Phantom-based performance comparison of two commercial deep learning CT reconstruction algorithms with super- and normal-resolution settings

The paper proposes an image super resolution reconstruction (SRR) algorithm based on Undecimated Morphological Wavelet (UMW). As UMW can maintain more gradient information and eliminate more artificial blocks, it is used to decompose low resolution (LR) images. High frequency coefficients from the decomposition contain the gradients that reflect the horizontal, vertical, and diagonal domain information. Then both the low frequency and high frequency coefficients are interpolated to produce an extended group of coefficient sequence. Finally, the extended coefficients are fused and using inverse transformation to produce a super resolution image. Experimental results show that that the proposed algorithm have good performance in the image reconstruction quality and the computational efficiency.

In order to reduce greatly computational cost and storage requirement, three two-dimensional (2D) gradient algorithms for fast super-resolution image reconstruction was presented by minimizing the smooth cost function. To enhance estimation accuracy of super resolution image reconstruction, this paper proposes a 2D subgradient algorithm for super-resolution image reconstruction. Compared with the existing 2D gradient algorithms, the proposed 2D subgradient algorithm can minimize a novel nonsmooth cost function with an l 1 -norm learning term and TV term. Simulation results show that the proposed 2D subgradient algorithm has better performance in terms of both PSNR and visual quality than the existing 2D gradient algorithms.

To characterize the noise and spatial resolution properties of a commercially available deep learning-based computed tomography (CT) reconstruction algorithm. Two phantom experiments were performed. The first used a multisized image quality phantom (Mercury v3.0, Duke University) imaged at five radiation dose levels (CTDIvol : 0.9, 1.2, 3.6, 7.0, and 22.3mGy) with a fixed tube current technique on a commercial CT scanner (GE Revolution CT). Images were reconstructed with conventional (FBP), iterative (GE ASiR-V), and deep learning-based (GE True Fidelity) reconstruction algorithms. Noise power spectrum (NPS), high-contrast (air-polyethylene interface), and intermediate-contrast (water-polyethylene interface) task transfer functions (TTF) were measured for each dose level and phantom size and summarized in terms of average noise frequency (fav ) and frequency at which the TTF was reduced to 50% (f50% ), respectively. The second experiment used a custom phantom with low-contrast rods and lung texture sections for the assessment of low-contrast TTF and noise spatial distribution. The phantom was imaged at five dose levels (CTDIvol : 1.0, 2.1, 3.0, 6.0, and 10.0mGy) with 20 repeated scans at each dose, and images reconstructed with the same reconstruction algorithms. The local noise stationarity was assessed by generating spatial noise maps from the ensemble of repeated images and computing a noise inhomogeneity index, , following AAPM TG233 methods. All measurements were compared among the algorithms. Compared to FBP, noise magnitude was reduced on average (± one standard deviation) by 74±6% and 68±4% for ASiR-V (at "100%" setting) and True Fidelity (at "High" setting), respectively. The noise texture from ASiR-V had substantially lower noise frequency content with 55±4% lower NPS fav compared to FBP while True Fidelity had only marginally different noise frequency content with 9±5% lower NPS fav compared to FBP. Both ASiR-V and True Fidelity demonstrated locally nonstationary noise in a lung texture background at all radiation dose levels, with higher noise near high-contrast edges of vessels and lower noise in uniform regions. At the 1.0mGy dose level values were 314% and 271% higher in ASiR-V and True Fidelity compared to FBP, respectively. High-contrast spatial resolution was similar between all algorithms for all dose levels and phantom sizes (<3% difference in TTF f50% ). Compared to FBP, low-contrast spatial resolution was lower for ASiR-V and True Fidelity with a reduction of TTF f50% of up to 42% and 36%, respectively. The deep learning-based CT reconstruction demonstrated a strong noise magnitude reduction compared to FBP while maintaining similar noise texture and high-contrast spatial resolution. However, the algorithm resulted in images with a locally nonstationary noise in lung textured backgrounds and had somewhat degraded low-contrast spatial resolution similar to what has been observed in currently available iterative reconstruction techniques.

Medical image super-resolution reconstruction algorithms based on deep learning: A survey

Impact of a new deep-learning image reconstruction algorithm on potential dose reduction and quality of chest CT images: a phantom study.

The objective of this study was to compare the image quality of a standard single-source (SSS) computed tomography (CT) with that of a virtual single-source CT (VSS-CT) data set reconstructed from 2 raw data sets obtained by dual-source CT acquisition in abdominal CT to establish a radiation dose-neutral approach for the intraindividual comparison of 3 acquisition protocols at different radiation dose levels (RDLs). An abdominal phantom representing an 80-kg male was imaged using dual-source CT (SOMATOM Definition; Siemens Healthcare) at 3 RDLs with 120 kV(p) and different tube currents (low, standard, and high milliampere-second protocol). For each RDL, raw data were obtained once in single-source mode using x-ray tube A only and 5 times in dual-source mode using different ratios for tube current of x-ray tubes A and B (same total radiation dose; A/B: 90%/10%, 80%/20%, 70%/30%, 60%/40%, 50%/50%). For each RDL, SSS-CT and 5 virtual single-source image data sets (VSS-CT50 - 90) were reconstructed. To compare SSS-CT and VSS-CT data sets, image quality was assessed in terms of high- and low-contrast performance by calculating the modulation transfer function, image noise, noise power spectrum, and, for low contrast lesion detectability, the modified multiscale structural similarity index (MS-SSIM*). A maximum decrease of Δ = 5% of image quality compared with SSS-CT was defined as acceptable, and a noninferiority analysis with Δ was performed. For modulation transfer function, noninferiority was observed for all VSS-CT data sets and RDL (P < 0.05). Image noise demonstrated an acceptable increase (<3.2%, P < 0.05) for each RDL and noise power spectrum showed only minor differences in the midfrequency range. The MS-SSIM* index demonstrated for the high RDL protocol a minor decrease for VSS-CT data sets (<2%, P < 0.05). For the standard and low RDL, the relative differences of the MS-SSIM* index increased and were only in 1 case above Δ (standard RDL, mean VSS-CT80 5.1%, P > 0.05). The image quality obtained by virtual and SSS reconstruction using equivalent total radiation exposure to the patient showed only negligible differences in image quality. Therefore, this technique might allow an intraindividual comparison of full and reduced radiation dose protocols within 1 image acquisition step by simply splitting the radiation dose between the 2 x-ray tubes of a dual-source CT.

Purpose To determine the effect of radiation dose and iterative reconstruction (IR) on noise, contrast, resolution, and observer-based detectability of subtle hypoattenuating liver lesions and to estimate the dose reduction potential of the IR algorithm in question. Materials and Methods This prospective, single-center, HIPAA-compliant study was approved by the institutional review board. A dual-source computed tomography (CT) system was used to reconstruct CT projection data from 21 patients into six radiation dose levels (12.5%, 25%, 37.5%, 50%, 75%, and 100%) on the basis of two CT acquisitions. A series of virtual liver lesions (five per patient, 105 total, lesion-to-liver prereconstruction contrast of -15 HU, 12-mm diameter) were inserted into the raw CT projection data and images were reconstructed with filtered back projection (FBP) (B31f kernel) and sinogram-affirmed IR (SAFIRE) (I31f-5 kernel). Image noise (pixel standard deviation), lesion contrast (after reconstruction), lesion boundary sharpness (average normalized gradient at lesion boundary), and contrast-to-noise ratio (CNR) were compared. Next, a two-alternative forced choice perception experiment was performed (16 readers [six radiologists, 10 medical physicists]). A linear mixed-effects statistical model was used to compare detection accuracy between FBP and SAFIRE and to estimate the radiation dose reduction potential of SAFIRE. Results Compared with FBP, SAFIRE reduced noise by a mean of 53% ± 5, lesion contrast by 12% ± 4, and lesion sharpness by 13% ± 10 but increased CNR by 89% ± 19. Detection accuracy was 2% higher on average with SAFIRE than with FBP (P = .03), which translated into an estimated radiation dose reduction potential (±95% confidence interval) of 16% ± 13. Conclusion SAFIRE increases detectability at a given radiation dose (approximately 2% increase in detection accuracy) and allows for imaging at reduced radiation dose (16% ± 13), while maintaining low-contrast detectability of subtle hypoattenuating focal liver lesions. This estimated dose reduction is somewhat smaller than that suggested by past studies. © RSNA, 2017 Online supplemental material is available for this article.

Image super-resolution (SR) reconstruction is to reconstruct a high-resolution (HR) image from one or a series of low-resolution (LR) images in the same scene with a certain amount of prior knowledge. Learning based algorithm is an effective one in image super-resolution reconstruction algorithm. The core idea of the algorithm is to use the training examples of image to increase the high frequency information of the test image to achieve the purpose of image super-resolution reconstruction. This paper presents a novel algorithm for image super resolution based on morphological component analysis (MCA) and dictionary learning. The MCA decomposition based SR algorithm utilizes MCA to decompose an image into the texture part and the structure part and only takes the texture part to train the dictionary. The reconstruction of the texture part is based on sparse representation, while that of the structure part is based on more faster method, the bicubic interpolation. The proposed method improves the robustness of the image, while for different characteristics of textures and structure parts, using a different reconstruction algorithm, better preserves image details, improve the quality of the reconstructed image.

In order to further improve the reconstruction quality of super-resolution image, a new reconstruction algorithm of adaptive super-resolution image sequence is proposed. The algorithm could adaptively estimate the regularization coefficients for each low-resolution image, and self-adaptively update the iterative step size in the iterative process. In the process of super-resolution image reconstruction, the prior information of image are sufficiently utilized, thus the reconstruction effects would be better. In the conventional bilateral total variation (BTV) regularized algorithm, the regularization parameter can be selected adaptively, but the iterative step size is not adaptive. The simulation experiments are performed for Lena image and Fish image, which using the proposed algorithm and the conventional regularized algorithm. And the experimental results indicate that, compared with the conventional adaptive BTV algorithm, this proposed algorithm has a relatively higher peak signal-to-noise ratio (PSNR) and better visual reconstruction effects of image.

In order to solve the ill-posed problem in super-resolution image reconstruction, this paper proposes an adaptive regularization way use sparse representation. We build a new L1/2 non-convex optimization model and apply reweighted L2 Norm for the adaptive algorithm in this paper. Experimental results show the significant effect in denoising and preserving edge details. It outperforms some traditional methods in the value of peak signal to noise ratio and structural similarity.

Brain image quality according to beam collimation width and image reconstruction algorithm: A phantom study

To compare the spectral performance of three rapid kV switching dual-energy CT (DECT) systems on virtual monoenergetic images (VMIs) at low-energy levels on abdominal imaging. A multi-energy phantom was scanned on three DECT systems equipped with three different gemstone spectral imaging (GSI) platforms: GSI (1st generation, GSI-1st), GSI-Pro (2nd generation, GSI-2nd ), and GSI-Xtream (3rd generation, GSI-3rd). Acquisitions on the phantom were performed with a CTDIvol close to 11mGy. For all platforms, raw data were reconstructed using filtered-back projection (FBP) and a hybrid iterative reconstruction algorithm (ASIR-V at 50%; AV50). A deep-learning image reconstruction (DLR) algorithm (TrueFidelity) was used only for the GSI-3rd. Noise power spectrum (NPS) and task-based transfer function (TTF) were evaluated from 40 to 80 keV of VMIs. A detectability index (d') was computed to assess the detection of two contrast-enhanced lesions according to the keV level used. For all GSI platforms, the noise magnitude decreased from 40 to 70 keV, and using AV50 compared to FBP. The average NPS spatial frequency (fav ) and spatial resolution (TTF50% ) were similar from 40 to 70 keV and decreased with AV50 compared to FBP. Compared to AV50, using DLR reduced the noise magnitude (-27% ± 3%) and improved fav values (10% ± 0%) and altering spatial resolution (2% ± 5%). For the two lesions, d' values peaked at 70 keV for GSI-1st and GSI-2nd platforms and at 40/50 keV for GSI-3rd, for all reconstruction algorithms. The highest d' values were found for the GSI-3rd with DLR. Differences in image quality were found between the GSI platforms for VMIs at low keV. The new DLR algorithm on the GSI-3rd platform reduced noise and improved spatial resolution and detectability without changing the noise texture for VMIs at low keV. The choice of the best energy level in VMIs depends on the platform and the reconstruction algorithm.

To evaluate the image quality produced by six different iterative reconstruction (IR) algorithms in four CT systems in the setting of brain CT, using different radiation dose levels and iterative image optimisation levels. An image quality phantom, supplied with a bone mimicking annulus, was examined using four CT systems from different vendors and four radiation dose levels. Acquisitions were reconstructed using conventional filtered back-projection (FBP), three levels of statistical IR and, when available, a model-based IR algorithm. The evaluated image quality parameters were CT numbers, uniformity, noise, noise-power spectra, low-contrast resolution and spatial resolution. Compared with FBP, noise reduction was achieved by all six IR algorithms at all radiation dose levels, with further improvement seen at higher IR levels. Noise-power spectra revealed changes in noise distribution relative to the FBP for most statistical IR algorithms, especially the two model-based IR algorithms. Compared with FBP, variable degrees of improvements were seen in both objective and subjective low-contrast resolutions for all IR algorithms. Spatial resolution was improved with both model-based IR algorithms and one of the statistical IR algorithms. The four statistical IR algorithms evaluated in the study all improved the general image quality compared with FBP, with improvement seen for most or all evaluated quality criteria. Further improvement was achieved with one of the model-based IR algorithms. The six evaluated IR algorithms all improve the image quality in brain CT but show different strengths and weaknesses.

To assess the effect of a new lung enhancement filter combined with deep learning image reconstruction (DLIR) algorithm on image quality and ground-glass nodule (GGN) sharpness compared to hybrid iterative reconstruction or DLIR alone. Five artificial spherical GGNs with various densities (-250, -350, -450, -550, and -630 Hounsfield units) and 10 mm in diameter were placed in a thorax anthropomorphic phantom. Four scans at four different radiation dose levels were performed using a 256-slice CT (Revolution Apex CT, GE Healthcare). Each scan was reconstructed using three different reconstruction algorithms: adaptive statistical iterative reconstruction-V at a level of 50% (AR50), Truefidelity (TF), which is a DLIR method, and TF with a lung enhancement filter (TF + Lu). Thus, 12 sets of reconstructed images were obtained and analyzed. Image noise, signal-to-noise ratio, and contrast-to-noise ratio were compared among the three reconstruction algorithms. Nodule sharpness was compared among the three reconstruction algorithms using the full-width at half-maximum value. Furthermore, subjective image quality analysis was performed. AR50 demonstrated the highest level of noise, which was decreased by using TF + Lu and TF alone (P = 0.001). TF + Lu significantly improved nodule sharpness at all radiation doses compared to TF alone (P = 0.001). The nodule sharpness of TF + Lu was similar to that of AR50. Using TF alone resulted in the lowest nodule sharpness. Adding a lung enhancement filter to DLIR (TF + Lu) significantly improved the nodule sharpness compared to DLIR alone (TF). TF + Lu can be an effective reconstruction technique to enhance image quality and GGN evaluation in ultralow-dose chest CT scans.

Performance evaluation of deep learning image reconstruction algorithm for dual-energy spectral CT imaging: A phantom study.