length (mm)
40
40
beam pitch
0.984:1
0.984:1
table feed (mm)
39.37
39.37
Processing and selection of images
AS the HD reconstruction kernel was not available for the baseline scanner and HD
reconstructions are limited to a maximum scan field-of-view (sFOV) of 25 cm, HD reconstruction
was not used in this study. Instead, the sFOV was identical for each HD750 study,
and the display field-of-view (dFOV) was matched to the baseline study given. The
raw data was reconstructed in both standard and lung kernels, resulting in nine settings
for each kernel: One setting with an ASIR of 0 % (solely with FBP), and the remaining
sets in both slice and volume mode, with ASIRs of 30 %, 50 %, 70 %, and 100 %, respectively.
Therefore, a total of n = 828 series were available (46 CT scans * 2 kernels * 9 differing
image reconstructions). The duration of image reconstruction was measured manually.
The resulting slices with a thickness of 0.625 mm were used to reformat image series
matched to those used in the baseline examinations given:
Standard kernel: axial, sagittal and coronal with a slice thickness of 3 mm
Lung kernel: axial with a slice thickness of 2.5 mm
This reformation process was chosen because 0.625 mm slices were a prerequisite for
ASIR reconstructions in the volume mode and for multiplanar reformations.
Three standard images were defined for each image series for the evaluation of body
diameters and noise measurements ([Fig. 1 ]):
Fig. 1 Objective image evaluation: Lines indicate patient’s diameters; circles show the
region of interest used. For standard kernel: a Axial: through the carina, c Coronal: through the aortic valve, d Sagittal: centric through the spine. For lung kernel: b Axial: through the carina.
Abb. 1 Standardisierte Bilder für die quantitative (objektive) Auswertung: Die Linien zeigen
die Patienten-Durchmesser, Kreise repräsentieren die verwendeten Region ́s of Interest
(ROI). a Axial: durch die Carina b Coronar: durch die Aortenklappe c Sagittal: mittig durch die Wirbelsäule.
Axial: at the level of the carina
Coronal: through the middle of the aortic valve
Sagittal: centric through the spine
Overall, these procedures ensured that the resulting HD750 datasets differed only
with respect to ASIR level and mode, that the slices were identical with respect to
position and size, and that the ASIR image series were as comparable as possible to
the corresponding baseline series.
Reading of images and collection of data
A single workstation (Advantage workstation version 4.5, GE) with two identically
calibrated displays was used for the data collection and the reading process. Viewing
conditions were held constant according to the AAPM TG18 report [26 ] and checked daily for consistency.
The mean image noise was assessed as the unweighted mean value of standard deviations (SD) of Hounsfield
Units using 100 [90 – 110] mm2 circular regions of interest (ROIs). Measurements of soft tissue were taken at areas
of muscle [1 ] and fat [2 ] that were as homogenous as possible in standard kernel images. The lung kernel images
were used to assess the noise in the lung tissue in the same way ([Fig. 1 ]). Although the SD is a complex combination of normal texture and noise-generated
inhomogeneities of the tissue samples, it still enables the estimation of the level
of statistical noise as reasonably as possible. In addition, sagittal and coronal
diameters of every patient were measured using a standardized procedure ([Fig. 1a ]).
Image quality was evaluated by three radiologists, blinded to the ASIR level and mode. The image
quality for ASIR follow-up studies was compared to the corresponding baseline series
using a five-point scale (+ 2: presumably diagnostically superior, + 1: superior,
0: equal, -1: inferior, -2: presumably diagnostically inferior). In addition, the
readers had to elucidate whether too much image noise, oversmoothing/blurring or another
reason caused the diagnostically inferior image quality in the case of a rating of
-2. A total image quality score was calculated as the mean value of ratings in the axial, sagittal and coronal
planes. The lung kernel evaluation criteria included the noise, contrast and sharpness
of lung parenchyma, as well as the sharpness and contrast of central and peripheral
lung vessels and airways. The criteria for the standard kernel included the noise,
contrast and sharpness of the mediastinum, chest wall and muscle [27 ].
Statistical analysis
The Mann-Whitney U Test was used to compare ASIR images to the baseline, whereas the
Wilcoxon rank-sum test was used to detect differences within the ASIR groups. The
intraclass correlation coefficient (ICC) was calculated as a measure of inter-reader
variability. All statistical testing was performed with SPSS 18.0.0 (PASW, Chicago,
IL, USA).
Results
Phantom scans
No visual differences could be detected for the lung kernel (neither between the scanners
nor between HD and non-HD scanning). Also no differences between HD and non-HD scans
for the standard kernel were detected, but the HD750 image shows visually better contrast
resolution compared to the LS-VCT scan ([Fig. 2 ]).
Fig. 2 A CATPHAN Phantom scanned with 2496 projections per rotation (HD scan) and with 984
projections per rotation (non-HD scan). a , b , d , and e were scanned on the Discovery HD750, while c and f were scanned on the LightSpeed VCT XT. Top: in lung kernel for high-contrast resolution
(windowing: W = 2000, L = 350), and below: in standard kernel for low-contrast resolution
(windowing: W = 50, L = 50). No differences are observable between non-HD and HD scans
in high-contrast resolution. In contrast, low-contrast resolution of the HD750 scans
was visually better compared to the LS-VCT data.
Abb. 2 Ein CATPHAN Phantom wurde, sowohl mit 2496 Projektionen pro Röhrenumlauf (HD Scans)
als auch mit 984 Projektionen pro Röhrenumlauf (no-HD Scans) gescannt. a , b, d and e wurden auf dem Discovery HD750 gescannt, wohingegen c und f auf dem Lightspeed VCT XT gescannt wurden. Im oberen Bildanteil ist die hoch-Kontrast
Auflösung im Lung-Kernel dargestellt (Fensterung: W = 2000, L = 350) und darunter
für die niedrig-Kontrast-Auflösung im Standard-Kernel. (Fensterung: W = 50, L = 50).
Für die hoch-Kontrast Auflösung war kein Unterschied zwischen HD und no-HD Scans erkennbar.
Im Gegensatz dazu war das Ergebnis des Discovery HD750 für die niedrig-Kontrast Auflösung
visuell besser im Vergleich zum Lightspeed VCT XT.
Patient scans
The resulting follow-up collective had the following parameters: age was 60 ± 15 [23 – 85]
years, 26 % female. The mean body weight was 83.1 ± 14.5 kg. The mean sagittal diameter
at the level of the tracheal bifurcation was 24.84 ± 3.42 cm (baseline 24.64 ± 3.33 cm,
p = 0.18), and the mean coronal diameter was 37.95 ± 3.0 cm (baseline 37.94 ± 2.96 cm,
p = 0.98).
The image reconstruction time for the standard and lung kernel was 58 ± 9 seconds for both slice mode and FBP,
increasing to 102 ± 15 seconds for the corresponding volume mode series.
The mean CTDIvol
of the patient studies significantly decreased by 53 % from 17.2 ± 6.9 (baseline)
to 8.0 ± 2.3 mGy (ASIR follow-up; p < 0.001). The CTDIvol of the phantom scans was 7.9 ± 0.1 mGy.
Objective image noise
Increasing ASIR levels resulted in an approximately linear reduction of image noise in slice and volume mode for the standard and lung kernel.
Lung kernel
The mean image noise of follow-ups was always higher compared to the baseline, but
only significant at an ASIR of 0 %. Axial image reconstructions in the volume mode
resulted in a slightly higher objective image noise than in the slice mode. However,
this was not significant, neither between the two modes nor at the baseline examination
(p > 0.053, [Fig. 3 ])
Fig. 3 Mean (SD) image noise in the lung kernel: Comparison of 50 % dose-reduced ASIR to
full-dose FBP baseline. Image noise of follow-ups was always higher compared to baseline
(FBP) image noise, despite ASIR influence. There was only a significant difference
compared to the baseline study at an ASIR of 0 %, but not between slice and volume
modes.
Abb. 3 Durchschnittliches (SD) Bildrauschen im Lung-Kernel: Vergleich von 50 % dosis-reduzierten
ASIR zu voll-dosis FBP-Baseline Untersuchungen. Das Bildrauschen der Folgeuntersuchungen
war unabhängig vom ASIR Einfluß immer höher als das der Baseline (FBP). Jedoch war
eine signifikante Differenz im Vergleich zur Baseline nur bei ASIR 0 % festzustellen,
aber nicht zwischen Slice- und Volumen–Modus.
Standard kernel
Axial image reconstructions (ASIR 30 %-100 %) in the volume mode resulted in significantly
higher image noise than in the slice mode (p < 0.01), whereas the values for MPRs
were nearly identical ([Fig. 4 ]). When analyzing the image noise of axial pictures of the follow-ups, the ASIR levels
of 30 % and 50 % in slice and volume mode and 70 % in volume mode were able to provide
image noise levels that were comparable (p > 0.089) to an ASIR 50 % slice and an ASIR
70 % volume and therefore almost identical to the full-dose baseline. ASIR levels
below these limits cause higher image noise (p = 0.001), whereas higher levels imply
a noise reduction (p ≤ 0.018). In contrast, the noise in MPR images was always higher
compared to the full-dose baseline (p ≤ 0.002), except for an ASIR of 100 % (slice:
p = 0.122; volume: p = 0.235).
Fig. 4 Mean image noise in standard kernel: Comparison of 50 % dose-reduced ASIR to full-dose
FBP baseline. Increasing ASIR levels resulted in an approximately linear decrease.
While follow-up image noise at an ASIR of 70 % in the axial plane approached the baseline
and undermatched with 100 %, the MPRs' image noise is always higher than that of FBP
reconstructed MPRs.
Abb. 4 Durschnittliches Bildrauschen im Standard-Kernel: Vergleich von 50 % dosis-reduzierten
ASIR Studien zu voll-dosis Baseline Untersuchungen. Bei ansteigendem Einfluß von ASIR
zeigte sich ein nahezu linear abnehmendes Bildrauschen. Die ASIR Folgestudien erreichten
den Wert des Bildrauschens der Voruntersuchung bei ASIR 70 % und lagen bei 100 %ASIR
Einfluß darunter. Das Bildrauschen der MPRs war immer höher als das Bildrauschen der
mit FBP rekonstruierten MPRs.
Image quality
Lung kernel
The image quality of the follow-ups outperformed the baseline examination at every ASIR level (p < 0.001),
and performed best at an ASIR of 30 % (slice: 0.70 ± 0.60, volume: 0.74 ± 0.61). In
particular, an ASIR of 30 % was rated significantly better than an ASIR of 0 % (i. e.,
solely calculated with FBP) in slice mode (p = 0.029) and volume mode (p = 0.047).
A significant difference between slice mode and volume mode was limited to an ASIR
of 100 % (p = 0.002) ([Fig. 5 ], [6 ]).
Fig. 5 Image quality in lung kernel: Comparison of 50 % dose-reduced ASIR to full-dose FBP
baseline studies. Every level of the ASIR performed significantly superior to the
baseline study. The best performance was evaluated at an ASIR of 30 %, whereas smoothing
effects of high ASIR levels led to inferior performance, especially in the slice mode.
Score: + 2 = presumably diagnostically superior, + 1 = superior, 0 = equal, -1 = inferior,
-2 = presumably diagnostically inferior.
Abb. 5 Bildqualität im Lung-Kernel: Vergleich von 50 % dosis-reduzierten ASIR- zu voll-dosis
FBP-Baseline Studien: Die mit ASIR berechneten Studien waren unabhängig vom ASIR Einfluß
alle der Baseline Untersuchung überlegen. Am besten wurde die Bildqualität bei ASIR
30 % bewertet, wohingegen vor allem im Slice-Modus bei höherem ASIR Einfluss Weichzeichnungseffekte
zu einer schlechteren Bewertung führten. Skala: + 2 = vermutlich diagnostisch besser,
+ 1 = besser, 0 = gleich, -1 = schlechter, -2 = vermutlich diagnostisch schlechter.
Fig. 6 Composed images from one patient in the standard kernel (all planes) and in the lung
kernel (axial plane). Examples of the full-dose baseline as the reference study and
ASIR calculated images at different ASIR levels of the dose-reduced follow-up study.
Compared to the baseline, an ASIR of 0 % showed up with more image noise according
to the reduced dose. The noise decreases with increasing ASIR blendings, but can result
in oversmoothing images, especially at ASIR 100 % in slice mode.
Abb. 6 Bildbeispiele eines Patienten im Standard-Kernel (alle Ebenen) und im Lung-Kernel
(axiale Ebene): Beispiele der voll-dosis Baseline-Untersuchung als Referenz und dosis-reduzierten
ASIR-Bildern, die mit unterschiedlichem ASIR-Einfluß berechnet worden sind. Verglichen
mit der Baseline zeigten sich ASIR0 % Bilder mit einer erhöhtem Bildrauschen was bei
der reduzierten Dosis verständlich ist. Das Bildrauschen verringert sich mit zunehmendem
Einfluß von ASIR auf die Bildberechnung, kann aber zu sehr weichgezeichneten Bildern
führen, vorallem bei ASIR 100 % im Slice-Modus.
Standard kernel
The total image quality reached the level of the baseline examination with ASIR blendings of 70 % (-0.07 ± 0.29;
p = 0.289) and 100 % (-0.10 ± 0.42; p = 0.177) solely in the volume mode. In contrast,
all ASIR levels were rated significantly inferior to the baseline examination (p < 0.01)
for the slice mode. Even though the volume mode was almost always rated better than
the slice mode, particularly with higher ASIR levels, this effect was only significant
for 70 % (p = 0.026) and 100 % (p < 0.001). The results are summarized in [Tab. 2, ]
[Fig. 6 ], [7 ].
Table 2
Mean subjective image quality ratings ± standard deviation of dose-reduced images
calculated with different Adaptive Statistical Iterative Reconstruction (ASIR) blending
levels compared to full-dose filtered back projection (FBP) baseline examination (semi-quantitative
5-point scale: + 2 = presumably diagnostically superior, + 1 = superior, 0 = equal,
-1 = inferior, -2 = presumably diagnostically inferior).
Tab. 2 Durchschnittlich subjektive Bewertungen der Bildqualität ± Standardabweichung von
Dosis-reduzierten Untersuchungen, die mit unterschiedlich starkem Einfluss des Adaptiven
Statistischen Iterativen Rekonstruktions-Algorithmus (ASIR) berechnet wurden, im Vergleich
zu einer Voruntersuchung mit voller Dosis, welche mittels gefilterter Rückprojektion
(FBP) berechnet wurden (semi-quantitative 5 Punkte-Skala: + 2 = vermutlich diagnostisch
besser, + 1 = besser, 0 = gleich, -1 = schlechter, -2 = vermutlich diagnostisch schlechter).
ASIR level (%)
ASIR mode
0
30
50
70
100
Standard kernel
Axial
Slice
–0.41 ± 0.52
–0.17 ± 0.38
–0.07 ± 0.40
–0.09 ± 0.59
–0.90 ± 0.49
p < 0.001***
p = 0.033*
p = 0.357
p = 0.418
p < 0.001***
Volume
–0.41 ± 0.52
–0.16 ± 0.58
–0.03 ± 0.38
–0.01 ± 0.68
–0.10 ± 0.60
p < 0.001***
p = 0.143
p = 0.701
p = 0.901
p = 0.352
Sagittal/Coronal
Slice
–0.64 ± 0.51
–0.32 ± 0.41
–0.16 ± 0.38
–0.17 ± 0.40
–0.38 ± 0.38
p < 0.001***
p < 0.001***
p = 0.020**
p = 0.014*
p < 0.001***
Volume
–0.64 ± 0.51
–0.30 ± 0.47
–0.14 ± 0.27
–0.09 ± 0.34
–0.09 ± 0.49
p < 0.001***
p = 0.003**
p = 0.012*
p = 0.164
p = 0.387
Total
Slice
–0.57 ± 0.40
–0.27 ± 0.30
–0.13 ± 0.26
–0.14 ± 0.36
–0.55 ± 0.32
p < 0.001***
p < 0.001***
p = 0.006**
p = 0.010*
p < 0.001***
Volume
–0.57 ± 0.40
–0.25 ± 0.38
–0.11 ± 0.22
–0.07 ± 0.29
–0.10 ± 0.42
p < 0.001***
p < 0.001***
p = 0.020**
p = 0.289
p = 0.177
Lung kernel
ASIR mode
0
30
50
70
100
Axial
Slice
0.57 ± 0.63
0.70 ± 0.60
0.59 ± 0.63
0.52 ± 0.63
0.36 ± 0.71
p < 0.001***
p < 0.001***
p < 0.001***
p < 0.001***
p = 0.003**
Volume
0.57 ± 0.65
0.74 ± 0.61
0.68 ± 0.61
0.62 ± 0.60
0.59 ± 0.67
p < 0.001***
p < 0.001***
p < 0.001***
p < 0.001***
p < 0.001***
Significance levels: (*) p < 0.05; (**) p < 0.01; (***) p < 0.001, p > 0.05 in bold.
Signifikanz-Level: (*) p < 0.05; (**) p < 0.01; (***) p < 0.001, p > 0.05 in Fettdruck.
Fig. 7 Image quality in the standard kernel: Comparison of 50 % dose-reduced ASIR to full-dose
FBP baseline studies. With increasing levels of ASIR, the image quality of the follow-up
study approached the value “0” as the reference image quality of the baseline and
performed best at an ASIR of 70 % in the volume mode. Axial images at an ASIR of 100 %
in slice mode performed even inferior to the ASIR of 0 %, i. e., the FBP series at
the same dose level. Score: + 2 = presumably diagnostically superior, + 1 = superior,
0 = equal, -1 = inferior, -2 = presumably diagnostically inferior.
Abb. 7 Bildqualität im Standard-Kernel: Vergleich von 50 % dosis-reduzierten ASIR- zu voll-dosis
FBP-berechneten Studien. Mit ansteigendem Einfluß von ASIR erreichte die Bildqualität
der Folgeuntersuchungen den Wert „0“, definiert als Referenz durch die Baseline-Untersuchung,
und stellte sich am besten bei einem ASIR-Einfluss von 70 % im Volumen Modus dar.
Die axialen Bilder bei ASIR 100 % im Slice-Modus jedoch sogar schlechter als ASIR
0 %, welches ausschließlich mit FBP berechnete Bildern bei identischer Dosis entspricht.
Skala: + 2 = vermutlich diagnostisch besser, + 1 = besser, 0 = gleich, -1 = schlechter,
-2 = vermutlich diagnostisch schlechter.
Axial reformations in the slice mode performed significantly superior compared to
MPRs with ASIRs of 0 %-30 % (0 %: p = 0.004; 30 %: p = 0.021), whereas ASIRs of 50 %-70 %
performed comparably (p ≥ 0.347) and an ASIR of 100 % was inferior (p < 0.001). In
this manner, axial images in the volume mode only performed superior with an ASIR
of 0 % compared to MPRs (p = 0.004).
Overall, the best particular results for MPRs were reached with an ASIR of 70 % in
the volume mode (-0.09 ± 0.34; p = 0.164). The best ratings for axial images resulted
from an ASIR of 70 % blending in the volume mode (-0.10 ± 0.68; p = 0.90). In contrast,
an ASIR of 100 % in the slice mode performed worst (p < 0.001) and even ranged below
the ASIR of 0 % value ([Tab. 2, ]
[Fig. 6 ], [7 ]). Diagnostically inferior images (i. e., ratings with scores of -2) were limited
to an ASIR of 0 % and axial reconstructions with an ASIR of 100 % in the slice mode.
Within the ASIR 0 % (i. e., FBP) subgroup, these low ratings occurred in 9 % of the
images and were all caused by inadequate image noise. In contrast, for the subgroup
of axial reconstructions with an ASIR of 100 % in the slice mode, low rating scores
occurred in 22 % of the images and were all caused by oversmoothing.
Inter-reader variability
Acceptable inter-reader variability was observed for the lung kernel with an ICC of
0.67 (0.57 – 0.75; 95 % confidence interval CI, p < 0.001), and for the standard kernel,
with an ICC of 0.70 (0.62 – 0.76; 95 % CI, p < 0.001).
Discussion:
As vendors especially advertise the dose-saving potential of their new systems and
supply recommendations for new protocols, this study particularly aims to get the
results of the typical situation when CT systems are upgraded. In this manner, the
installation of a newly designed CT platform (such as the HD750) and its performance
compared to its predecessor (such as the LS-VCT) were the inducement for closer inspection.
In order to check whether the scanning protocols as supplied by the vendor are appropriate,
we strictly followed the manufacturer’s recommendations for CT scanning.
According to the highest ethical standards, every examination required a medical indication
and the patient dose exposure had to be strongly minimized within the limits of the
used system. We had to compare possible influences of HD scanning on the results,
as data from HD scans was required as input for the volume mode and this option was
not available on the baseline LS-VCT platform. It was not possible for ethical reasons
to acquire HD and non-HD scans from the same patient during one examination. Instead,
we compared these effects using a phantom.
Overall, the LS-VCT did not offer HD scanning but the protocols on the HD 750 were
designed to match those protocols as closely as possible. Among others, this implicated
the slice thicknesses used and also caused an unavoidable time interval between the
two patient examinations. Comparability of the whole group is given, as the differences
of the diameters of the patients’ bodies are not significant, which are known to be
the most important factors for modulation of the dose in CT scanning [28 ].
Three major changes between the two scanners contributed to the effects which were
investigated in this study: V-Res/Gemstone detector, non-HD/HD scanning, and FBP/ASIR
image calculation. A phantom was additionally scanned with both settings in order
to estimate the particular contributions. HD scanning is known to be superior as long
as images are additionally calculated with the HD kernel because the image matrix
is able to display the benefits of higher resolution [17 ]. As long as no HD reconstruction is performed, no relevant differences between HD
and non-HD scanning on the HD750 machine were observed. This implies no or only minor
contribution of this parameter to image quality ([Fig. 2 ]). In comparison, the low-contrast resolution of the LS-VCT scans was visually slightly
inferior compared to the HD750 scans. In contrast to HD scanning, this implies a relevant
contribution of the detector to the image quality, which is concordant with a recent
study where the Gemstone detector outperformed the V-Res detector regarding image
noise by 7.9 – 20.6 %, depending on the slice thickness and reconstruction kernel
[23 ]
[29 ].
The mean image noise in the standard kernel decreased roughly linearly with increasing
influence of ASIR ([Fig. 4 ]). Similar results were observed in other publications [18 ]
[19 ]. In this context, a higher noise reduction effect was observed solely for axial
images in the slice mode. Since this mode does not take the z-axis into consideration,
the observed effect could be attributed to a more effective but less realistic noise
reduction (this is supported by the lower image quality especially of those images
with high ASIR levels in [Fig. 7 ]). As the dose of both groups is far above doses used in dedicated low-dose protocols
for the lung, the dose difference between the baseline and the follow-up group can
be expected to be negligible (of course this does not apply for soft tissue evaluation).
Thus, the superior ratings compared to the FBP baseline with a higher dose (and thus
lower image noise) might have to be credited to the new Gemstone detector in the HD750
machine. In addition, image noise is only one of obviously many factors which influence
image quality. In detail, only the portion of “real statistical noise” is relevant
and should be minimized. Other noise components, such as tissue inhomogeneities, show
completely different influences and thus have to be preserved. The influence of edge
enhancement for high-contrast images, such as the lung in the lung kernel, is much
more relevant to the image quality than image noise. Once again, this does not apply
to soft tissue evaluation. Nevertheless, the problem of noise measurements remains
the same. The performed measurements are not able to distinguish between real statistical
noise and the “noise” caused by tissue inhomogeneities. As a consequence, the additional
coverage of the z-axis in the volume mode may help the algorithm to differentiate
between these two major noise components and may, therefore, focus the suppression
of total noise more specifically on the real statistical component. Although it is
not possible to prove this statement with the protocols used, it may be supported
by identical findings within the abdominal part of the study. The higher noise of
axial images compared to MPRs could be attributed to the primary voxel size with a
higher resolution in the z-axis (0.625 mm axial slice thickness < FOV 50 cm/512 matrix = 0.975 mm),
a somewhat sub-isotropic resolution [30 ].
The image quality in the lung kernel was rated best at an ASIR of 30 % and is significantly
superior compared to conventionally FBP-reconstructed image sets. This finding is
also supported by other published recommendations of an ASIR of 30 % in chest CT imaging,
even though sharper kernels, such as the detail or bone kernel, were used instead
of the lung kernel [13 ]
[16 ]
[17 ]
[18 ]
[20 ]. The image quality for soft tissue evaluation in the standard kernel was rated best
at an ASIR of 70 %. Even though this seems to be contrary to results in abdominal
CT imaging which recommend the use of a 30 %-50 % ASIR, the findings of both studies
are concordant considering that the abdominal studies were performed with a higher
CDTI [19 ]
[21 ] and that the more the dose is reduced, the more the particular ASIR effects seem
to contribute to the appropriate enhancement of the image quality [18 ]. Furthermore, a recent study with a comparable design supports corresponding results
for the CT examination of the abdomen and pelvis [14 ]
[23 ].
An oversmoothing effect is known to appear when using very high ASIR levels, particularly
in the slice mode, resulting in both low image noise and inferior image quality [13 ]
[15 ]
[17 ]
[21 ]. Obviously, image quality is not only a consequence of image noise but also of edge
preservation [13 ]. This effect of high-contrast kernels, such as the lung kernel, lead to a reduction
of the image quality at ASIR levels of 50 –100 %, especially in the slice mode. A
comparable effect for the standard kernel appeared for an ASIR of 100 % in the volume
mode, and for 70 –100 % in the slice mode.
Limitations of the study:
Limitations of the study:
For ethical reasons, follow-up examinations had a medically indicated minimum time
delay and, therefore, dose reduction was not a variable but a consequence of the protocols
provided by the vendor.
The study design allowed the influence of three different major variables. However,
results indicate no/negligible influence of HD scanning (as long as it was performed
without HD image reconstruction). The detector might be awarded 10 – 20 % of the effects
detected and led to a corresponding minor shifting of curves [29 ]. The remaining major part is contributed to the ASIR.
Apart from changes in the patients’ body weights, the underlying disease may also
influence image quality. However, none of the patients had to be excluded for this
reason.
Although the rating criteria were according to guidelines on quality criteria for
chest CT examinations, this does not allow final conclusions regarding diagnostic
safety. Despite the attempt to keep these effects as low as possible, ratings of image
quality were restricted to a subjective image impression.
Results are limited to a specific vendor, here GE Healthcare.
Overall, the vendors’ recommendations for upgrading to iterative image reconstruction
were fair, but not ideal. With a 50 % dose reduction, an ASIR of 30 % even enhances
image quality for the lung kernel in both modes, while for the standard kernel, an
ASIR of 70 % in the volume mode leads to an image quality that is similar to full-dose
FBP images. Comparable to the setup of this study, most users follow the vendors’
recommendations and thus the results of this study should also apply for their specific
environment. For different baseline dose levels there might be a varying dose reduction
potential. However, although different algorithms and study protocols were used, the
reported dose reduction potential of ASIR ranged from 35 % [31 ] to 64 % [32 ]. When considering the particular upgrade scenario of our study, the iterative image
reconstruction algorithm was the major factor for dose reduction in CT imaging of
the chest, whereas a more sensitive detector might have contributed to additional
minor dose savings as well. Although high-definition imaging (i. e., HD scanning and
HD reconstruction) is proven to be superior [17 ], it showed no relevant impact in this study with a standard FOV and without HD data
reconstruction. Iterative reconstruction algorithms significantly enhance image quality
(XH HU), even for low-dose examinations, and a combination of low-dose protocols,
HD imaging and iterative image reconstruction may be highly promising. However, this
would require separate reconstructions of the left/right lungs in HD-FOVs of < 25 cm
[17 ].
In the future, vendors will have to consider the implementation of upgraded computational
power to keep this procedure fast enough for the daily routine, especially for the
promising upcoming model-based approaches, which are even more complex, as they additionally
consider exact geometrical features of the patient, the detectors and the beam distribution
[33 ].