Four-dimensional noise reduction using the time series of medical computed tomography datasets with short interval times: a static-phantom study

Introduction

Four-dimensional noise reduction (4DNR) is a powerful post-processing method of reducing noise by employing spatial–temporal analysis. Large noise reduction using time analysis is well known in other fields such as video processing. Noise and artifacts can appear randomly in the temporal direction. By adopting temporal filtering with an appropriate estimation of motion (Brailean et al., 1995), the effect of random noise is minimized.

Several filtering methods have been recently proposed for the 4DNR of medical images (Liu et al., 2009) and can be used for a four dimensional dataset with any medical imaging modality such as CT. For the CT datasets, because not only the volume dataset but also temporal axis data are required, only multi-data acquisition examination has been considered for the application of 4DNR. Thus, 4DNR is generally performed for time series of volume datasets having a relatively long interval time (e.g., 1000 ms), such as datasets of computed tomography (CT) perfusion examinations (Li et al., 2014a; Li et al., 2014b; Corcuera-Solano et al., 2014). However, the applicability of 4DNR to time series of CT volume datasets with short interval times has not been fully discussed (Otton et al., 2013; Tatsugami et al., 2015). In contrast to the case for datasets having long time intervals, the strong advantage of applying 4DNR in the case of short interval times is considered to be the ability to scan datasets constructed by single data acquisition. Thus, we developed a method of applying 4DNR to datasets having short interval times (≤50 ms), which we refer to as the method of obtaining a noisele ss imag e by a daptive phase-shifted to pological coherence analysis or legato.

Cardiac CT (CCT) usually acquires datasets that include “padding” phases centered on the mid-diastole. From these “padding” phases, time-series datasets with short interval times can be easily reconstructed (Tatsugami et al., 2015). However, the additional information provided by additional phases has largely been ignored in post-processing noise reduction for CCT. Legato can be applied to such datasets to reduce noise and the radiation dose in CCT. The present study conducts quantitative image quality analysis using static ex vivo phantoms and in vivo retrospective analysis to examine the hypothesis that post processing with legato reduces the noise in CCT images and permits a lower radiation dose when using “padding” phases.

Materials and Methods

Results

Ex vivo preliminary analysis of legato

The CNR improved with an increase in interval time from 10 to 50 ms. However, the CNR reached a plateau at an interval time of 50 ms, beyond which it did not change considerably (Fig. 1). The CNRs were 1.33 ± 0.16 for the reference image, 1.48 ± 0.18 for an interval time of 10 ms, 1.90 ± 0.22 for 30 ms, 2.41 ± 0.38 for 50 ms, 2.61 ± 0.27 for 70 ms, and 2.65 ± 0.27 for 90 ms. In addition, the MSE for each interval time of 10, 30, 50, 70, and 90 was 14.0 ± 0.7, 64.9 ± 4.5, 114.2 ± 5.1, 151.0 ± 5.8, and 149.3 ± 4.9 respectively. Although, the MSE increased significantly with the interval time until 70 ms (P < 0.001), the MSEs for intervals of 70 and 90 ms were not significantly different.

Ex vivo quantitative image analysis of legato

The CNR obtained with legato was significantly better than that of the reference image (1.18 ± 0.15 versus 2.08 ± 0.19, P < 0.001) (Figs. 2A, 2B and 3A). Even when employing the IR method, significant noise reduction was possible with legato (1.25 ± 0.15 versus 2.21 ± 0.22, P < 0.001) (Figs. 2C, 2D and 3B). In contrast, the CT attenuation did not change significantly when using legato (Table 2).

Table 2:

Results on the equivalence test and non-inferiority test.

	Reference	Legato	Margin	Difference [95% CI]	P value
CT value (HU)
FBP	63.4 ± 0.6	63.4 ± 0.4	2	−0.02 [−0.51–0.46]	<0.001
IR	63.6 ± 0.6	63.9 ± 0.4	2	−0.05 [−0.57–0.47]	<0.001
MTF10% (cycles/mm)	0.55 ± 0.01	0.56 ± 0.02	0.05	0.005 [−0.010–0.019]	<0.001

	Reference	Quarter dose +legato	Margin	Difference [95% CI]	P value
Ex vivo
CNR	1.25 ± 0.15	1.37 ± 0.09	0.1	0.113 [−0.003–0.230]	<0.001
CT value (HU)	63.6 ± 0.6	63.9 ± 0.9	2	−0.3 [−0.96–0.40]	<0.001
In vivo
SD (HU)	11.46 ± 1.99	11.01 ± 1.84	1	−0.455 [−1.89–0.980]	0.027

DOI: 10.7717/peerj.1680/table-2

Notes:

FBP

filtered back projection

IR

iterative reconstruction

CI

confidence interval

MTF

modulation transfer function

SD

standard deviation

Legato

the method of obtaining a noisele ss imag e by a daptive phase-shifted to pological coherence analysis

In the assessment of spatial resolution, MTF_10% obtained with legato was not inferior to that of the reference image (Table 2). Each MTF curve is shown in Fig. 4A.

According to NPS analysis shown in Fig. 4B, noise was reduced equally throughout the frequency band.

Ex vivo assessment of the ability to reduce the radiation dose using legato

With legato, CNRs for the quarter radiation dose significantly improved from 0.71 ± 0.15 to 1.37 ± 0.09 (P < 0.001) (Figs. 2E and 2F) and were not inferior to those of the reference dose image (Table 2 and Fig. 3D). CT attenuation did not differ among reference dose image, quarter radiation dose image, and quarter radiation dose image with legato (63.6 ± 0.6 HU, 63.7 ± 0.5 HU and 63.9 ± 0.9 HU, respectively, P < 0.001 for each). Furthermore, the subjective image score had a median value of 16 (range 13–18) for the reference, 6 (range 4–9) for the quarter radiation dose, and 15 (range 13–18) for the quarter radiation dose with legato.

In vivo assessment of the ability to reduce the radiation dose using legato

The details of the subjects in the in vivo study are listed in Table 3. The legato images were obtained successfully and the post-processing time was approximately 10 min in each case. A representative image for the case of a 70-year-old male is shown in Fig. 5. The noise level in the image acquired at a quarter radiation dose and processed with legato was significantly lower than that in the image not processed legato (17.11 ± 2.73 versus 11.01 ± 1.84, P < 0.001) and no worse than that in the image acquired with the reference radiation dose (Table 2 and Fig. 3D).

Table 3:

Subject characteristics in-vivo study.

Variables
Age(years)	66.3 ± 13.8
Female/male (n)	2/13
Body mass index	22.4 ± 3.97
Heart rate (bpm)	56.6 ± 7.00
R–R interval time (ms)	1,066 ± 147.7
CTDI (mGy)	50.2 ± 8.49
Contrast material volume (mL)	60.4 ± 13.0
Injection flow rate (mL/s)	3.60 ± 0.61

DOI: 10.7717/peerj.1680/table-3

Notes:

CTDI

computed tomography dose index

Discussions

Our present study using static phantoms showed that legato could reduce noise throughout the investigated frequency band without degrading the CT numbers or spatial resolution. Further, the noise level of images obtained by scanning with a quarter radiation dose and post processing with legato was not inferior that of the reference images obtained with a full radiation dose.

CCT images provide a good test of the performance of legato in that several cardiac phases with short interval times are usually obtained as a padding phase (Tatsugami et al., 2015). In addition, the noise level would be higher than that in full reconstruction because there are fewer photons in the data acquisition of the half reconstruction that accounts for temporal resolution. Employing temporal filtering, this noise resulting from the half reconstruction could be compensated. Moreover, artifacts in half reconstruction are well known to mainly depend on the acquisition degree. Using a different phase means the use of a different acquisition degree. Thus, theoretically, these artifacts could also be minimized employed temporal noise reduction.

Interval times are important parameters when using legato in CCT. Small movements might be better for the precise estimation of movement. Thus, datasets with relative short interval times are required. However, it is not easy to distinguish noise in datasets for which the intervals are too short because the images and the noise they contain are almost the same. Our preliminary study showed that noise reduction for interval times from 10 to 30 ms was relatively low. In contrast, the noise reduction reached a plateau at interval times exceeding 50 ms. Larger interval times can thus improve noise reduction but possibly only to a certain degree. If the interval time is set as 70 ms, the three phases are separated by 90°for our scanner (having a rotation time of 280 ms). In dual-source CT, two detectors are aligned at almost 90°, and datasets having phases shifted by 90°contain data of completely different images with completely different acquisition. This would affect the limit of noise reduction of legato for CT images. Moreover, in coronary artery analysis, interval times of 50 ms rather than longer interval times are considered the most appropriate in a clinical setting. Because larger interval times mean departing from the mid-diastolic phase, these separated datasets could be adversely affected by a motion artifact of the coronary arteries. However, to investigate the appropriate interval times for legato in a clinical setting, further study using moving cardiac phantoms is needed. Moreover, the appropriate interval time would depend on the rotation speed and whether dual- or single-source CT was employed.

Image filtering can reduce the resolution and produce artifacts by oversmoothing. There is thus a trade-off between such artifacts and the reduced rate of noise. However, our legato can uphold the CT attenuation and the resolution with powerful noise reduction for static phantoms. Temporal filtering methods are known to avoid the oversmoothing of voxels because of their spatial independence (Brailean et al., 1995). Thus, legato improves the image quality by reducing noise while preserving structural detail, which is important for CCT. The importance of the spatial resolution has been well discussed in the detection of coronary stenosis (Leber et al., 2005). Recently, the CT attenuation of plaque has also been an important research topic. Plaque that is vulnerable to rupturing is known to have a low CT attenuation, and a small ulcer having a napkin-ring sign would be an important finding (Hoffmann et al., 2006; Motoyama et al., 2009). If the CT value changes, the definition of the plaque could be vague. Additionally, noise would easily affect these relatively low values. Noise-less images that maintain the CT attenuation and resolution would thus meet the clinical demands of CCT.

Our results also showed the possibility of combining legato with IR. Recently, powerful noise reduction using IR methods has been reported (Renker et al., 2011; Utsunomiya et al., 2012; Yin et al., 2013; Meyer et al., 2014). The IR methods are mainly used to reduce the radiation dose in clinical settings. Whereas the IR methods are used in the image reconstruction stage, legato is performed in the post-processing stage. Thus, legato could be applied to the IR images and further reduce noise independently of the IR methods.

Another possibility of legato is taking the same noise images with a lower radiation dose. Even at a quarter dose, the CNR obtained using legato was not less than that of the reference image. In addition, our subjective image quality assessment supports this finding. The results are thus reasonable. Without legato, the noise level would be twice that of the reference image for a quarter radiation dose. However, with legato, the noise level can be reduced by half, as previously discussed. Our result theoretically suggests that a three-quarter reduction of the radiation dose is possible when we focus on the image noise level, but a larger investigation of clinical cases is needed to assess the diagnostic performance.

There were limitations to the present study. First, in the ex vivo study, a static phantom was used to investigate the potential of legato. However, the effect of motion and the analysis of temporal resolution need to be further investigated for the use of CCT in clinical cases. Second, the application of legato to other scanners, such as a single-source scanner, and with other reconstruction kernels and other IR methods needs to be investigated.

In summary, our proposed 4DNR method can be applied to data series having short interval times (legato) to reduce noise without changing the CT attenuation or spatial resolution. Thus, applying this post-processing method to the “padding” phases of CCT images would have a potential of theoretically achieving a 75% reduction at the maximum in the radiation dose while maintaining the noise level as that of the baseline radiation dose. However, further studies are needed for the clinical usage.