Performance of clinically available deep learning image reconstruction in computed tomography: a phantom study

Abstract.

Purpose: To assess the physical performance of deep learning image reconstruction (DLIR) compared with those of filtered back projection (FBP) and iterative reconstruction (IR) and to estimate the dose reduction potential of the technique.

Approach: A cylindrical water bath phantom with a diameter of 300 mm including two rods composed of acrylic and soft tissue-equivalent material was scanned using a clinical computed tomography (CT) scanner at four dose levels (CT dose index of 20, 15, 10, and 5 mGy). Phantom images were reconstructed using FBP, DLIR, and IR. The in-plane and $z$ axis task transfer functions (TTFs) and in-plane noise power spectrum (NPS) were measured. The dose reduction potential was estimated by evaluating the system performance function calculated from TTF and NPS. The visibilities of a bar pattern phantom placed in the same water bath phantom were compared.

Results: The use of DLIR resulted in a notable decrease in noise magnitude. The shift in peak NPS frequency was reduced compared with IR. Preservation of in-plane TTF was superior using DLIR than using IR. The estimated dose reduction potentials of DLIR and IR were 39% to 54% and 19% to 29%, respectively. However, the $z$ axis resolution was decreased with DLIR by 6% to 21% compared with FBP. The bar pattern visibilities were approximately consistent with the TTF results in both planes.

Conclusions: The in-plane edge-preserving noise reduction performance of DLIR is superior to that of IR. Moreover, DLIR enables approximately half-dose acquisitions with no deterioration in noise texture in cases that permit some $z$ axis resolution reduction.

1. Introduction

In recent years, a number of techniques have been proposed for the reduction of noise in computed tomography (CT) images to enable radiation doses to be decreased.^1–4 The use of non-linear algorithms such as iterative reconstruction (IR) and three-dimensional image filtering have facilitated low-dose CT image acquisition and have had a substantial impact on patient dose reduction. Recently, deep learning image reconstructions (DLIRs) have been introduced as novel noise reducing techniques for CT images. DLIRs are designed to suppress noise without impacting anatomical structures by using a deep convolutional neural network trained to differentiate signals from noise.⁵ Presently, two DLIR techniques are commercially available, and their beneficial performances have been reported in clinical studies.^6–10

Objective evaluation of image quality is a suitable method of interpreting the characteristics of noise reduction algorithms such as IRs. Because these involve non-linear processes, the image quality varies depending on the image contrast and radiation dose.^11–13 Moreover, DLIRs are presumed to involve non-linear processes in more complex functions. Task specific image quality metrics have been developed for evaluating the performances of such non-linear techniques;^14,15 these are recommended in current guidelines for the characterization of noise magnitude, noise texture, and resolution, all of which vary depending on image contrast and noise.¹⁶ However, to the best of our knowledge, the physical image characteristics of DLIRs have not been sufficiently investigated. Although some researchers have addressed these in-plane characteristics,^17–20 there have been no reports on the evaluation of $z$ axis resolution. Even if the in-plane performance of DLIR outperforms those of existing techniques, the performance may be pseudo-one when its edge preservation in the $z$ direction is inferior to those of the existing techniques. The inferiority in the $z$ direction causes quality degradation of the coronal and sagittal multiplaner reformation images that are frequently used in the CT image interpretation. Therefore, the evaluation of $z$ axis resolution is required to correctly address the noise reduction performance of DLIR.

The present study aimed to evaluate the physical performance of a DLIR [true fidelity imaging (TFI); GE Healthcare, Waukesha, WI, USA] at different dose levels and compare the results with those obtained using filtered back projection (FBP) and an existing IR technique. The potential for dose reduction offered by TFI based on the performance of FBP was estimated from the measured results to investigate whether DLIR could be an alternative reconstruction method to FBP and IR, which have been used in daily clinical practice.

2. Materials and Methods

2.1. Phantoms

We used a cylindrical water bath phantom with a diameter of 300 mm, which resembled the x-ray absorption of the adult abdomen, for performance evaluations. For task transfer function (TTF) measurements at in-plane resolution, we placed two rod-shaped objects with a 30-mm diameter and a 40-mm height into the cylindrical water bath phantom. The two rods comprised acrylic and a soft tissue-equivalent material [STEM-06 (custom-ordered phantom), Kyoto-kagaku, Kyoto, Japan], and their CT numbers were $\sim 130$ and 60 Hounsfield units (HU), respectively, at 120 kV. These objects were offset at 40 mm from the phantom center [Fig. 1(a)]. This phantom configuration was determined with reference to phantoms used for CT image quality measurements, such as the Mercury phantom (Gammex, Middleton, WI, USA) and the ACR phantom (Gammex, Middleton, WI, USA). The alignment of the rod with the phantom was adjusted such that each rod axis was completely parallel to the rotation axis of the CT system.

Fig. 1.

Diagrams of the phantom structures and representative images. Phantom structures and representative CT images used for the measurements to evaluate (a) in-plane resolution and (b) $z$ axis resolution. The regions of interest for NPS measurement are depicted (described in Sec. 2). For evaluation of $z$ axis resolution, a sagittal image was reformatted from axial images obtained by scanning the phantom for $z$ axis TTF. Multiple sagittal images were obtained from repeated scanning and averaged to achieve a high CNR ($>25$).

For measuring $z$ axis TTF, a larger rod-shaped object (70-mm diameter and 50-mm height) made of acrylic or the soft tissue-equivalent material was instead placed centrally in the water bath phantom. During scanning, the phantom was tilted by $\sim 4\mathrm{deg}$ with respect to the rotation axis to obtain sagittal reformation images presenting as slanted edge images [Fig. 1(b)]. The detailed processing method for obtaining low-noise sagittal images is described in the $z$ axis resolution subsection.

We used these custom phantom combinations because rod objects can be positioned freely in the phantom depending on measurement purposes. Moreover, any material can be used for the objects. To the best of our knowledge, there is no phantom with the diameter corresponding to the adult abdomen including objects for the $z$ axis resolution measurement.

2.2. Data Acquisition and Image Reconstruction

Phantom scanning was conducted using a clinical CT scanner (Revolution CT; GE Healthcare) equipped with reconstruction hardware for TFI. Moreover, an IR technique (adaptive statistical iterative reconstruction, ASiR-V) is implemented in this scanner. TFI has three noise reduction strengths: low, middle, and high; ASiR-V yields images with FBP-IR blending at percentages ranging from 0% to 100% (100% offers the strongest noise reduction).²¹ For the present study, we reconstructed phantom images by FBP; TFI at low, middle, and high strengths (hereafter referred to as DLIR-L, DLIR-M, and DLIR-H, respectively); and ASiR-V with 40% and 80% blending (hereafter referred to as IR40% and IR80%, respectively). A “standard” kernel was used for all reconstructions. The scan parameters were as follows: 120 kV, 0.5 s per rotation, spiral pitch: 0.508, coverage: 80 mm ($128\text{rows}\times 0.625\mathrm{mm}$). Tube current settings of 300, 225, 150, and 75 mA were set to obtain CT dose index (${\mathrm{CTDI}}_{\mathrm{vol}}$) values of 20, 15, 10, and 5 mGy, respectively. The CT images were reconstructed with a slice thickness of 1.25 mm without a slice gap and a display field of view of 30 cm. The slice gap was set half of the slice thickness (i.e., 0.625 mm) only for the data of $z$ axis resolution measurement to increase the number of data sampling. We used 20 mGy as a reference dose to estimate dose reduction potentials. This radiation dose value was determined according to the abdominal CT dose level in the ACR Dose Index Registry (20 mGy for the 75th percentile).²²

2.3. In-Plane Resolution and Image Noise

An in-plane TTF was measured using images of the rods as an index of task-specific resolution. To improve TTF accuracy, the phantom was repeatedly scanned to obtain an ensemble average image with a contrast-to-noise ratio (CNR) of more than 25.^23,24 Preliminarily, we confirmed that the repeated scanning did not lead to any resolution loss. An edge spread function (ESF) was generated from the averaged rod image using the circular edge method described by Richard et al.²⁵. The derivative of the ESF provided the line spread function (LSF), and then, the TTF was computed from a discrete Fourier transform of the LSF.

The noise power spectrum (NPS) was analyzed through an established method using the 2D fast Fourier transform from three regions of interest of $128\times 128\text{pixels}$^26,27 [Fig. 1(a)]. The 2D NPS measurements were radially averaged and binned into 40 frequency bins; 80 consecutive axial slices obtained in one scan were averaged to reduce variability. The area under the curve and peak frequency of the NPS were calculated as indices of noise magnitude and noise texture characteristics, respectively.

2.4. System Performance Function and Dose Reduction Potential

To compare the overall performance of reconstructions, we estimated the system performance function (SPF),^27–33 which was defined using the modulation transfer function (MTF) and the NPS as follows:

$${\mathrm{SPF}}^2(f)=\frac{{\mathrm{MTF}}^2(f)}{\mathrm{NPS}(f)},$$ (1)

where $f$ denotes the spatial frequency assuming radial symmetry (i.e., ${[u^2+v^2]}^{1/2}$, where $u$ and $v$ represent spatial frequencies in the $x$ and $y$ directions, respectively). In this study, TTF was used as MTF, as used in previous studies.^2,13,33 Although this metric is based on the ideal observer model for imaging systems with linearity, it has been exploited for IR images and non-linear image-based techniques. Furthermore, its value in the proximity of zero frequency is adequate to evaluate system sensitivity, similar to noise equivalent quanta of CT, where the SPF value can be a surrogate to the relative number of quantum contribute to an image.^27,30 We estimated potential dose reduction capabilities between FBP and noise reduction algorithms (i.e., DLIR and IR) using the ${\mathrm{SPF}}^2$ value at $0.02{\mathrm{mm}}^{-1}$, which was the lowest frequency achieved in our calculations. That value of the FBP protocol at ${\mathrm{CTDI}}_{\mathrm{vol}}$ of 20 mGy was set a reference value. The magnitude of dose required to achieve that same ${\mathrm{SPF}}^2$ value with DLIR and IR was assessed. The ratio of the difference between the two dose levels determined the percentage–dose reduction potentials.

2.5. z Axis Resolution

As shown in Fig. 1(b), a sagittal image stack was acquired within the width of the rod object, and then, images were averaged into one sagittal image to reduce noise. The CNR of the averaged image for one scan did not reach the required CNR of $>25$, which was the reference level mentioned earlier;^23,24 therefore, phantom scanning was repeated three, four, five, and eight times at 20, 15, 10, and 5 mGy, respectively, to obtain averaged sagittal images. A synthetic edge profile was generated by plotting pixel intensity of the slanted edge against the distance from the edge plane, and then, the derived ESF was differentiated to yield the task specific slice sensitivity profile (SSP). Moreover, the $z$ axis TTF was calculated from the obtained SSP using a discrete Fourier transform and subsequent normalization using a Fourier coefficient at zero frequency.

2.6. Bar Pattern Observations

The same two bar pattern phantoms composed of acrylic were instead placed inside the water bath phantom with an orthogonal relationship so that one was parallel to the slice plane (in-plane bar pattern), and the other was parallel to the coronal plane. The bar pattern phantom consisted of six segments with bar sizes of 0.5 to 5.0 mm with corresponding square wave frequencies of 1.0 to $0.1{\mathrm{mm}}^{-1}$. The edge-preserving noise reduction performances were visually checked from axial images of the slice plane bars and coronal multiplanar reformation images of the coronal plane bars, respectively. This aimed to assess the shape and edge preservation performance of the phantom with more complicated structures than the simple rod shape that was used for TTF measurements.

3. Results

3.1. Resolution

Figure 2 shows the in-plane and $z$ axis TTFs from the acrylic and soft tissue contrast materials. Graphs at 10 mGy are presented as representatives. The 50%TTFs for all dose levels are summarized in Table 1. For in-plane resolution, all levels of DLIR improved the 50%TTF of images obtained from the acrylic contrast, and the percentage increases compared with FBP were 9% to 17%, 9% to 15%, and 8% to 13% for DLIR-L, DLIR-M, and DLIR-H, respectively. The TTFs for the three strengths of DLIR with the soft tissue contrast were almost equivalent to those of FBP. By contrast, IR tended to reduce the in-plane TTFs, with the reduction rate being more obvious with the soft tissue contrast compared with the acrylic contrast, with IR80% compared with IR40%, and with decreased dose.

Fig. 2.

Graphs show the (a), (b) in-plane and (c), (d) $z$ axis TTFs for (a), (c) acrylic and (b), (d) soft tissue contrast at a ${\mathrm{CTDI}}_{\mathrm{vol}}$ of 10 mGy.

Table 1.

In-plane and $z$ axis 50% TTFs for all dose levels.

${\mathrm{CTDI}}_{\mathrm{vol}}$ (mGy)	50% TTFs with acrylic contrast (${\mathrm{mm}}^{-1}$)						50% TTFs with soft tissue contrast (${\mathrm{mm}}^{-1}$)
	FBP	DLIR low	DLIR middle	DLIR high	IR 40%	IR 80%	FBP	DLIR low	DLIR middle	DLIR high	IR 40%	IR 80%
In-plane TTF
20	0.37	0.40	0.40	0.40	0.34	0.31	0.35	0.36	0.36	0.35	0.29	0.24
15	0.36	0.40	0.40	0.39	0.33	0.30	0.37	0.37	0.36	0.35	0.29	0.23
10	0.35	0.40	0.39	0.38	0.30	0.26	0.35	0.36	0.35	0.34	0.28	0.23
5	0.32	0.37	0.37	0.36	0.27	0.23	0.31	0.33	0.32	0.30	0.25	0.21
$z$ axis TTF
20	0.31	0.29	0.29	0.28	0.30	0.28	0.31	0.27	0.26	0.26	0.28	0.25
15	0.32	0.29	0.29	0.28	0.29	0.27	0.32	0.28	0.27	0.26	0.28	0.25
10	0.31	0.29	0.28	0.28	0.29	0.26	0.31	0.28	0.27	0.26	0.28	0.25
5	0.30	0.29	0.28	0.27	0.27	0.24	0.30	0.26	0.25	0.24	0.27	0.24

In contrast, the $z$ axis TTFs were reduced by both DLIR and IR compared with FBP. The percentage decreases of the 50%TTF for DLIRs were 6% to 12% for the acrylic contrast and 10% to 22% for the soft tissue contrast. The 50%TTFs for IR40% was slightly higher than that for DLIR. For IR80%; the 50%TTFs for the acrylic contrast at frequencies of $<0.5{\mathrm{mm}}^{-1}$ and for the soft tissue contrast were similar to those of DLIR. Through IR, TTFs for acrylic contrast increased at frequencies of $>0.5{\mathrm{mm}}^{-1}$.

3.2. Noise

The area under the curve of the NPS that corresponds to noise magnitude was notably decreased with both DLIR and IR compared with FBP at all dose levels (Table 2). The largest reduction was observed in the case of DLIR-H, followed by IR80%, DLIR-M, DLIR-L, and IR40%. Figure 3 illustrates the NPS at 10 mGy; DLIRs reduced the NPS at all frequencies; particularly, the reduction at low-frequency noise was more remarkable than IRs. By contrast, noise reduction of IR predominantly occurred in the middle to high frequencies; the tendency was stronger for IR80% than for IR40%. The NPS peak frequency shifts toward lower frequencies compared with FBP were mostly negligible ($\le 0.02{\mathrm{mm}}^{-1}$) with DLIR at all dose levels except for DLIR-H at 5 mGy, which exhibited a shift of $0.06{\mathrm{mm}}^{-1}$. The IR40% caused relatively larger shifts ($\le 0.04{\mathrm{mm}}^{-1}$) compared with DLIR. Notable shifts ranging from 0.08 to $0.12{\mathrm{mm}}^{-1}$ were observed using IR80%.

Table 2.

Noise magnitude and peak spatial frequencies of NPS for all reconstruction methods and dose levels.

${\mathrm{CTDI}}_{\mathrm{vol}}$ (mGy)	Area under the curve of NPS (${\mathrm{HU}}^2{\mathrm{mm}}^2$)						NPS peak frequencies (${\mathrm{mm}}^{-1}$)
	FBP	DLIR low	DLIR middle	DLIR high	IR 40%	IR 80%	FBP	DLIR low	DLIR middle	DLIR high	IR 40%	IR 80%
20	265	103	67	41	129	56	0.24	0.23	0.23	0.22	0.22	0.15
15	346	133	90	55	182	74	0.23	0.22	0.22	0.22	0.20	0.11
10	500	196	135	83	259	112	0.23	0.22	0.22	0.22	0.19	0.11
5	776	370	261	160	410	191	0.21	0.21	0.19	0.15	0.19	0.11

Fig. 3.

NPSs for all reconstruction methods at a ${\mathrm{CTDI}}_{\mathrm{vol}}$ of 10 mGy.

3.3. System Performance Function and Dose Reduction Potential

Figure 4 depicts the ${\mathrm{SPF}}^2$ values at 10 mGy. Similar to FBP, the ${\mathrm{SPF}}^2$ values for DLIR and IR40% gradually decreased with increasing frequency. The difference compared with FBP increased at higher frequencies, except in the case of IR40% with soft tissue contrast. The curve shapes of IR80% were different from all others; particularly, the curve for the acrylic contrast was characteristic, where the ${\mathrm{SPF}}^2$ values were almost constant at frequencies of $>0.2{\mathrm{mm}}^{-1}$ unlike the typical SPF curves that monotonically decreased with increasing the spatial frequency.

Fig. 4.

Graphs show results squared SPFs for (a) acrylic and (b) soft tissue contrasts for all reconstruction methods.

Figure 5 illustrates the variation of ${\mathrm{SPF}}^2$ values at the lowest frequency ($0.02{\mathrm{mm}}^{-1}$) as a function of ${\mathrm{CTDI}}_{\mathrm{vol}}$. Overall, DLIR outperformed IR. The average increases in ${\mathrm{SPF}}^2$ compared with FBP were 65%, 91%, 144%, 32%, and 50% for DLIR-L, DLIR-M, DLIR-H, IR40%, and IR80%, respectively. The estimated average dose reduction percentages for the two contrasts were 39%, 47%, 54%, 19%, and 30% for DLIR-L, DLIR-M, DLIR-H, IR40%, and IR80%, respectively.

Fig. 5.

Graphs show results the SPF2 values at the lowest frequency ($0.02{\mathrm{mm}}^{-1}$) as a function of ${\mathrm{CTDI}}_{\mathrm{vol}}$ for (a) acrylic and (b) soft tissue contrasts. Each red dashed line shows the reference level (SPF2 of FBP at 20 mGy).

3.4. Phantom Visibility

Figure 6 presents the bar pattern phantom images acquired in the axial and coronal planes for FBP at 20 mGy as well as DLIR-M at 10 mGy because its dose reduction percentage was estimated to be $\sim 50\%$. Additionally, IR80% images at 10 mGy were used for comparison. In the axial plane, DLIR-M images at 10 mGy (the half dose) showed notable noise reduction while maintaining the noise texture of the FBP. The DLIR-M preserved the in-plane bar visibilities at $0.25{\mathrm{mm}}^{-1}$ and less (bar widths of $\ge 2.0\mathrm{mm}$). However, slight blurring occurred at $\ge 0.5{\mathrm{mm}}^{-1}$ (bar widths of $\le 1.0\mathrm{mm}$), which was not sufficiently consistent with the TTF for the acrylic contrast. In the coronal plane, resolution preservation was not achieved by DLIR, in agreement with the $z$ axis TTFs. The IR80% images showed notable changes in noise texture and blurring in both in-plane and $z$ axis direction, consistent with the TTF for the acrylic material.

Fig. 6.

Bar pattern images. Images show comparisons of (a), (d) FBP at 20 mGy (the reference dose) and (b), (e) DLIR-M and (c), (f) IR80% at 10 mGy. The upper (a)–(c) and bottom (d)–(f) rows present images acquired in the axial and coronal planes, respectively. The texts embedded in (a) show square wave frequencies of the respective segments. All window widths/levels were set to 300/30.

Figure 7 shows the bar pattern phantom images for the low-dose situation (5 mGy). In the axial plane, the bar visibilities of DLIR-M images were better compared with FBP. However, the disrupted shapes of the bars in the FBP image caused by high noise persisted in the DLIR-M image. Though IR80% reduced the noise, the bar sharpness deteriorated compared with FBP, corresponding to the in-plane TTF results. In the coronal plane, the bar sharpness deteriorated with DLIR-M compared with FBP. The bar sharpness of DLIR-M also seemed to be inferior to IR80%, in line with the $z$ axis results. Persistence of disrupted shape was also observed in coronal plane images.

Fig. 7.

Bar pattern images. Images present reconstructions using (a), (d) FBP; (b), (e) DLIR-M; and (c), (f) IR80% in the low-dose condition of 5 mGy. The upper (a)–(c) and bottom (d)–(f) rows present images acquired in the axial and coronal planes, respectively. The red arrows highlight the disrupted shapes of the bars. The texts embedded in (a) show square wave frequencies of the respective segments. All window widths/levels were set to 300/30.

4. Discussion

This study compared the image characteristics of DLIR with those of FBP and IR through TTF and NPS measurements, including $z$ axis spatial resolution, which has not been performed in the previous studies. Overall, the in-plane resolution of DLIR was superior to that of IR. Compared with FBP, DLIR improved the TTF for the acrylic contrast and preserved the TTF for the soft tissue contrast. The noise magnitudes of DLIR at all three strengths at 10 mGy were lower than that of FBP at 20 mGy; moreover, even at 5 mGy, DLIR-H presented a noise magnitude lower than that of FBP at 20 mGy. Although the noise magnitude of IR80% at 10 mGy was significantly lower than that of FBP at 20 mGy, its performance was lower than that of DLIR-H. Furthermore, the peak NPS frequency shifts of DLIR-L and DLIR-M relative to FBP were almost negligible at all doses ($\le 0.02{\mathrm{mm}}^{-1}$), which were smaller than those of IR. These attributes of DLIR are similar to the previous reports.^17,19,20 Our in-plane results were consistent with the DLIR’s development design to differentiate noise from signals in order to reduce reconstructed image noise without changing its texture.

Conversely, the $z$ axis resolution for each contrast was lower using DLIR than FBP. The performance of DLIR was not necessarily superior to that of IR80% that presented noise magnitudes comparable to those of DLIR. The insufficient $z$ axis edge preservation of DLIR reflects the difficulty of completely accurate differentiation between noise and signals in both the axial plane and $z$ direction, suggesting that DLIR did not work well as expected in the $z$ axis direction. It is well known that $z$ axis filtering (a broadening of slice thickness) results in in-plane NPS reduction over the entire frequency range.^30,34 On FBP (linear) images, the in-plane noise is nearly inversely proportional to the square root of the full-width at half-maximum of SSP. Therefore, it was presumed that some processes in DLIR that act as $z$ axis filtering contributed to the in-plane NPS reduction observed in our results; however, analysis of the $z$ axis filtering process is beyond the scope of this study.

Our experiments using bar pattern images for the acrylic contrast revealed image characteristics approximately consistent with the measurement results. The notable decrease in noise magnitude and the negligible NPS peak shift were able to be subjectively verified with the bar-pattern images. The visibility of in-plane bar pattern phantom was consistent with the comparisons of SPF that of DLIR-M at 10 mGy was equivalent to FBP at 20 mGy. These results suggest that DLIR-M can facilitate an $\sim 50\%$ dose reduction protocol. At the same low-dose (5 mGy), although the bar pattern visibility was improved by DLIR, the disrupted shapes of the bars in the DLIR-M image were unfavorable. For coronal bar pattern images, the worse TTF results of DLIR were reflected in the blurred bar patterns. The bar edge sharpness appeared to be improved using IR compared with DLIR, suggesting that this effect is due to the higher $z$ axis TTFs of IR compared with DLIR for the acrylic contrast.

Deep learning reconstruction intelligently differentiates noise from signals in CT images using filtering processes that have been optimized through deep learning operations.⁵ The effects of optimized process can be seen in our results, which demonstrate the improved in-plane TTF preservation and negligible changes in noise texture (i.e., NPS shift). Although $z$ axis TTFs were somewhat reduced by DLIR, we consider this to be a cost of the sufficient in-plane edge-preserving noise reduction, resulting from the difficulty in achieving complete separation of noise from signals. Moreover, the disrupted bar shapes in the images acquired at 5 mGy also reveal the difficulty in achieving suitable noise reduction. Recovery of object edge shape is ideally required for noise reduction; however, in DLIR images, the disrupted bar shapes of the original FBP images were preserved.

This study has some limitations that should be acknowledged. First, we evaluated the in-plane and $z$ axis TTFs for only two contrasts of 130 and 60 HU, obtained by the acrylic and soft-tissue equivalent material. Contrasts other than 130 and 60 HU should be investigated for evaluating the performance of DLIR on different clinical tasks. Second, compared with real anatomical structures, the object shape was a simple rod, and the acrylic bar pattern phantom does not also have complicated shapes. Thus the TTF results may not sufficiently represent the spatial resolutions of clinical images, and the bar pattern image might be an insufficient model. It is expected that more robust methods will be developed for evaluating the spatial resolution in clinically relevant models. Third, although we evaluated the $z$ axis TTF, we did not evaluate the $z$ axis NPS. The dose reduction potential of DLIR was estimated considering the non-negligible effects of $z$-directional edge blurring. Detailed evaluations using three-dimensional SPF are required to accurately estimate the dose reduction potential. Finally, we evaluated the DLIR’s negligible noise texture change as a suitable property based on the FBP-like images that are preferred by the radiology community. In a previous report, however, a strange-looking image caused by noise reduction processing did not influence diagnostic performance and instead improved low-contrast detectability.³⁵ A validation study using actual clinical images should be conducted by radiologists.

5. Conclusions

Our evaluations using TTF and NPS demonstrate that DLIR offers considerably better performance for in-plane edge-preserving noise reduction than IR. In addition, the change in noise texture evaluated by NPS peak frequency is negligible in DLIR images. However, we observed a slight but non-negligible reduction in $z$ axis resolution in DLIR images. Taking into account the in-plane ($x\text{–}y$) and $z$ axis performance characteristics, DLIR can potentially enable approximately half-dose acquisitions in cases that permit some $z$ axis resolution reduction.

Biographies

Hiroki Kawashima is an assistant professor at Kanazawa University, Japan. His research interests include x-ray computed tomography and digital radiography, image quality assessment, and optimization of imaging technique.

Biographies of the other authors are not available.

Disclosures

No conflicts of interest, financial or otherwise, are declared by the authors.