Quality evaluation of image‐based iterative reconstruction for CT: Comparison with hybrid iterative reconstruction

Abstract The purpose of this study is to evaluate the physical image quality of a commercially available image‐based iterative reconstruction (IIR) system for two object contrasts to resemble a soft tissue (60 HU) and an enhanced vessel (270 HU), and compare the results with those of filtered back projection (FBP) and iterative reconstruction (IR). A 192‐slice computed tomography (CT) scanner was used for data acquisitions. IIR images were processed from the FBP images. Task‐based in‐plane transfer function (TTF) and slice sensitivity profile (SSPtask) were measured from rod objects inside of a 25‐cm diameter water phantom at four dose levels (2.5, 5, 10, and 20 mGy). Noise power spectrum (NPS) was measured from the water‐only part. System performance (SP) function was calculated as TTF2/NPS over FBP, IR, and IIR for comparison. In addition, an image subtraction was performed using images of rod objects, a bar‐pattern phantom, and a clinical abdomen case to observe the noise reduction performance of IIR. As a results, IIR mostly preserved TTF and SSPtask of FBP, whereas IR exhibited enhanced TTF at 10 and 20 mGy for 60 HU contrast and at all doses for 270 HU contrast. SP of IIR at 2.5, 5, 10 mGy (half doses) were similar to those of FBP at 5, 10, 20 mGy, respectively. IR exhibited enhanced SP at medium to high frequencies. The subtracted images showed weak remained edge signals in the bar‐pattern and abdominal images. In conclusion, IIR uniformly improved the task‐based image quality of FBP over the entire frequency range, whereas IR improved the characteristics over medium to high frequencies. The dose reduction potential of IIR estimated from SP is approximately 50%, when allowing the slight signal reductions.

The quality of IR images is affected by nonlinear characteristics, and spatial resolution varies according to noise level (related to radiation dose level) and target object contrast. 7,8 Thus, it is difficult to adapt conventional measuring methods to this type of images, because these methods assume FBP images with linear characteristics. Consequently, task-based methods that use specific object contrasts and radiation doses have been proposed to consistently evaluate the quality of images processed using IR. 2,3,[7][8][9] In this study, we evaluated the physical image quality of IIR using a phantom and including objects with two different contrasts scanned at four dose levels, and compared the results with those of a state-of-the-art IR technique.

2.A | CT system and IR
For acquiring CT images, we employed a SOMATOM Force dual source CT scanner (Siemens Healthcare, Erlangen, Germany) equipped with advanced modeled IR (ADMIRE), which has five noise reduction levels from 1 to 5, with 5 corresponding to the most intense noise reduction. The CT images considered for comparison were FBP, ADMIRE and IIR. IIR images were processed from the FBP images.

2.B | Measurement setup
We used two rod-shaped objects with 60 and 270 HU contrasts at 120 kV for measuring in-plane task-based transfer function (TTF), which has been employed for spatial resolution measurements of images processed with IR. 7,8 Each object with diameter of 3 cm and height of 4 cm was placed in a cylindrical acrylic case with diameter of 25 cm filled with water, as shown in Fig. 1(a). In this study, we approximated the rod size of the ACR phantom and used two contrasts of 60 and 270 HU to resemble a soft tissue and an enhanced vessel with a 12 mg iodine (mgI)/ml concentration, respectively. Moreover, the phantom diameter resembled the absorption of adult abdomen. 10 The rod containing iodine for this study was commercially available as a custom order supplied by the phantom manufacturer Kyoto Kagaku Corporation (Kyoto, Japan). The central axis of the phantom was accurately positioned in parallel to the rotation axis of the CT system with a 10 mm offset in the y-axis direction to avoid the specific modulation transfer function induced when the central axis matches the rotation axis of the CT system. 8 The region of the phantom without the rod was used for measuring the noise power spectrum (NPS).
Two other rods with diameter of 10 cm and contrasts of 60 and 270 HU were used for measuring task-based slice sensitivity profile (SSP task ). The objects were placed in the phantom as shown in Fig. 2(a). The phantom was tilted by approximately 3°with respect to the rotation axis to apply an established edge method that provides a sufficiently fine effective sampling for Fourier analysis. 11 An averaged sagittal image was created from a stack of axial images, as shown in Fig. 2(b), and SSP task was measured from the object edge formed by the top surface of the rod.

2.C | Data acquisition
The imaging conditions were as follows: applied voltage of 120 kVp, rotation time of 0.5 s, pitch factor of 0.6 using a detector of 196 × 0.6 mm. The CT images obtained from FBP and ADMIRE were reconstructed with a display field of view of 250 mm, nominal slice thickness of 1 mm, and abdominal standard reconstruction kernel Br40d. The noise reduction level of ADMIRE was set to 3 and 5 (ADMIRE 3 and ADMIRE 5, respectively). The radiation doses in volume CT dose index CTDI vol were set to 2.5, 5, 10, and 20 mGy for

2.D | In-plane TTF
TTF was determined from disc images of each rod. Image averaging can effectively improve the accuracy of transfer function measurements, and suitable contrast-to-noise ratios values should be above 25. 8 Hence, we used between 150 and 500 images obtained from 5 to 15 acquisitions. Then, a one-dimensional edge spread function from an averaged disc image was obtained using the circular edge method proposed by Richard et al. 7 We set the bin width to onefifth of the pixel pitch to create equidistant data for the edge spread function and reduce noise.

2.E | Noise power spectrum
The NPS was determined from the central 256 × 256 pixels from the area of the phantom images that did not contain the rod and by using the radial frequency method based on the two-dimensional Fourier transform. The two-dimensional NPS measurements were radially averaged and split into 40 frequency bins. To reduce the NPS variability, the results from 80 consecutive images were averaged. 2,12 2.F | System performance function Samei and Richard used the following detectability index, d', to assess the IR techniques' imaging performance: where u denotes the spatial frequency and S(u) is the spectrum of the signal to be detected. The d′ 2 value is a figure of merit that incorporates square of system performance (SP) TTF 2 (u)/NPS(u) and imaging task S 2 (u). 2 This index is similar to the prewhitening signalto-noise ratio that is based on an ideal observer model. 13 In this study, we focused on this SP function expressed as Since the TTFs we measured were specific for rod objects with the soft tissue and iodine contrasts presenting circular edges, we used the SP function to evaluate the noise reduction performance for the specific conditions that did not cover various contrasts of tissues such as bones and fats in clinical CT images.

2.G | Slice sensitivity profile
From the longitudinal edge surface in the phantom shown in Both the averaging and edge synthesizing effectively reduce noise, and the tilting with respect to the z axis provides an oversampled profile to accurately detect the edge. The SSP task was calculated from the derivative of the edge spread function, and then the full width at half maximum (FWHM) was determined from the obtained SSP task .

2.H | Image subtraction for evaluating noise reduction
We used a subtraction technique based on FBP-processed images for IIR to determine noise reduction. The rod images for TTF at

3.E | Image subtraction
Images obtained from FBP and IIR processing, and their subtraction images are shown in Fig. 6  | 203 excluded during IIR processing because they exhibit uniform areas with pixels having zero value.

| DISCUSSION
We evaluated the physical image quality using IIR at four radiation dose levels on objects presenting soft-tissue and iodine contrasts and compared the outcomes with those using FBP and ADMIRE implemented on a state-of-art dual source CT scanner. Overall, IIR achieved notable noise reduction and preserved both the in-plane and longitudinal resolutions with a negligible NPS peak shift related to a change in noise texture. ADMIRE reduced image noise, but its NPS exhibited peak shifts and the resolution was strongly dependent on the dose and contrast. This trend was more remarkable when using the stronger ADMIRE 5.