Assessment of Image Quality in Chest CT With Precision Matrix and Increased Framing Rate Using Single Source CT: A Phantom Study

Abstract

Background/Aim: The aim was to evaluate the effect of a combined precision matrix and high sampling rate on the delineation of anatomical structures and objective image quality in single source CT in a qualitative approach.

Materials and Methods: An anthropomorphic thoracic phantom was used to evaluate the objective image quality parameters, including image noise, noise power spectrum, image stepness and Q for different CT scanners including high/ standard matrix and framing frequency setups. Scan parameters were standardized over all scanners. Additional subjective quality assessment was also performed.

Results: A linear mixed effects model was used to determine the effect of sampling rate and image matrix on objective image quality parameters. Noise power spectrum and image noise were significantly influenced by both framing frequency and image matrix. There were significant differences between high and standard frequency/ matrix acquisitions.

Conclusion: Higher framing frequency and image matrix allows for improved image noise texture and objective image quality in CT.

Similarly to other modalities, spatial resolution in CT represents the ability to distinguish and identify small, closely-neighboring objects on an image. High spatial resolution is crucial for proper assessment and correct identification of sub-millimetric anatomical structures and pathologies in the lung parenchyma. High resolution computed tomography (HRCT) is the most commonly used cross-sectional imaging technique for the evaluation of interstitial lung disease (ILD) (1). The spatial resolution of a CT system is predominantly determined by hardware restrictions such as e.g., number of detectors and focal spot size (2). Since the CT image reconstruction is achieved mathematically, the reconstruction algorithm and the matrix size can further impact spatial resolution. Fine high-contrast detail kernels are used to obtain high spatial resolution of lung parenchyma. Usually, the image matrix is predefined in an institutional setup and fixed at a standard of 512×512 pixels. When reconstructing images with a large field of view (FOV) covering the chest, the spatial resolution is reduced with increasing FOV and fixed image matrix. The introduction of larger image matrices may consequently improve image quality (3). Typically, spatial resolution degrades in the outer FOV in systems with lower sampling rates. An increase of the sampling rate may therefore substantially impact the image quality for scans with fast resolution (4,5). This is of utter importance for lung and trauma imaging as fast rotation and large FOV are key. Although focal spot size does affect CT spatial resolution, CT resolution is generally limited by the size of the detector measurements (referred to as the aperture size) and by the spacing of detector measurements used to reconstruct the image (6). Another way to enhance spatial resolution is to improve the sampling of detector units by deflecting the focal spot on the x-ray tube anode along the longitudinal and fan angle direction (7). CT detectors have a fixed sampling or reading rate, that is, the speed at which data can be processed and transferred from the detector to the image reconstruction system. At a given sampling rate, the amount of data, i.e., the number of projections available for a dedicated rotation decreases for faster rotation times. For example, with the traditional sampling rate of 4 kHz, a detector can process ~2,000 projections in one rotation when operating with a rotation time of 0.5 seconds, while the number of projections is reduced to 1,000 in one rotation when operating at a rotation time of 0.25 s. This influences the (azimuthal) image resolution, especially in regions away from the isocenter, where the overlap of projections is smaller, with higher reading sampling allowing for better resolution. Our prototype scanner used for phantom evaluation offers a very fast rotation of 0.25 s that allows scan speeds of up to 261 mm/s. Its detector has been designed with a sampling rate of 8 kHz, in order to maintain image quality and spatial resolution in procedures focused on speed, like cardiac and lung visualizations or emergencies, even for large patients and across the whole FOV.

The aim of our phantom study is to evaluate the impact of this new scanning technology with a higher framing rate and increased image matrix on the objective and subjective image quality.

Materials and Methods

Phantom description. An anthropomorphic phantom of the male thorax (Lungman, Kyoto Kagaku, Tokyo, Japan) that has been used before in the literature was used for analysis (8). The standardized anatomical model in life size was built up of an artificial thoracic wall and an artificial inlay including the heart, diaphragm, the mediastinum and both lungs with the pulmonary vessels (Figure 1). The soft tissue substitute materials were made of polyurethane resin composites and the synthetic bones were made of epoxy resin with X-ray absorption rates very close to those of human tissue. The space between the pulmonary vessels in the thoracic cavity consisted of air (8).

Figure 1. Anthropomorphic thoracic phantom with artificial thoracic cavity and inlay with lungs and pulmonary vessels.

Scan protocol. A newest generation multidetector single source CT (SOMATOM x.ceed, Siemens, Forchheim, Germany) was used for image acquisition (A). The phantom was imaged in single-energy mode using a low-dose protocol, as proposed for lung cancer screening (9), with z-flying focal spot using two different radiation doses by altering the tube current time product (mAs of 80 and 20). For both acquisitions, a high data sampling rate was used, as proposed by the manufacturer (8,064 readings/s). Tube rotation time was set at 1, 0.5 and 0.25 s. The same pitch of 1.5 was used for all acquisitions.

The same acquisitions were performed on a multidetector single source CT and a dual source CT, both with standard data sampling rate of 4032 readings/s [SOMATOM drive (B), mBiography (C), SOMATOM definition AS+ (D), Siemens]. Details of the scan parameters are summarized in Table I. All acquisitions on scanner A were performed twice, one with an empty phantom cavity and one with a complete vessel inlay.

Table I. Detailed computed tomography (CT) scan parameters for scanners A-D used in phantom evaluation.

All CT images (Scanner A-D) were reconstructed with a slice thickness of 1 mm and increment of 1 mm with 512 pixel matrix using sharp reconstruction kernels (Br62, Br64 and I70 respectively). The transfer of reconstruction sharpness from SAFIRE to ADMIRE was performed as proposed by the manufacturer. The FOV was kept constant with 316 mm for all datasets. For Scan A, an additional 1,024 pixel matrix (*) was used for reconstructions.

Image analysis. Images were loaded onto a clinical viewing workstation and displayed in a 2×3 fashion with the Scan A (512/1) images (clinical reference) located in the upper left panel. The remaining acquisitions were randomly assigned to the other panels.

Subjective image quality. Two radiologists (SM and TN with 2 and 17 years of experience, respectively), blinded to the CT acquisition and reconstruction method of the randomly assigned images, were asked to compare the randomized images to the reference 512/1 images. The readers were free to enlarge the images for better visualization of the structures. The radiologists were asked to assess their ability to visualize the morphologic features of the vessels (1). Depiction of vessels was assessed on a 5-point Likert-scale (−2=definitely worse, probable decreased ability to see small vascular structures; −1=definitely worse, unclear effect on potential diagnosis; 0=about the same or unclear benefit/decrement; +1=definitely better, unclear effect on potential diagnosis; +2=definitely better) (1,10). The score was not given to the reference images.

Objective image quality. Image noise was measured for each acquisition and reconstruction method by placing a circular region of interest in the center of the empty phantom at the same anatomic site for all measurements. The standard deviation of this measurement was reported to represent image noise. Image sharpness was assessed according to a previously established method (11,12). The attenuation profile was measured along perpendicular lines to the left paravertebral stripe of the phantom using an open-source software (Fiji ImageJ ver. 2.1.0/1.53c) (13). For each dataset the attenuation profile was measured at a standardized position at the 5th thoracic spine vertrebral body. The line width was 10 pixels, and the line length 20 mm. Image sharpness was defined as maximum steepness of the attenuation profile in CT numbers per mm (ΔCTmm) (3) (Figure 2).

Figure 2. Estimation of image sharpness, defined as maximum steepness of the attenuation profile in CT numbers per mm (ΔCT_mm) using Fiji ImageJ ver. 2.1.0/1.53c.

Additionally, the normalized noise power spectrum (nNPS) was assessed as an objective image quality parameter as proposed before (14,15). NPS describes the spatial frequency content of image noise and was measured for each combination of radiation dose level, rotation time and reconstruction setting based on standard methods. The nNPS was analyzed for each dataset using four 32 × 32 square regions of interest (ROIs) placed at different positions in the uniformly aerated empty phantom. Measurements were performed in 60 slices in each dataset (4 ROIs per slice × 60=240 ROIs) using an open-source software (imQuest ver. 1, Duke University, Durham, NC, USA). NPS gives noise power within a ROI as a function of spatial frequency (14). NPS peak frequency (f_peak) served as a scalar metric for comparing noise texture across different imaging conditions. A higher fpeak implies noise with finer texture, and a lower f_peak implies noise with grainier or coarser texture. To quantify the changes of magnitude and texture of the image noise, the average spatial frequency (f_av) of the NPS curve was measured. f_av values describe the overall frequency content of the NPS (10).

As radiological image quality is usually evaluated using noise, contrast and spatial resolution, in order to assess the whole CT image acquisition process, one must also consider specific CT acquisition parameters, such as slice thickness and CT dose index (CTDI). Volumetric computed tomography dose index (CTDI_vol) is a well-known measure of radiation dose in computed tomography (16) and is shown on the CT console while making scans. From the Rose theory, it is possible to build a figure of merit (Q). In a single quantitative index, several quantitative parameters are included, defined and present in different ways, including the Contrast-to-Noise-Ratio (CNR) and the dosimetric quantity (CTDI_vol) (17). Q was defined as CNR²/CTDI_Vol, as described before (17,18), where CNR is contrast to noise ratio. Contrast was evaluated as the difference between mean CT numbers in standardized ROIs of 2.5-3 cm³ measured in the thoracic spine in T5 and in the basal central part of the phantom. Noise is mean noise in the considered details, evaluated as standard deviation in the same ROIs. Higher Q values indicate higher image quality relative to radiation dose.

Statistics. The ability to visualize the pulmonary vessels was compared among the different acquisitions and reconstructions using the Wilcoxon signed rank test. Statistical significance was defined as p<0.05. Objective image quality was assessed using a linear mixed effects model to determine the effect of sampling rate and image matrix on the image sharpness, noise power spectrum and Q.

Results

For assessment of objective image quality, data were assessed for normality and an analysis of variance (One-way ANOVA) was performed separately for the variables of noise, f_peak, f_av, Q, and steepness. In all models, the scan mode was included in the model as a batch variable. There were significant differences between the scanners in all 5 tested variables (Figure 3).

Figure 3. Boxplots for the different scanners, for the variables of noise, peak frequency (f_peak), average frequency (f_av), Quality index (Q) and steepness. The scan modes were combined for the boxplots. The box corresponds to the area where 50% of the data are located. The line represents the median. Points are marked as outliers. Reconstructions with 1,024 pixel matrix are marked with (*).

Image noise (represented by measurement of standard deviation of attenuation) was significantly influenced by both the sampling rate (effect size=0.131, p=0.049) and matrix size (effect size=0.15, p=0.034). Noise power spectrum was significantly influenced by sampling rate (effect size=0.309, p=0.001 for f_peak and effect size=0.42, p<0.001 for f_av). Matrix size showed no significant influence. Mean f_peak/f_av was 0.7±0.02/0.63±0.01 for high sampling rate and 0.57±0.11/0.54±0.07 for low sampling rate acquisitions, respectively (Figure 4).

Figure 4. Estimation of noise and noise texture using ImQuest ver. 1. Standardized squared regions of interests were measured in 60 slices each.

Pixel size was 0.69 mm for the 512 matrix and 0.35 mm for the 1024 matrix, respectively. The linear mixed effects model identified image sharpness (represented by the attenuation profile as the increase in CT number per pixel size) to be significantly influenced by sampling rate (effect size=0.35, p<0.001). There was no significant influence of matrix size (effect size=0.36, p=0.553). There was no significant influence of sampling rate (effect size=0.2, p=0.453). There was no significant influence of sampling rate (effect size=0.2, p=0.453) on objective image quality Q, nor of matrix size (effect size=0.33, p=0.051), even though a trend can be identified for the latter, which was very close to reaching statistical significance.

In the assessment of subjective image quality there was no significant difference in the highest-order bronchus visible in the upper, middle, and lower lobe of the right lung, respectively, between the 1,024- and 512-matrix (p=0.102).

Discussion

Our study demonstrated significantly improved noise texture for scan acquisitions with increased sampling rate for both image matrices used (512 and 1,024 pixels respectively). As previously published, a default image matrix of 512 pixels may limit the spatial resolution, depending on the resolution of the reconstruction kernel and the size of the field of view. At a given number of pixels, the volumetric image pixel is increasing with larger fields of view (3). If the kernel is sufficiently hard and the FOV is large enough, the volumetric image pixel will be smaller with a higher matrix which results in a visually sharper image but also increases the image noise (19). The effect of increasing image noise in larger image matrices has been evaluated before (20). This is in line with our results that showed the same correlation between enlarged matrix size and increasing image noise. FOV and the reconstruction kernel were kept constant over all acquisitions. In our study, the matrix size significantly influenced image noise over all dose variations, as has already been published by Euler et al. and Hata et al. (3,20).

The objective image quality Q was not significantly influenced by sampling rate of the scanner for the higher dose acquisitions (effect size=0.202, p=0.142), whereas the effect was significant for the dose reduced acquisitions (effect size=0.424, p=0.028. Image sharpness was inversely correlated with the sampling rate. This is in line with the analysis of noise power spectrum that showed an opposing trend. Increasing f_peak implies finer noise texture and fav represents the overall frequency content. A larger image matrix results in less stepness compared to coarser image noise texture and frequency content that yields higher differences in attenuation per mm for the same object imaged. Even if objective image quality analysis showed no significant differences between the 512 and 1,024 pixel matrix, the subjective differences were obvious and clearly in favor of the 1024 matrix (Figure 5). We contribute this to the choice of subjective scoring system that focused on global depictability of vascular structures and the diagnostic implication and less on obvious differences of spatial resolution, as has been observed before (3).

Figure 5. Assessment of objective image quality of the upper lobe. Left: 1024 matrix with finer details, right: 512 matrix with coarser noise texture and higher stepness.

In the literature, several studies have been performed concerning image quality in complex CT acquisitions (21-23), but without focusing on obese patients. Our study has several limitations. Firstly, the chest phantom was placed on the scan table in supine position without bearing aids, so slight differences in the positioning between different scanners are possible and that might have a minor impact on image quality. Secondly, since technology is evolving over time, we included scanners with different software surfaces and different reconstruction kernels (e.g., SAFIRE vs. ADMIRE). We used vendor given recommendations to harmonize the kernels used and to correlate the same sharpness of kernels available. Effects on image quality as given by closely adjacent kernel variance have shown to have minor effects on objective assessments (19).

Conclusion

In conclusion, in our study we could clearly demonstrate the beneficial effects of the newest sampling rate scanners on objective image quality parameters, such as noise spectrum analysis and global image noise. Furthermore, formerly described effects of increased image matrix on image noise and image sharpness could be supported. High resolution applications, such as chest CT with fine reticular structures may largely benefit from these new techniques. Sharper details may also increase image quality in musculoskeletal CT imaging.

Conflicts of Interest

JFL and FK are employees of Siemens Healthcare GmbH. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. No funding or grant support was received for this study. Siemens Healthcare provided support in the form of salaries for author JFL and FK but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Authors’ Contributions

Conceptualization, MS, EGM, KRA, HL, NT; methodology, MS, EG, KF, LF, NT.; software, KF, FJ; formal analysis, MS, EGM, NT; resources, KRA; data curation, MS, NT; writing – original draft preparation, MS, NT; writing – review and editing, MS, EGM, KRA, KF, JF, HL, NT, EA; visualization, MS, NT; supervision, NT, KRA. All Authors have read and agreed to the published version of the manuscript.

Acknowledgements

SM received a scientific grant from Guerbet AG, Switzerland. We thank Lars Bosshard of NEXUS Personalized Health Technologies, ETH Zürich, and the Swiss Institute for Bioinformatics, Zürich, for support with statistics.