Adaptive Gradient Balancing for Undersampled MRI Reconstruction and Image-to-Image Translation
0
Citation
0
Reference
10
Related Paper
Abstract:
Recent accelerated MRI reconstruction models have used Deep Neural Networks (DNNs) to reconstruct relatively high-quality images from highly undersampled k-space data, enabling much faster MRI scanning. However, these techniques sometimes struggle to reconstruct sharp images that preserve fine detail while maintaining a natural appearance. In this work, we enhance the image quality by using a Conditional Wasserstein Generative Adversarial Network combined with a novel Adaptive Gradient Balancing (AGB) technique that automates the process of combining the adversarial and pixel-wise terms and streamlines hyperparameter tuning. In addition, we introduce a Densely Connected Iterative Network, which is an undersampled MRI reconstruction network that utilizes dense connections. In MRI, our method minimizes artifacts, while maintaining a high-quality reconstruction that produces sharper images than other techniques. To demonstrate the general nature of our method, it is further evaluated on a battery of image-to-image translation experiments, demonstrating an ability to recover from sub-optimal weighting in multi-term adversarial training.Keywords:
Image translation
Hyperparameter
Image noise
Radon transform
Reconstruction algorithm
Cite
Citations (63)
Abstract Introduction Cadaveric studies provide a means of safely assessing new technologies and optimizing scanning prior to clinical validation. Reducing radiation exposure in a clinical setting can entail incremental dose reductions to avoid missing important clinical findings. The use of cadavers allows assessment of the impact of more substantial dose reductions on image quality. Our aim was to identify a suitable low‐dose abdominopelvic CT protocol for subsequent clinical validation. Methods Five human cadavers were scanned at one conventional dose and three low‐dose settings. All scans were reconstructed using three different reconstruction algorithms: filtered back projection ( FBP ), hybrid iterative reconstruction (60% FBP and 40% adaptive statistical iterative reconstruction ( ASIR 40)), and model‐based iterative reconstruction ( MBIR ). Two readers rated the image quality both quantitatively and qualitatively. Results Model‐based iterative reconstruction images had significantly better objective image noise and higher qualitative scores compared with both FBP and ASIR 40 images at all dose levels. The greatest absolute noise reduction, between MBIR and FBP , of 34.3 HU (equating to a 68% reduction) was at the lowest dose level. MBIR reduced image noise and improved image quality even in CT images acquired with a mean radiation dose reduction of 62% compared with conventional dose studies reconstructed with ASIR 40, with lower levels of objective image noise, superior diagnostic acceptability and contrast resolution, and comparable subjective image noise and streak artefact scores. Conclusion This cadaveric study demonstrates that MBIR reduces image noise and improves image quality in abdominopelvic CT images acquired with dose reductions of up to 62%.
Image noise
Cadaveric spasm
Radon transform
Cite
Citations (7)
The aim of this study was to evaluate image quality and radiation dose in low-dose head and neck CT comparing two different commercially available iterative reconstruction algorithms: sinogram-affirmed iterative reconstruction (SAFIRE) and advanced modeled iterative reconstruction (ADMIRE) with fixed and automated tube voltage adaptation (TVA).CT examinations of 103 patients were analysed. 58 patients were examined on a single-source CT at fixed tube voltage of 120 kV and reconstructed with filtered back projection (FBP) and SAFIRE (Strength Level 3). 45 patients were examined in a single-source mode on a dual-source CT with automated TVA and reconstructed with FBP and ADMIRE (Strength Levels 2 and 3). Image noise was calculated in seven anatomical volumes of interest. Subjective evaluation of the CT images was performed using a four-grade scale.Mean CT numbers of FBP and the corresponding iterative reconstruction did not differ significantly (p = 0.74-0.99). Image noise was lower with both iterative reconstruction techniques than with FBP (SAFIRE 3: -22.3%; ADMIRE 2: -14.9%; ADMIRE 3: -24.2%; all p < 0.05); hence, the signal-to-noise ratio and the contrast-to-noise values were higher. Subjective image quality revealed a more favourable result for the iterative reconstruction. ADMIRE 3 in combination with automated TVA showed 14.4% (p < 0.05) less image noise with a 7.5% lower radiation dose than SAFIRE 3 with fixed tube voltage.Higher image quality at lower radiation dose can be achieved using ADMIRE in combination with automated TVA.
Image noise
Radon transform
Contrast-to-noise ratio
Cite
Citations (19)
Iterative computed tomography (CT) image reconstruction shows high potential for the preservation of image quality in diagnostic CT while reducing patients' exposure; it has become available for low-dose CT (LD-CT) in high-end hybrid imaging systems (e.g. single-photon emission computed tomography [SPECT]-CT).To examine the effect of an iterative CT reconstruction algorithm on image quality, image noise, detectability, and the reader's confidence for LD-CT data by a subjective assessment.The LD-CT data were validated for 40 patients examined by an abdominal hybrid SPECT-CT (U = 120 kV, I = 40 mA, pitch = 1.375). LD-CT was reconstructed using either filtered back projection (FBP) or an iterative image reconstruction algorithm (Adaptive Statistical Iterative Reconstruction [ASIR]®) with different parameters (ASIR levels 50% and 100%). The data were validated by two independent blinded readers using a scoring system for image quality, image noise, detectability, and reader confidence, for a predefined set of 16 anatomic substructures.The image quality was significantly improved by iterative reconstruction of the LD-CT data compared with FBP (P ≤ 0.0001). While detectability increased in only 2/16 structures (P ≤ 0.03), the reader's confidence increased significantly due to iterative reconstruction (P ≤ 0.002). Meanwhile, at the ASIR level of 100%, the detectability in bone structure was highly reduced (P = 0.003).An ASIR level of 50% represents a good compromise in abdominal LD-CT image reconstruction. The specific ASIR level improved image quality (reduced image noise) and reader confidence, while preserving detectability of bone structure.
Image noise
Radon transform
Reconstruction algorithm
Cite
Citations (11)
Objective: To investigate the impact of iterative reconstruction method(iDose4) on the CT image quality and dose reduction by using phantom studies. Methods: A Catphan phantom was scanned by a 256-slice CT(Philips Brilliance iCT) with varied mAs from 20 to 200, the axial images were reconstructed by filtered back projection(FBP) and iDose4with a level of 1 to 6. Image quality was assessed with following aspect: CT value accuracy and consistency, image noise and signal-noise-ratio(SNR), low density resolution. Meanwhile the volume CT dose index CTDIvol and Dose length product DLP were recorded for dose evaluation. Results: The axial image CT values has a high degree of consistency(P0.05) for both iDose4and FBP reconstruction method. The image noise can be reduced in different degree with different iDose4level 1~6 reconstruction(9.46%~43.30%). When tube current reduced to 40%~50%(80~100 mAs), iDose4level 6 got image without significant difference regarding to the image noise compared to 200 mAs FBP image(P0.05); under the conditions of 50%~60% initial tube current(100~120 mAs), low density resolution could remain unchanged. Conclusion: iDose4can significantly reduce image noise and improve image quality, under the condition of 40% ~50% scanning dose can get satisfactory image quality. Thus iterative reconstruction has larger space to lower the scanning dose.
Image noise
Contrast-to-noise ratio
Cite
Citations (0)
Radon transform
Image noise
Hounsfield scale
Computed Tomography Angiography
Contrast-to-noise ratio
Cite
Citations (28)
The aim of this study was to evaluate the impact on dose reduction and image quality of the new iterative reconstruction technique: adaptive statistical iterative reconstruction (ASIR-V).Fifty consecutive oncologic patients acted as case controls undergoing during their follow-up a computed tomography scan both with ASIR and ASIR-V. Each study was analyzed in a double-blinded fashion by 2 radiologists. Both quantitative and qualitative analyses of image quality were conducted.Computed tomography scanner radiation output was 38% (29%-45%) lower (P < 0.0001) for the ASIR-V examinations than for the ASIR ones. The quantitative image noise was significantly lower (P < 0.0001) for ASIR-V. Adaptive statistical iterative reconstruction-V had a higher performance for the subjective image noise (P = 0.01 for 5 mm and P = 0.009 for 1.25 mm), the other parameters (image sharpness, diagnostic acceptability, and overall image quality) being similar (P > 0.05).Adaptive statistical iterative reconstruction-V is a new iterative reconstruction technique that has the potential to provide image quality equal to or greater than ASIR, with a dose reduction around 40%.
Image noise
Statistical noise
Cite
Citations (27)
To evaluate the performance of sinogram-affirmed iterative (SAFIRE) reconstruction on image quality of low-dose lung computed tomographic (CT) screening compared with filtered back projection (FBP).Three hundred four patients for annual low-dose lung CT screening were examined by a dual-source CT system at 120 kilovolt (peak) with reference tube current of 40 mA·s. Six image serials were reconstructed, including one data set of FBP and 5 data sets of SAFIRE with different reconstruction strengths from 1 to 5. Image noise was recorded; and subjective scores of image noise, images artifacts, and the overall image quality were also assessed by 2 radiologists.The mean ± SD weight for all patients was 66.3 ± 12.8 kg, and the body mass index was 23.4 ± 3.2. The mean ± SD dose-length product was 95.2 ± 30.6 mGy cm, and the mean ± SD effective dose was 1.6 ± 0.5 mSv. The observation agreements for image noise grade, artifact grade, and the overall image quality were 0.785, 0.595 and 0.512, respectively. Among the overall 6 data sets, both the measured mean objective image noise and the subjective image noise of FBP was the highest, and the image noise decreased with the increasing of SAFIRE reconstruction strength. The data sets of S3 obtained the best image quality scores.Sinogram-affirmed iterative reconstruction can significantly improve image quality of low-dose lung CT screening compared with FBP, and SAFIRE with reconstruction strength 3 was a pertinent choice for low-dose lung CT.
Radon transform
Image noise
Cite
Citations (42)
Purpose: To demonstrate the image-quality benefits and potential for significant dose reduction with Model-Based Iterative Reconstruction (MBIR) technique incorporating physical model of computed tomography (CT) systems. Method and Materials: A model based iterative reconstruction (MBIR), a maximum a posteriori (MAP) estimate with edge-preserving prior, has been developed for x-ray CT image reconstruction. It utilizes a more accurate physical model of the imaging chain accounting for system-optics, noise and non-idealities in the data, hence improves image quality compared to conventional filtered backprojection (FBP) at significantly reduced dose levels. In this work, a GE multi-slice CT system was used to acquire a set of multi-dose data and standard FBP reconstruction. For resolution assessment, a Catphan600® phantom was scanned at three dose levels (40, 20, and 10 mGy with 120kVp spectrum), and images were reconstructed using two methods: FBP with ASiR, and the MBIR. For artifact and image-quality evaluations, an anthropomorphic CT abdomen phantom (Kyoto Kagaku Co., Ltd) was scanned at four dose levels (120kVp spectrum with 225, 112, 54, and 27 mAs), and a comparative image-quality study between standard FBP and MBIR in slice and multi-planar reformat (MPR) modes was made. In addition, few clinical case studies were also used to compare the imaging performance in actual clinical data. Results: From the resolution study, we found that even at 1/4th dose, MBIR images have improved resolution at significantly reduced noise compared to standard state-of-the-art FBP with ASiR. Use of ASIR provides up to 50% dose reduction with equivalent FBP image-quality. For anthropomorphic phantom, even below 1/8th dose, MBIR images outperformed the corresponding FBP images in both, slice and MPR modes, demonstrating immense potential for dose reduction, yet improved image quality, in clinical CT. Conclusion: Results of the MBIR method demonstrated significant potential for dose reduction and image-quality improvements in clinical CT.
Image noise
Artifact (error)
Cite
Citations (13)
Purpose: To reduce radiation dose to patients undergoing computed tomography (CT) for lung cancer screening while maintaining overall diagnostic image quality and definition of ground‐glass opacities (GGOs). Methods: A Catphan phantom, a Kyoto Kagaku lung screening phantom, and a Kyoto Kagaku multipurpose chest phantom were scanned on a GE Discovery CT750 HD scanner to quantitatively assess the performance of two image reconstruction algorithms (adaptive statistical iterative reconstruction [ASIR] and model‐based iterative reconstruction [MBIR]) used in conjunction with reduced tube currents relative to a standard clinical lung cancer screening protocol (51 mAs and filtered back projection [FBP] reconstruction; CTDIvol = 3.9 mGy). To further assess the algorithms performances, qualitative image analysis was conducted (in the form of a reader study) using the multipurpose chest phantom, which was implanted with eight GGOs of two densities. Results: Our quantitative image analysis indicated that tube current, and thus radiation dose, could be reduced by at least 60% from MBIR while maintaining similar noise and spatial resolution (as determined from the CT images) compared with conventional FBP. The reader study indicated that dose could be reduced by 40% (to 30 mAs or 2.3 mGy) or 60% (to 20 mAs or 1.5 mGy) from using ASIR or MBIR, respectively, while maintaining GGO definition. Additionally, the readers ratings for overall image quality were equal or better (for a given dose) when using ASIR or MBIR compared with FBP. Conclusion: Combining ASIR or MBIR with reduced tube current may allow for lower doses while maintaining overall diagnostic image quality, as well as GGO definition during CT lung cancer screening. Dr. Adam Chandler is an employee of GE Healthcare. Dr. Tinsu Pan is the owner of Texas Medical Imaging Consultants.
Radon transform
Image noise
Cite
Citations (0)