Noise and spatial resolution properties of a commercially available deep learning-based CT reconstruction algorithm

Deep learning image reconstruction algorithm reduces image noise while alters radiomics features in dual-energy CT in comparison with conventional iterative reconstruction algorithms: a phantom study

OBJECTIVES

To compare image quality between a deep learning image reconstruction (DLIR) algorithm and conventional iterative reconstruction (IR) algorithms in dual-energy CT (DECT) and to assess the impact of these algorithms on radiomics robustness.

METHODS

A phantom with clinical-relevant densities was imaged on seven DECT scanners with the same voxel size using typical abdominal-pelvis examination protocols. On one DECT scanner, raw data were reconstructed using both conventional IR (adaptive statistical iterative reconstruction-V, ASIR-V) and DLIR. Nine sets of corresponding images were generated on other six DECT scanners using scanner-equipped conventional IR. Regions of interest were delineated through rigid registrations. Image quality was compared. Pyradiomics platform was used for radiomics feature extraction. Test-retest repeatability was assessed by Bland-Altman analysis for repeated scans. Inter-reconstruction algorithm reproducibility between conventional IR and DLIR was tested by intraclass correlation coefficient (ICC) and concordance correlation coefficient (CCC). Inter-scanner reproducibility was evaluated by coefficient of variation (CV) and quartile coefficient of dispersion (QCD). Robust features were identified.

RESULTS

DLIR significantly improved image quality. Ninety-four radiomics features were extracted and nine features were considered as robust. 93.87% features were repeatable between repeated scans. ASIR-V images showed higher reproducibility to other conventional IR than DLIR (ICC mean, 0.603 vs 0.558, p = 0.001; CCC mean, 0.554 vs 0.510, p = 0.004). 7.45% and 26.83% features were reproducible among scanners evaluated by CV and QCD, respectively.

CONCLUSIONS

DLIR improves quality of DECT images but may alter radiomics features compared to conventional IR. Nine robust DECT radiomics features were identified.

KEY POINTS

• DLIR improves DECT image quality in terms of signal-to-noise ratio and contrast-to-noise ratio compared with ASIR-V and showed the highest noise reduction rate and lowest peak frequency shift. • Most of radiomics features are repeatable between repeated DECT scans, while inter-reconstruction algorithm reproducibility between conventional IR and DLIR, and inter-scanner reproducibility, are low. • Although DLIR may alter radiomics features compared to IR algorithms, nine radiomics features survived repeatability and reproducibility analysis among DECT scanners and reconstruction algorithms, which allows further validation and clinical-relevant analysis.

Safety Considerations in MRI and CT

OBJECTIVE

MRI and CT are indispensable imaging modalities for the evaluation of patients with neurologic disease, and each is particularly well suited to address specific clinical questions. Although both of these imaging modalities have excellent safety profiles in clinical use as a result of concerted and dedicated efforts, each has potential physical and procedural risks that the practitioner should be aware of, which are described in this article.

LATEST DEVELOPMENTS

Recent advancements have been made in understanding and reducing safety risks with MR and CT. The magnetic fields in MRI create risks for dangerous projectile accidents, radiofrequency burns, and deleterious interactions with implanted devices, and serious patient injuries and deaths have occurred. Ionizing radiation in CT may be associated with shorter-term deterministic effects on biological tissues at extremely high doses and longer-term stochastic effects related to mutagenesis and carcinogenesis at low doses. The cancer risk of radiation exposure in diagnostic CT is considered extremely low, and the benefit of an appropriately indicated CT examination far outweighs the potential risk. Continuing major efforts are centered on improving image quality and the diagnostic power of CT while concurrently keeping radiation doses as low as reasonably achievable.

ESSENTIAL POINTS

An understanding of these MRI and CT safety issues that are central to contemporary radiology practice is essential for the safe and effective treatment of patients with neurologic disease.

"Image quality evaluation of the Precise image CT deep learning reconstruction algorithm compared to Filtered Back-projection and iDose4: a phantom study at different dose levels"

PURPOSE

To characterize the performance of the Precise Image (PI) deep learning reconstruction (DLR) algorithm for abdominal Computed Tomography (CT) imaging.

METHODS

CT images of the Catphan-600 phantom (equipped with an external annulus) were acquired using an abdominal protocol at four dose levels and reconstructed using FBP, iDose⁴ (levels 2,5) and PI ('Soft Tissue' definition, levels 'Sharper','Sharp','Standard','Smooth','Smoother'). Image noise, image non-uniformity, noise power spectrum (NPS), target transfer function (TTF), detectability index (d'), CT numbers accuracy and image histograms were analyzed.

RESULTS

The behavior of the PI algorithm depended strongly on the selected level of reconstruction. The phantom analysis suggested that the PI image noise decreased linearly by varying the level of reconstruction from Sharper to Smoother, expressing a noise reduction up to 80% with respect to FBP. Additionally, the non-uniformity decreased, the histograms became narrower, and d' values increased as PI reconstruction levels changed from Sharper to Smoother. PI had no significant impact on the average CT number of different contrast objects. The conventional FBP NPS was deeply altered only by Smooth and Smoother levels of reconstruction. Furthermore, spatial resolution was found to be dose- and contrast-dependent, but in each analyzed condition it was greater than or comparable to FBP and iDose⁴ TTFs.

CONCLUSIONS

The PI algorithm can reduce image noise with respect to FBP and iDose⁴; spatial resolution, CT numbers and image uniformity are generally preserved by the algorithm but changes in NPS for the Smooth and Smoother levels need to be considered in protocols implementation.

Contribution of an artificial intelligence deep-learning reconstruction algorithm for dose optimization in lumbar spine CT examination: A phantom study

PURPOSE

The purpose of this study was to assess the impact of the new artificial intelligence deep-learning reconstruction (AI-DLR) algorithm on image quality and radiation dose compared with iterative reconstruction algorithm in lumbar spine computed tomography (CT) examination.

MATERIALS AND METHODS

Acquisitions on phantoms were performed using a tube current modulation system for four DoseRight Indexes (DRI) (i.e., 26/23/20/15). Raw data were reconstructed using the Level 4 of iDose⁴ (i4) and three levels of AI-DLR (Smoother/Smooth/Standard) with a bone reconstruction kernel. The Noise power spectrum (NPS), task-based transfer function (TTF) and detectability index (d') were computed (d' modeled detection of a lytic and a sclerotic bone lesions). Image quality was subjectively assessed on an anthropomorphic phantom by two radiologists.

RESULTS

The Noise magnitude was lower with AI-DLR than i4 and decreased from Standard to Smooth (-31 ± 0.1 [SD]%) and Smooth to Smoother (-48 ± 0.1 [SD]%). The average NPS spatial frequency was similar with i4 (0.43 ± 0.01 [SD] mm^-1) and Standard (0.42 ± 0.01 [SD] mm^-1) but decreased from Standard to Smoother (0.36 ± 0.01 [SD] mm^-1). TTF values at 50% decreased as the dose decreased but were similar with i4 and all AI-DLR levels. For both simulated lesions, d' values increased from Standard to Smoother levels. Higher detectabilities were found with a DRI at 15 and Smooth and Smoother levels than with a DRI at 26 and i4. The images obtained with these dose and AI-DLR levels were rated satisfactory for clinical use by the radiologists.

CONCLUSION

Using Smooth and Smoother levels with CT allows a significant dose reduction (up to 72%) with a high detectability of lytic and sclerotic bone lesions and a clinical overall image quality.

Comparison of image quality of two versions of deep-learning image reconstruction algorithm on a rapid kV-switching CT: a phantom study

BACKGROUND

To assess the impact of the new version of a deep learning (DL) spectral reconstruction on image quality of virtual monoenergetic images (VMIs) for contrast-enhanced abdominal computed tomography in the rapid kV-switching platform.

METHODS

Two phantoms were scanned with a rapid kV-switching CT using abdomen-pelvic CT examination parameters at dose of 12.6 mGy. Images were reconstructed using two versions of DL spectral reconstruction algorithms (DLSR V1 and V2) for three reconstruction levels. The noise power spectrum (NSP) and task-based transfer function at 50% (TTF₅₀) were computed at 40/50/60/70 keV. A detectability index (d') was calculated for enhanced lesions at low iodine concentrations: 2, 1, and 0.5 mg/mL.

RESULTS

The noise magnitude was significantly lower with DLSR V2 compared to DLSR V1 for energy levels between 40 and 60 keV by -36.5% ± 1.4% (mean ± standard deviation) for the standard level. The average NPS frequencies increased significantly with DLSR V2 by 23.7% ± 4.2% for the standard level. The highest difference in TTF₅₀ was observed at the mild level with a significant increase of 61.7% ± 11.8% over 40-60 keV energy with DLSR V2. The d' values were significantly higher for DLSR V2 versus DLSR V1.

CONCLUSIONS

The DLSR V2 improves image quality and detectability of low iodine concentrations in VMIs compared to DLSR V1. This suggests a great potential of DLSR V2 to reduce iodined contrast doses.

COMBINING HI-RESOLUTION SCAN MODE WITH DEEP LEARNING RECONSTRUCTION ALGORITHMS IN CARDIAC CT

To investigate the impact of combining the high-resolution (Hi-res) scan mode with deep learning image reconstruction (DLIR) algorithm in CT. Two phantoms (Catphan600® and Lungman, small, medium, large size) were CT scanned using combinations of Hi-res/standard mode and high-definition (HD)/standard kernels. Images were reconstructed with ASiR-V and three levels of DLIR. Spatial resolution, noise and contrast-to-noise ratio (CNR) were assessed. The radiation dose was recorded. The spatial resolution increased using Hi-res & HD. Image noise in the Catphan600® (69%) and the Lungman (10-70%) significantly increased when Hi-res & HD was applied. DLIR reduced the mean noise (54%). The CNR was reduced (64%) for Hi-res & HD. The radiation dose increased for both small (+70%) and medium (+43%) Lungman phantoms but decreased slightly for the large ones (-3%) when Hi-res was applied. In conclusion, the Hi-res scan mode improved the spatial resolution. The HD kernel significantly increased the image noise. DLIR improved the image noise and CNR and did not affect the spatial resolution.

Pulmonary nodule volumetric accuracy of a deep learning-based reconstruction algorithm in low-dose computed tomography: A phantom study

PURPOSE

To compare the image properties and pulmonary nodule volumetric accuracies among deep learning-based reconstruction (DLR), filtered back projection (FBP), and hybrid iterative reconstruction (hybrid IR) in low-dose computed tomography (LDCT).

METHODS

A multipurpose chest phantom containing artificial spherical pulmonary nodules with 5-, 8-, 10-, and 12-mm diameters and Hounsfield units (HUs) of -630 and +100 HU was scanned 20 times at a standard dose, based on a low-dose screening CT trial, and at 1/2, 1/6, and 1/12 of the standard dose. To assess noise reduction performance and volumetric accuracy, the standard deviations (SDs) of the pixel values and volumetric percentage errors (PEs) were compared among FBP, hybrid IR, and DLR. The noise non-stationarity index (NNSI) was calculated from 20 image replicates and compared among FBP, hybrid IR, and DLR to evaluate noise stationarity.

RESULTS

The SD reduction rates for FBP in hybrid IR and DLR were 62 %-85 % and 79 %-90 %, respectively. For the four nodules with +100 HU, the PE of all reconstruction methods was <±25 % (not clinically relevant). For the four nodules with -630 HU, the PEs were equivalent or lower for hybrid IR and DLR than for FBP, and the PE difference between hybrid IR and DLR ranged from 0 % to 7%. The NNSI was significantly higher for DLR than for FBP and hybrid IR (p < 0.01).

CONCLUSIONS

Greater noise suppression was achieved with DLR than with hybrid IR without compromising nodule volumetric accuracy in LDCT despite the representative noise non-stationarity.

Cardiac CTA image quality of adaptive statistical iterative reconstruction-V versus deep learning reconstruction "TrueFidelity" in children with congenital heart disease

BACKGROUND

Several recent studies have reported that deep learning reconstruction "TrueFidelity" (TF) improves computed tomography (CT) image quality. However, no study has compared adaptive statistical repeated reconstruction (ASIR-V) using TF in pediatric cardiac CT angiography (CTA) with a low peak kilovoltage.

OBJECTIVE

This study aimed to determine whether ASIR-V or TF CTA image quality is superior in children with congenital heart disease (CHD).

MATERIALS AND METHODS

Fifty children (median age, 2 months; interquartile range, 0-5 months; 28 men) with CHD who underwent CTA were enrolled between June and September 2020. Images were reconstructed using 2 ASIR-V blending factors (80% and 100% [AV-100]) and 3 TF settings (low, medium, and high [TF-H] strength levels). For the quantitative analyses, 3 objective image qualities (attenuation, noise, and signal-to-noise ratio [SNR]) were measured of the great vessels and heart chambers. The contrast-to-noise ratio (CNR) was also evaluated between the left ventricle and the dial wall. For the qualitative analyses, the degree of quantum mottle and blurring at the upper level to the first branch of the main pulmonary artery was assessed independently by 2 radiologists.

RESULTS

When the ASIR-V blending factor level and TF strength were higher, the noise was lower, and the SNR was higher. The image noise and SNR of TF-H were significantly lower and higher than those of AV-100 (P < .01), except for noise in the right atrium and left pulmonary artery and SNR of the right ventricle. Regarding CNR, TF-H was significantly better than AV-100 (P < .01). In addition, in the objective assessment of the degree of quantum mottle and blurring, TF-H had the best score among all examined image sets (P < .01).

CONCLUSION

TF-H is superior to AV-100 in terms of objective and subjective image quality. Consequently, TF-H was the best image set for cardiac CTA in children with CHD.

Is There Any Improvement in Image Quality in Obese Patients When Using a New X-ray Tube and Deep Learning Image Reconstruction in Coronary Computed Tomography Angiography?

Deep learning image reconstruction (DLIR) is a technique that should reduce noise and improve image quality. This study assessed the impact of using both higher tube currents as well as DLIR on the image quality and diagnostic accuracy. The study consisted of 51 symptomatic obese (BMI > 30 kg/m2) patients with low to moderate risk of coronary artery disease (CAD). All patients underwent coronary computed tomography angiography (CCTA) twice, first with the Revolution CT scanner and then with the upgraded Revolution Apex scanner with the ability to increase tube current. Images were reconstructed using ASiR-V 50% and DLIR. The image quality was evaluated by an observer using a Likert score and by ROI measurements in aorta and the myocardium. Image quality was significantly improved with the Revolution Apex scanner and reconstruction with DLIR resulting in an odds ratio of 1.23 (p = 0.017), and noise was reduced by 41%. A total of 88% of the image sets performed with Revolution Apex + DLIR were assessed as good enough for diagnosis compared to 69% of the image sets performed with Revolution Apex/CT + ASiR-V. In obese patients, the combination of higher tube current and DLIR significantly improves the subjective image quality and diagnostic utility and reduces noise.

Application of medical imaging methods and artificial intelligence in tissue engineering and organ-on-a-chip

Organ-on-a-chip (OOC) is a new type of biochip technology. Various types of OOC systems have been developed rapidly in the past decade and found important applications in drug screening and precision medicine. However, due to the complexity in the structure of both the chip-body itself and the engineered-tissue inside, the imaging and analysis of OOC have still been a big challenge for biomedical researchers. Considering that medical imaging is moving towards higher spatial and temporal resolution and has more applications in tissue engineering, this paper aims to review medical imaging methods, including CT, micro-CT, MRI, small animal MRI, and OCT, and introduces the application of 3D printing in tissue engineering and OOC in which medical imaging plays an important role. The achievements of medical imaging assisted tissue engineering are reviewed, and the potential applications of medical imaging in organoids and OOC are discussed. Moreover, artificial intelligence - especially deep learning - has demonstrated its excellence in the analysis of medical imaging; we will also present the application of artificial intelligence in the image analysis of 3D tissues, especially for organoids developed in novel OOC systems.

A Review of Deep Learning CT Reconstruction: Concepts, Limitations, and Promise in Clinical Practice

Purpose of Review

Deep Learning reconstruction (DLR) is the current state-of-the-art method for CT image formation. Comparisons to existing filter back-projection, iterative, and model-based reconstructions are now available in the literature. This review summarizes the prior reconstruction methods, introduces DLR, and then reviews recent findings from DLR from a physics and clinical perspective.

Recent Findings

DLR has been shown to allow for noise magnitude reductions relative to filtered back-projection without suffering from “plastic” or “blotchy” noise texture that was found objectionable with most iterative and model-based solutions. Clinically, early reader studies have reported increases in subjective quality scores and studies have successfully implemented DLR-enabled dose reductions.

Summary

The future of CT image reconstruction is bright; deep learning methods have only started to tackle problems in this space via addressing noise reduction. Artifact mitigation and spectral applications likely be future candidates for DLR applications.

Comparative Assessment of Noise Properties for Two Deep Learning CT Image Reconstruction Techniques and Filtered Back Projection

BACKGROUND

Two deep-learning image reconstruction (DLIR) techniques from two different CT vendors have recently been introduced into clinical practice.

PURPOSE

To characterize the noise properties of two DLIR techniques with different training methods, using a phantom containing a simple uniform and a complex non-uniform region.

METHODS

A water-bath phantom with a diameter of 300 mm was used as a base phantom. A textured phantom with a diameter of 128 mm, which was made of two materials, one equivalent to water and the other being 12 mg/mL diluted iodine, irregularly mixed to create a complex texture (non-uniform region), was placed in the base phantom. Thirty repeated phantom scans were performed using two CT scanners (GE, Revolution CT with Apex Edition; Canon, Aquilion One PRISM Edition) at two dose levels (CTDI: 5 and 15 mGy). Images were reconstructed with each CT system's filtered back projection (FBP) and DLIR [GE, TrueFidelity (TF); Canon, Advanced intelligent Clear-IQ Engine Body Sharp (AC)] for three process strengths. For basic characteristics of noise, the standard deviation (SD) and noise power spectrum (NPS) were measured for the uniform (water) region. A noise magnitude map was generated by calculating the inter-image SD at each pixel position across the 30 images. Then, a noise reduction map (NRM), which visualizes the relative differences in noise magnitude between FBP and DLIR, was calculated. The NRM values ranged from 0.0 to 1.0. A low NRM value represents a less aggressive noise reduction. The histograms of the NRM value were analyzed for the uniform and non-uniform regions.

RESULTS

The reduction in noise magnitude compared with FBP tended to be greater with AC (45%-85%) than with TF (32%-65%). The average NPS frequencies of TF and AC were almost comparable to those of FBP, except for the low-dose condition and the high noise reduction strength for AC. The NRM values of TF and AC were higher in the uniform region than in the non-uniform region. In the non-uniform region, TF's average NRM values (0.21-0.48) tended to be lower than AC's (0.39-0.78). The histograms for TF showed a small overlap between the uniform and the non-uniform regions; in contrast, those for AC showed a greater overlap. This difference seems to indicate that TF processes the uniform and non-uniform regions more differently than AC does.

CONCLUSION

This study has revealed a distinct difference in characteristics between the two DLIR techniques: TF tends to offer less aggressive noise reduction in non-uniform regions and preserve the original signals, whereas AC tends to prioritize noise filtering over edge-preservation, especially at the low-dose condition and with the high noise reduction strength. This article is protected by copyright. All rights reserved.

Radiation Dose Reduction for 80-kVp Pediatric CT Using Deep Learning-Based Reconstruction: A Clinical and Phantom Study

BACKGROUND. Deep learning-based reconstruction (DLR) may facilitate CT radiation dose reduction, but a paucity of literature has compared lower-dose DLR images with standard-dose iterative reconstruction (IR) images or explored application of DLR to low-tube-voltage scanning in children. OBJECTIVE. The purpose of this study was to assess whether DLR can be used to reduce radiation dose while maintaining diagnostic image quality in comparison with hybrid IR (HIR) and model-based IR (MBIR) for low-tube-voltage pediatric CT. METHODS. This retrospective study included children 6 years old or younger who underwent contrast-enhanced 80-kVp CT with a standard-dose or lower-dose protocol. Standard images were reconstructed with HIR, and lower-dose images were reconstructed with HIR, MBIR, and DLR. Size-specific dose estimate (SSDE) was calculated for both protocols. Image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) were quantified. Two radiologists independently evaluated noise magnitude, noise texture, streak artifact, edge sharpness, and overall quality. Interreader agreement was assessed, and mean values were calculated. To evaluate task-based object detection performance, a phantom was imaged with 80-kVp CT at six doses (SSDE, 0.6-5.3 mGy). Detectability index (d') was calculated from the noise power spectrum and task-based transfer function. Reconstruction methods were compared. RESULTS. Sixty-five children (mean age, 25.0 ± 25.2 months) who underwent CT with standard- (n = 31) or lower-dose (n = 34) protocol were included. SSDE was 54% lower for the lower-dose than for the standard-dose group (1.9 ± 0.4 vs 4.1 ± 0.8 mGy). Lower-dose DLR and MBIR yielded lower image noise and higher SNR and CNR than standard-dose HIR (p < .05). Interobserver agreement on subjective features ranged from a kappa coefficient of 0.68 to 0.78. The readers subjectively scored noise texture, edge sharpness, and overall quality lower for lower-dose MBIR than for standard-dose HIR (p < .001), though higher for lower-dose DLR than for standard-dose HIR (p < .001). In the phantom, DLR provided higher d' than HIR and MBIR at each dose. Object detectability was greater for 2.0-mGy DLR than for 4.0-mGy HIR for low-contrast (3.67 vs 3.57) and high-contrast (1.20 vs 1.04) objects. CONCLUSION. Compared with IR algorithms, DLR results in substantial dose reduction with preserved or even improved image quality for low-tube-voltage pediatric CT. CLINICAL IMPACT. Use of DLR at 80 kVp allows greater dose reduction for pediatric CT than do current IR techniques.

Improved image quality and dose reduction in abdominal CT with deep-learning reconstruction algorithm: a phantom study

OBJECTIVES

To assess the impact of a new artificial intelligence deep-learning reconstruction (Precise Image; AI-DLR) algorithm on image quality against a hybrid iterative reconstruction (IR) algorithm in abdominal CT for different clinical indications.

METHODS

Acquisitions on phantoms were performed at 5 dose levels (CTDI_vol: 13/11/9/6/1.8 mGy). Raw data were reconstructed using level 4 of iDose⁴ (i4) and 3 levels of AI-DLR (Smoother/Smooth/Standard). Noise power spectrum (NPS), task-based transfer function (TTF) and detectability index (d') were computed: d' modelled detection of a liver metastasis (LM) and hepatocellular carcinoma at portal (HCCp) and arterial (HCCa) phases. Image quality was subjectively assessed on an anthropomorphic phantom by 2 radiologists.

RESULTS

From Standard to Smoother levels, noise magnitude and average NPS spatial frequency decreased and the detectability (d') of all simulated lesions increased. For both inserts, TTF values were similar for all three AI-DLR levels from 13 to 6 mGy but decreased from Standard to Smoother levels at 1.8 mGy. Compared to the i4 used in clinical practice, d' values were higher using the Smoother and Smooth levels and close for the Standard level. For all dose levels, except at 1.8 mGy, radiologists considered images satisfactory for clinical use for the 3 levels of AI-DLR, but rated images too smooth using the Smoother level.

CONCLUSION

Use of the Smooth and Smoother levels of AI-DLR reduces the image noise and improves the detectability of lesions and spatial resolution for standard and low-dose levels. Using the Smooth level is apparently the best compromise between the lowest dose level and adequate image quality.

KEY POINTS

• Evaluation of the impact of a new artificial intelligence deep-learning reconstruction (AI-DLR) on image quality and dose compared to a hybrid iterative reconstruction (IR) algorithm. • The Smooth and Smoother levels of AI-DLR reduced the image noise and improved the detectability of lesions and spatial resolution for standard and low-dose levels. • The Smooth level seems the best compromise between the lowest dose level and adequate image quality.

Fully automated image quality evaluation on patient CT: Multi-vendor and multi-reconstruction study

While the recent advancements of computed tomography (CT) technology have contributed in reducing radiation dose and image noise, an objective evaluation of image quality in patient scans has not yet been established. In this study, we present a patient-specific CT image quality evaluation method that includes fully automated measurements of noise level, structure sharpness, and alteration of structure. This study used the CT images of 120 patients from four different CT scanners reconstructed with three types of algorithm: filtered back projection (FBP), vendor-specific iterative reconstruction (IR), and a vendor-agnostic deep learning model (DLM, ClariCT.AI, ClariPi Inc.). The structure coherence feature (SCF) was used to divide an image into the homogeneous (RH) and structure edge (RS) regions, which in turn were used to localize the regions of interests (ROIs) for subsequent analysis of image quality indices. The noise level was calculated by averaging the standard deviations from five randomly selected ROIs on RH, and the mean SCFs on RS was used to estimate the structure sharpness. The structure alteration was defined by the standard deviation ratio between RS and RH on the subtraction image between FBP and IR or DLM, in which lower structure alterations indicate successful noise reduction without degradation of structure details. The estimated structure sharpness showed a high correlation of 0.793 with manually measured edge slopes. Compared to FBP, IR and DLM showed 34.38% and 51.30% noise reduction, 2.87% and 0.59% lower structure sharpness, and 2.20% and -12.03% structure alteration, respectively, on an average. DLM showed statistically superior performance to IR in all three image quality metrics. This study is expected to contribute to enhance the CT protocol optimization process by allowing a high throughput and quantitative image quality evaluation during the introduction or adjustment of lower-dose CT protocol into routine practice.

Comparison of Three 3D Segmentation Software Tools for Hip Surgical Planning

In hip arthroplasty, preoperative planning is fundamental to reaching a successful surgery. Nowadays, several software tools for computed tomography (CT) image processing are available. However, research studies comparing segmentation tools for hip surgery planning for patients affected by osteoarthritic diseases or osteoporotic fractures are still lacking. The present work compares three different software from the geometric, dimensional, and usability perspectives to identify the best three-dimensional (3D) modelling tool for the reconstruction of pathological femoral heads. Syngo.via Frontier (by Siemens Healthcare) is a medical image reading and post-processing software that allows low-skilled operators to produce prototypes. Materialise (by Mimics) is a commercial medical modelling software. 3D Slicer (by slicer.org) is an open-source development platform used in medical and biomedical fields. The 3D models reconstructed starting from the in vivo CT images of the pathological femoral head are compared with the geometries obtained from the laser scan of the in vitro bony specimens. The results show that Mimics and 3D Slicer are better for dimensional and geometric accuracy in the 3D reconstruction, while syngo.via Frontier is the easiest to use in the hospital setting.

Comparison of a Deep Learning-Based Reconstruction Algorithm with Filtered Back Projection and Iterative Reconstruction Algorithms for Pediatric Abdominopelvic CT

Objective

To compare a deep learning-based reconstruction (DLR) algorithm for pediatric abdominopelvic computed tomography (CT) with filtered back projection (FBP) and iterative reconstruction (IR) algorithms.

Materials and Methods

Post-contrast abdominopelvic CT scans obtained from 120 pediatric patients (mean age ± standard deviation, 8.7 ± 5.2 years; 60 males) between May 2020 and October 2020 were evaluated in this retrospective study. Images were reconstructed using FBP, a hybrid IR algorithm (ASiR-V) with blending factors of 50% and 100% (AV50 and AV100, respectively), and a DLR algorithm (TrueFidelity) with three strength levels (low, medium, and high). Noise power spectrum (NPS) and edge rise distance (ERD) were used to evaluate noise characteristics and spatial resolution, respectively. Image noise, edge definition, overall image quality, lesion detectability and conspicuity, and artifacts were qualitatively scored by two pediatric radiologists, and the scores of the two reviewers were averaged. A repeated-measures analysis of variance followed by the Bonferroni post-hoc test was used to compare NPS and ERD among the six reconstruction methods. The Friedman rank sum test followed by the Nemenyi-Wilcoxon-Wilcox all-pairs test was used to compare the results of the qualitative visual analysis among the six reconstruction methods.

Results

The NPS noise magnitude of AV100 was significantly lower than that of the DLR, whereas the NPS peak of AV100 was significantly higher than that of the high- and medium-strength DLR (p < 0.001). The NPS average spatial frequencies were higher for DLR than for ASiR-V (p < 0.001). ERD was shorter with DLR than with ASiR-V and FBP (p < 0.001). Qualitative visual analysis revealed better overall image quality with high-strength DLR than with ASiR-V (p < 0.001).

Conclusion

For pediatric abdominopelvic CT, the DLR algorithm may provide improved noise characteristics and better spatial resolution than the hybrid IR algorithm.

Accuracy of two deep learning-based reconstruction methods compared with an adaptive statistical iterative reconstruction method for solid and ground-glass nodule volumetry on low-dose and ultra-low-dose chest computed tomography: A phantom study

No published studies have evaluated the accuracy of volumetric measurement of solid nodules and ground-glass nodules on low-dose or ultra-low-dose chest computed tomography, reconstructed using deep learning-based algorithms. This is an important issue in lung cancer screening. Our study aimed to investigate the accuracy of semiautomatic volume measurement of solid nodules and ground-glass nodules, using two deep learning-based image reconstruction algorithms (Truefidelity and ClariCT.AI), compared with iterative reconstruction (ASiR-V) in low-dose and ultra-low-dose settings. We performed computed tomography scans of solid nodules and ground-glass nodules of different diameters placed in a phantom at four radiation doses (120 kVp/220 mA, 120 kVp/90 mA, 120 kVp/40 mA, and 80 kVp/40 mA). Each scan was reconstructed using Truefidelity, ClariCT.AI, and ASiR-V. The solid nodule and ground-glass nodule volumes were measured semiautomatically. The gold-standard volumes could be calculated using the diameter since all nodule phantoms are perfectly spherical. Subsequently, absolute percentage measurement errors of the measured volumes were calculated. Image noise was also calculated. Across all nodules at all dose settings, the absolute percentage measurement errors of Truefidelity and ClariCT.AI were less than 11%; they were significantly lower with Truefidelity or ClariCT.AI than with ASiR-V (all P<0.05). The absolute percentage measurement errors for the smallest solid nodule (3 mm) reconstructed by Truefidelity or ClariCT.AI at all dose settings were significantly lower than those of this nodule reconstructed by ASiR-V (all P<0.05). Furthermore, the lowest absolute percentage measurement errors for ground-glass nodules were observed with Truefidelity or ClariCT.AI at all dose settings. The absolute percentage measurement errors for ground-glass nodules reconstructed with Truefidelity at ultra-low-dose settings were significantly lower than those of all sizes of ground-glass nodules reconstructed with ASiR-V (all P<0.05). Image noise was lowest with Truefidelity (all P<0.05). In conclusion, the deep learning-based algorithms were more accurate for volume measurements of both solid nodules and ground-glass nodules than ASiR-V at both low-dose and ultra-low-dose settings.

Impact of an artificial intelligence deep-learning reconstruction algorithm for CT on image quality and potential dose reduction: A phantom study

Background

Recently, computed tomography (CT) manufacturers have developed deep‐learning‐based reconstruction algorithms to compensate for the limitations of iterative reconstruction (IR) algorithms, such as image smoothing and the spatial resolution's dependence on contrast and dose levels.

Purpose

To assess the impact of an artificial intelligence deep‐learning reconstruction (AI‐DLR) algorithm on image quality and dose reduction compared with a hybrid IR algorithm in chest CT for different clinical indications.

Methods

Acquisitions on the CT American College of Radiology (ACR) 464 and CT Torso CTU‐41 phantoms were performed at five dose levels (CTDI_vol: 9.5/7.5/6/2.5/0.4 mGy) used for chest CT conditions. Raw data were reconstructed using filtered backprojection, two levels of IR (iDose⁴ levels 4 (i4) and 7 (i7)), and five levels of AI‐DLR (Precise Image; Smoother, Smooth, Standard, Sharp, Sharper). Noise power spectrum (NPS), task‐based transfer function, and detectability index (d′) were computed: d′‐modeled detection of a soft tissue mediastinal nodule (low‐contrast soft tissue chest nodule within the mediastinum [LCN]), ground‐glass opacity (GGO), or high‐contrast pulmonary (HCP) lesion. The subjective image quality of chest anthropomorphic phantom images was independently evaluated by two radiologists. They assessed image noise, image smoothing, contrast between vessels and fat in the mediastinum for mediastinal images, visual border detection between bronchus and lung parenchyma for parenchymal images, and overall image quality using a commonly used four‐ or five‐point scale.

Results

From Standard to Smoother levels, on average, the noise magnitude decreased (for all dose levels: −66.3% ± 0.5% for mediastinal images and −63.1% ± 0.1% for parenchymal images), the average NPS spatial frequency decreased (for all dose levels: −35.3% ± 2.2% for mediastinal images and −13.3% ± 2.2% for parenchymal images), and the detectability (d′) of the three lesions increased. The opposite pattern was found from Standard to Sharper levels. From Smoother to Sharper levels, the spatial resolution increased for the low‐contrast polyethylene insert and the opposite for the high‐contrast air insert. Compared to the i4 used in clinical practice, d′ values were higher using Smoother (mean for all dose levels: 338.7% ± 29.4%), Smooth (103.4% ± 11.2%), and Standard (34.1% ± 6.6%) levels for the LCN on mediastinal images and Smoother (169.5% ± 53.2% for GGO and 136.9% ± 1.6% for HCP) and Smooth (36.4% ± 22.1% and 24.1% ± 0.9%, respectively) levels for parenchymal images. Radiologists considered the images satisfactory for clinical use at these levels, but adaptation to the dose level of the protocol is required.

Conclusion

With AI‐DLR, the smoothest levels reduced the noise and improved the detectability of chest lesions but increased the image smoothing. The opposite was found with the sharpest levels. The choice of level depends on the dose level and type of image: mediastinal or parenchymal.

Radiation dose optimization potential of deep learning-based reconstruction for multiphase hepatic CT: A clinical and phantom study

PURPOSE

This clinical and phantom study aimed to evaluate the impact of deep learning-based reconstruction (DLR) on image quality and its radiation dose optimization capability for multiphase hepatic CT relative to hybrid iterative reconstruction (HIR).

METHODS

Task-based image quality was assessed with a physical evaluation phantom; the high- and low-contrast detectability of HIR and DLR images were computed from the noise power spectrum and task-based transfer function at five different size-specific dose estimate (SSDE) values in the range 5.3 to 18.0-mGy. For the clinical study, images of 73 patients who had undergone multiphase hepatic CT under both standard-dose (STD) and lower-dose (LD) examination protocols within a time interval of about four-months on average, were retrospectively examined. STD images were reconstructed with HIR, while LD with HIR (LD-HIR) and DLR (LD-DLR). SSDE, quantitative image noise, and contrast-to-noise ratio (CNR) were compared between protocols. The noise magnitude, noise texture, streak artifact, image sharpness, interface smoothness, and overall image quality were subjectively rated by two independent radiologists.

RESULTS

In phantom study, the high- and low-contrast detectability of DLR images obtained at 5.3-mGy and 7.3-mGy, respectively, were slightly higher than those obtained with HIR at the STD protocol dose (18.0-mGy). In clinical study, LD-DLR yielded lower image noise, higher CNR, and higher subjective scores for all evaluation criteria than STD (all, p ≤ 0.05), despite having 52.8% lower SSDE (8.0 ± 2.5 vs. 16.8 ± 3.4-mGy).

CONCLUSIONS

DLR improved the subjective and objective image quality of multiphase hepatic CT compared with HIR techniques, even at approximately half the radiation dose.

Low-Cost Probabilistic 3D Denoising with Applications for Ultra-Low-Radiation Computed Tomography

We propose a pipeline for synthetic generation of personalized Computer Tomography (CT) images, with a radiation exposure evaluation and a lifetime attributable risk (LAR) assessment. We perform a patient-specific performance evaluation for a broad range of denoising algorithms (including the most popular deep learning denoising approaches, wavelets-based methods, methods based on Mumford−Shah denoising, etc.), focusing both on accessing the capability to reduce the patient-specific CT-induced LAR and on computational cost scalability. We introduce a parallel Probabilistic Mumford−Shah denoising model (PMS) and show that it markedly-outperforms the compared common denoising methods in denoising quality and cost scaling. In particular, we show that it allows an approximately 22-fold robust patient-specific LAR reduction for infants and a 10-fold LAR reduction for adults. Using a normal laptop, the proposed algorithm for PMS allows cheap and robust (with a multiscale structural similarity index >90%) denoising of very large 2D videos and 3D images (with over 107 voxels) that are subject to ultra-strong noise (Gaussian and non-Gaussian) for signal-to-noise ratios far below 1.0. The code is provided for open access.

Deep learning versus iterative image reconstruction algorithm for head CT in trauma

PURPOSE

To compare the image quality between a deep learning-based image reconstruction algorithm (DLIR) and an adaptive statistical iterative reconstruction algorithm (ASiR-V) in noncontrast trauma head CT.

METHODS

Head CT scans from 94 consecutive trauma patients were included. Images were reconstructed with ASiR-V 50% and the DLIR strengths: low (DLIR-L), medium (DLIR-M), and high (DLIR-H). The image quality was assessed quantitatively and qualitatively and compared between the different reconstruction algorithms. Inter-reader agreement was assessed by weighted kappa.

RESULTS

DLIR-M and DLIR-H demonstrated lower image noise (p < 0.001 for all pairwise comparisons), higher SNR of up to 82.9% (p < 0.001), and higher CNR of up to 53.3% (p < 0.001) compared to ASiR-V. DLIR-H outperformed other DLIR strengths (p ranging from < 0.001 to 0.016). DLIR-M outperformed DLIR-L (p < 0.001) and ASiR-V (p < 0.001). The distribution of reader scores for DLIR-M and DLIR-H shifted towards higher scores compared to DLIR-L and ASiR-V. There was a tendency towards higher scores with increasing DLIR strengths. There were fewer non-diagnostic CT series for DLIR-M and DLIR-H compared to ASiR-V and DLIR-L. No images were graded as non-diagnostic for DLIR-H regarding intracranial hemorrhage. The inter-reader agreement was fair-good between the second most and the less experienced reader, poor-moderate between the most and the less experienced reader, and poor-fair between the most and the second most experienced reader.

CONCLUSION

The image quality of trauma head CT series reconstructed with DLIR outperformed those reconstructed with ASiR-V. In particular, DLIR-M and DLIR-H demonstrated significantly improved image quality and fewer non-diagnostic images. The improvement in qualitative image quality was greater for the second most and the less experienced readers compared to the most experienced reader.

Reduced-Dose Deep Learning Reconstruction for Abdominal CT of Liver Metastases

Background Assessment of liver lesions is constrained as CT radiation doses are lowered; evidence suggests deep learning reconstructions mitigate such effects. Purpose To evaluate liver metastases and image quality between reduced-dose deep learning image reconstruction (DLIR) and standard-dose filtered back projection (FBP) contrast-enhanced abdominal CT. Materials and Methods In this prospective Health Insurance Portability and Accountability Act-compliant study (September 2019 through April 2021), participants with biopsy-proven colorectal cancer and liver metastases at baseline CT underwent standard-dose and reduced-dose portal venous abdominal CT in the same breath hold. Three radiologists detected and characterized lesions at standard-dose FBP and reduced-dose DLIR, reported confidence, and scored image quality. Contrast-to-noise ratios for liver metastases were recorded. Summary statistics were reported, and a generalized linear mixed model was used. Results Fifty-one participants (mean age ± standard deviation, 57 years ± 13; 31 men) were evaluated. The mean volume CT dose index was 65.1% lower with reduced-dose CT (12.2 mGy) than with standard-dose CT (34.9 mGy). A total of 161 lesions (127 metastases, 34 benign lesions) with a mean size of 0.7 cm ± 0.3 were identified. Subjective image quality of reduced-dose DLIR was superior to that of standard-dose FBP (P < .001). The mean contrast-to-noise ratio for liver metastases of reduced-dose DLIR (3.9 ± 1.7) was higher than that of standard-dose FBP (3.5 ± 1.4) (P < .001). Differences in detection were identified only for lesions 0.5 cm or smaller: 63 of 65 lesions detected with standard-dose FBP (96.9%; 95% CI: 89.3, 99.6) and 47 lesions with reduced-dose DLIR (72.3%; 95% CI: 59.8, 82.7). Lesion accuracy with standard-dose FBP and reduced-dose DLIR was 80.1% (95% CI: 73.1, 86.0; 129 of 161 lesions) and 67.1% (95% CI: 59.3, 74.3; 108 of 161 lesions), respectively (P = .01). Lower lesion confidence was reported with a reduced dose (P < .001). Conclusion Deep learning image reconstruction (DLIR) improved CT image quality at 65% radiation dose reduction while preserving detection of liver lesions larger than 0.5 cm. Reduced-dose DLIR demonstrated overall inferior characterization of liver lesions and reader confidence. Clinical trial registration no. NCT03151564 © RSNA, 2022 Online supplemental material is available for this article.

Evaluation of Apparent Noise on CT Images Using Moving Average Filters

This study aims to devise a simple method for evaluating the magnitude of texture noise (apparent noise) observed on computed tomography (CT) images scanned at a low radiation dose and reconstructed using iterative reconstruction (IR) and deep learning reconstruction (DLR) algorithms, and to evaluate the apparent noise in CT images reconstructed using the filtered back projection (FBP), IR, and two types of DLR (AiCE Body and AiCE Body Sharp) algorithms. We set a square region of interest (ROI) on CT images of standard- and obese-sized low-contrast phantoms, slid different-sized moving average filters in the ROI vertically and horizontally in steps of 1 pixel, and calculated the standard deviation (SD) of the mean CT values for each filter size. The SD of the mean CT values was fitted with a curve inversely proportional to the filter size, and an apparent noise index was determined from the curve-fitting formula. The apparent noise index of AiCE Body Sharp images for a given mAs value was approximately 58, 23, and 18% lower than that of the FBP, AIDR 3D, and AiCE Body images, respectively. The apparent noise index was considered to reflect noise power spectrum values at lower spatial frequency. Moreover, the apparent noise index was inversely proportional to the square roots of the mAs values. Thus, the apparent noise index could be a useful indicator to quantify and compare texture noise on CT images obtained with different scan parameters and reconstruction algorithms.

Artificial Intelligence in Radiation Therapy

Artificial intelligence (AI) has great potential to transform the clinical workflow of radiotherapy. Since the introduction of deep neural networks, many AI-based methods have been proposed to address challenges in different aspects of radiotherapy. Commercial vendors have started to release AI-based tools that can be readily integrated to the established clinical workflow. To show the recent progress in AI-aided radiotherapy, we have reviewed AI-based studies in five major aspects of radiotherapy including image reconstruction, image registration, image segmentation, image synthesis, and automatic treatment planning. In each section, we summarized and categorized the recently published methods, followed by a discussion of the challenges, concerns, and future development. Given the rapid development of AI-aided radiotherapy, the efficiency and effectiveness of radiotherapy in the future could be substantially improved through intelligent automation of various aspects of radiotherapy.

Effect of a new deep learning image reconstruction algorithm for abdominal computed tomography imaging on image quality and dose reduction compared with two iterative reconstruction algorithms: a phantom study

Background

New reconstruction algorithms based on deep learning have been developed to correct the image texture changes related to the use of iterative reconstruction algorithms. The purpose of this study was to evaluate the impact of a new deep learning image reconstruction [Advanced intelligent Clear-IQ Engine (AiCE)] algorithm on image-quality and dose reduction compared to a hybrid iterative reconstruction (AIDR 3D) algorithm and a model-based iterative reconstruction (FIRST) algorithm.

Methods

Acquisitions were carried out using the ACR 464 phantom (and its body ring) at six dose levels (volume computed tomography dose index 15/10/7.5/5/2.5/1 mGy). Raw data were reconstructed using three levels (Mild/Standard/Strong) of AIDR 3D, of FIRST and AiCE. Noise-power-spectrum (NPS) and task-based transfer function (TTF) were computed. Detectability index was computed to model the detection of a small calcification (1.5-mm diameter and 500 HU) and a large mass in the liver (25-mm diameter and 120 HU).

Results

NPS peaks were lower with AiCE than with AIDR 3D (-41%±6% for all levels) or FIRST (-15%±6% for Strong level and -41%±11% for both other levels). The average NPS spatial frequency was lower with AICE than AIDR 3D (-9%±2% using Mild and -3%±2% using Strong) but higher than FIRST for Standard (6%±3%) and Strong (25%±3%) levels. For acrylic insert, values of TTF at 50 percent were higher with AICE than AIDR 3D and FIRST, except for Mild level (-6%±6% and -13%±3%, respectively). For bone insert, values of TTF at 50 percent were higher with AICE than AIDR 3D but lower than FIRST (-19%±14%). For both simulated lesions, detectability index values were higher with AICE than AIDR 3D and FIRST (except for Strong level and for the small feature; -21%±14%). Using the Standard level, dose could be reduced by -79% for the small calcification and -57% for the large mass using AICE compared to AIDR 3D.

Conclusions

The new deep learning image reconstruction algorithm AiCE generates an image-quality with less noise and/or less smudged/smooth images and a higher detectability than the AIDR 3D or FIRST algorithms. The outcomes of our phantom study suggest a good potential of dose reduction using AiCE but it should be confirmed clinically in patients.

Virtual monoenergetic images from dual-energy CT: systematic assessment of task-based image quality performance

Background

To compare task-based image quality (TB-IQ) among virtual monoenergetic images (VMI) and linear-blended images (LBI) from dual-energy CT as a function of contrast task, radiation dose, size, and lesion diameter.

Methods

A TB-IQ phantom (Mercury Phantom 4.0, Sun Nuclear Corporation) was imaged on a third-generation dual-source dual-energy CT with 100/Sn150 kVp at three volume CT dose levels (5, 10, 15 mGy). Three size sections (diameters 16, 26, 36 cm) with subsections for image noise and spatial resolution analysis were used. High-contrast tasks (e.g., calcium-containing stone and vascular lesion) were emulated using bone and iodine inserts. A low-contrast task (e.g., low-contrast lesion or hematoma) was emulated using a polystyrene insert. VMI at 40-190 keV and LBI were reconstructed. Noise power spectrum (NPS) determined the noise magnitude and texture. Spatial resolution was assessed using the task-transfer function (TTF) of the three inserts. The detectability index (d') served as TB-IQ metric.

Results

Noise magnitude increased with increasing phantom size, decreasing dose, and decreasing VMI-energy. Overall, noise magnitude was higher for VMI at 40-60 keV compared to LBI (range of noise increase, 3-124%). Blotchier noise texture was found for low and high VMIs (40-60 keV, 130-190 keV) compared to LBI. No difference in spatial resolution was observed for high contrast tasks. d' increased with increasing dose level or lesion diameter and decreasing size. For high-contrast tasks, d' was higher at 40-80 keV and lower at high VMIs. For the low-contrast task, d' was higher for VMI at 70-90 keV and lower at 40-60 keV.

Conclusions

Task-based image quality differed among VMI-energy and LBI dependent on the contrast task, dose level, phantom size, and lesion diameter. Image quality could be optimized by tailoring VMI-energy to the contrast task. Considering the clinical relevance of iodine, VMIs at 50-60 keV could be proposed as an alternative to LBI.

Low-contrast detectability and potential for radiation dose reduction using deep learning image reconstruction—A 20-reader study on a semi-anthropomorphic liver phantom

Background

A novel deep learning image reconstruction (DLIR) algorithm for CT has recently been clinically approved.

Purpose

To assess low-contrast detectability and dose reduction potential for CT images reconstructed with the DLIR algorithm and compare with filtered back projection (FBP) and hybrid iterative reconstruction (IR).

Material and methods

A customized upper-abdomen phantom containing four cylindrical liver inserts with low-contrast lesions was scanned at CT dose indexes of 5, 10, 15, 20 and 25 mGy. Images were reconstructed with FBP, 50% hybrid IR (IR50), and DLIR of low strength (DLL), medium strength (DLM) and high strength (DLH). Detectability was assessed by 20 independent readers using a two-alternative forced choice approach. Dose reduction potential was estimated separately for each strength of DLIR using a fitted model, with the detectability performance of FBP and IR50 as reference.

Results

For the investigated dose levels of 5 and 10 mGy, DLM improved detectability compared to FBP by 5.8 and 6.9 percentage points (p.p.), and DLH improved detectability by 9.6 and 12.3 p.p., respectively (all p < .007). With IR50 as reference, DLH improved detectability by 5.2 and 9.8 p.p. for the 5 and 10 mGy dose level, respectively (p < .03). With respect to this low-contrast detectability task, average dose reduction potential relative to FBP was estimated to 39% for DLM and 55% for DLH. Relative to IR50, average dose reduction potential was estimated to 21% for DLM and 42% for DLH.

Conclusions:

Low-contrast detectability performance is improved when applying a DLIR algorithm, with potential for radiation dose reduction.

•

Deep learning image reconstruction improves low-contrast detectability in CT.

•

Performance improved with increasing strength of deep learning image reconstruction.

•

Results suggest potential for CT radiation dose reduction.

Performance of a deep learning-based CT image denoising method: Generalizability over dose, reconstruction kernel and slice thickness

PURPOSE

Deep learning (DL) is rapidly finding applications in low-dose CT image denoising. While having the potential to improve image quality (IQ) over the filtered back projection method (FBP) and produce images quickly, performance generalizability of the data-driven DL methods is not fully understood yet. The main purpose of this work is to investigate the performance generalizability of a low-dose CT image denoising neural network in data acquired under different scan conditions, particularly relating to these three parameters: reconstruction kernel, slice thickness and dose (noise) level. A secondary goal is to identify any underlying data property associated with the CT scan settings that might help predict the generalizability of the denoising network.

METHODS

We select the residual encoder-decoder convolutional neural network (REDCNN) as an example of a low-dose CT image denoising technique in this work. To study how the network generalizes on the three imaging parameters, we grouped the CT volumes in the Low-Dose Grand Challenge (LDGC) data into three pairs of training datasets according to their imaging parameters, changing only one parameter in each pair. We trained REDCNN with them to obtain six denoising models. We test each denoising model on datasets of matching and mismatching parameters with respect to its training sets regarding dose, reconstruction kernel and slice thickness, respectively, to evaluate the denoising performance changes. Denoising performances are evaluated on patient scans, simulated phantom scans and physical phantom scans using IQ metrics including mean squared error (MSE), contrast-dependent modulation transfer function (MTF), pixel-level noise power spectrum (pNPS) and low-contrast lesion detectability (LCD).

RESULTS

REDCNN had larger MSE when the testing data was different from the training data in reconstruction kernel, but no significant MSE difference when varying slice thickness in the testing data. REDCNN trained with quarter-dose data had slightly worse MSE in denoising higher-dose images than that trained with mixed-dose data (17-80%). The MTF tests showed that REDCNN trained with the two reconstruction kernels and slice thicknesses yielded images of similar image resolution. However, REDCNN trained with mixed-dose data preserved the low-contrast resolution better compared to REDCNN trained with quarter-dose data. In the pNPS test, it was found that REDCNN trained with smooth-kernel data could not remove high-frequency noise in the test data of sharp kernel, possibly because the lack of high-frequency noise in the smooth-kernel data limited the ability of the trained model in removing high-frequency noise. Finally, in the LCD test, REDCNN improved the lesion detectability over the original FBP images regardless of whether the training and testing data had matching reconstruction kernels.

CONCLUSIONS

REDCNN is observed to be poorly generalizable between reconstruction kernels, more robust in denoising data of arbitrary dose levels when trained with mixed-dose data, and not highly sensitive to slice thickness. It is known that reconstruction kernel affects the in-plane pNPS shape of a CT image whereas slice thickness and dose level do not, so it is possible that the generalizability performance of this CT image denoising network highly correlates to the pNPS similarity between the testing and training data. This article is protected by copyright. All rights reserved.

Intelligent Noise Reduction Algorithm to Evaluate the Correlation between Human Fat Deposits and Uterine Fibroids under Ultrasound Imaging

This study aimed to realize the automatic diagnosis of fatty degeneration of uterine fibroids. In this study, the traditional nonlocal means (NLM) algorithm was improved by changing the Euclidean distance and introducing a cosine function and applied to the ultrasonic imaging intelligent diagnosis of patients with fatty degeneration of uterine fibroids. Then, the noise reduction effect of the improved NLM algorithm was evaluated based on several indicators, such as peak signal-to-noise ratio (PSNR), mean square error (MSE), contrast-to-noise ratio (CNR), figure of merit (FOM), and structural similarity (SSIM). The accuracy, sensitivity, specificity, and F1 score were adopted to evaluate the improved NLM algorithm for diagnosing fatty degeneration of uterine fibroids, and the Perona-Malik (PM) algorithm and NLM algorithm were used for comparative analysis. The results showed that after the ultrasound images of patients with uterine fibroids were denoised using the improved NLM algorithm, the PSNR, MSE, CNR, FOM, and SSIM were obviously better than the same indicators of the image processed with the PM algorithm and the NLM algorithm, and the differences were statistically significant (P < 0.05). The diagnosis results of patients with fatty degeneration of uterine fibroids found that there was only one patient with missed diagnosis after the ultrasound image was processed with NLM algorithm, and there was no statistical difference between the improved NLM algorithm and the assisted diagnosis accuracy of the pathological examination results (P > 0.05). The average noise reduction time of the PM algorithm, NLM algorithm, and the improved NLM algorithm was 16.38 ± 4.33 s, 18.01 ± 5.14 s, and 23.81 ± 4.62 s, respectively. The diagnosis rate before improvement was 75.0%, the diagnosis accuracy rate for PM was 79.69%, and that after improvement was 85.94%. In summary, the improved NLM algorithm showed a good noise reduction effect on ultrasound images of patients with uterine fibroids, could improve the diagnosis accuracy of fatty degeneration of uterine fibroids, and could assist clinicians in the ultrasound imaging diagnosis of patients with uterine fibroids.

Image quality in liver CT: low-dose deep learning vs standard-dose model-based iterative reconstructions

OBJECTIVES

To compare the overall image quality and detectability of significant (malignant and pre-malignant) liver lesions of low-dose liver CT (LDCT, 33.3% dose) using deep learning denoising (DLD) to standard-dose CT (SDCT, 100% dose) using model-based iterative reconstruction (MBIR).

METHODS

In this retrospective study, CT images of 80 patients with hepatic focal lesions were included. For noninferiority analysis of overall image quality, a margin of - 0.5 points (scored in a 5-point scale) for the difference between scan protocols was pre-defined. Other quantitative or qualitative image quality assessments were performed. Additionally, detectability of significant liver lesions was compared, with 64 pairs of CT, using the jackknife alternative free-response ROC analysis, with noninferior margin defined by the lower limit of 95% confidence interval (CI) of the difference of figure-of-merit less than - 0.1.

RESULTS

The mean overall image quality scores with LDCT and SDCT were 3.77 ± 0.38 and 3.94 ± 0.34, respectively, demonstrating a difference of - 0.17 (95% CI: - 0.21 to - 0.12), which did not cross the predefined noninferiority margin of - 0.5. Furthermore, LDCT showed significantly superior quantitative results of liver lesion contrast to noise ratio (p < 0.05). However, although LDCT scored higher than the average score in qualitative image quality assessments, they were significantly lower than those of SDCT (p < 0.05). Figure-of-merit for lesion detection was 0.859 for LDCT and 0.878 for SDCT, showing noninferiority (difference: - 0.019, 95% CI: - 0.058 to 0.021).

CONCLUSION

LDCT using DLD with 67% radiation dose reduction showed non-inferior overall image quality and lesion detectability, compared to SDCT.

KEY POINTS

• Low-dose liver CT using deep learning denoising (DLD), at 67% dose reduction, provided non-inferior overall image quality compared to standard-dose CT using model-based iterative reconstruction (MBIR). • Low-dose CT using DLD showed significantly less noise and higher CNR _{lesion to liver} than standard-dose CT using MBIR and demonstrated at least average image quality score among all readers, albeit with lower scores than standard-dose CT using MBIR. • Low-dose liver CT showed noninferior detectability for malignant and pre-malignant liver lesions, compared to standard-dose CT.

PULMONARY NODULE DETECTION IN CHEST CT USING A DEEP LEARNING-BASED RECONSTRUCTION ALGORITHM

This study's aim was to assess whether deep learning image reconstruction (DLIR) techniques are non-inferior to ASIR-V for the clinical task of pulmonary nodule detection in chest computed tomography. Up to 6 (range 3-6, mean 4.2) artificial lung nodules (diameter: 3, 5, 8 mm; density: -800, -630, +100 HU) were inserted at different locations in the Kyoto Kagaku Lungman phantom. In total, 16 configurations (10 abnormal, 6 normal) were scanned at 7.6, 3, 1.6 and 0.38 mGy CTDIvol (respectively 0, 60, 80 and 95% dose reduction). Images were reconstructed using 50% ASIR-V and a deep learning-based algorithm with low (DL-L), medium (DL-M) and high (DL-H) strength. Four chest radiologists evaluated 256 series by locating and scoring nodules on a five-point scale. No statistically significant difference was found among the reconstruction algorithms (p = 0.987, average across readers AUC: 0.555, 0.561, 0.557, 0.558 for ASIR-V, DL-L, DL-M, DL-H).

Deep learning-based Hounsfield unit value measurement method for bolus tracking images in cerebral computed tomography angiography

BACKGROUND

Patient movement during bolus tracking (BT) impairs the accuracy of Hounsfield unit (HU) measurements. This study assesses the accuracy of measuring HU values in the internal carotid artery (ICA) using an original deep learning (DL)-based method as compared with using the conventional region of interest (ROI) setting method.

METHOD

A total of 722 BT images of 127 patients who underwent cerebral computed tomography angiography were selected retrospectively and divided into groups for training data, validation data, and test data. To segment the ICA using our proposed method, DL was performed using a convolutional neural network. The HU values in the ICA were obtained using our DL-based method and the ROI setting method. The ROI setting was performed with and without correcting for patient body movement (corrected ROI and settled ROI). We compared the proposed DL-based method with settled ROI to evaluate HU value differences from the corrected ROI, based on whether or not patients experienced involuntary movement during BT image acquisition.

RESULTS

Differences in HU values from the corrected ROI in the settled ROI and the proposed method were 23.8 ± 12.7 HU and 9.0 ± 6.4 HU in patients with body movement and 1.1 ± 1.6 HU and 3.9 ± 4.7 HU in patients without body movement, respectively. There were significant differences in both comparisons (P < 0.01).

CONCLUSION

DL-based method can improve the accuracy of HU value measurements for ICA in BT images with patient involuntary movement.

Comparison of two versions of a deep learning image reconstruction algorithm on CT image quality and dose reduction: A phantom study

PURPOSE

To compare the impact on CT image quality and dose reduction of two versions of a Deep Learning Image Reconstruction algorithm.

MATERIAL AND METHODS

Acquisitions on the CT ACR 464 phantom were performed at five dose levels (CTDI_vol : 10/7.5/5/2.5/1 mGy) using chest or abdomen pelvis protocol parameters. Raw data were reconstructed using the filtered-back projection (FBP), the enhanced level of AIDR 3D (AIDR 3De), and the three levels of AiCE (Mild, Standard, and Strong) for the two versions (AiCE V8 vs AiCE V10). The noise power spectrum (NPS) and task-based transfer function (TTF) for bone (high-contrast insert) and acrylic (low-contrast insert) inserts were computed. To quantify the changes of noise magnitude and texture, the square root of the area under the NPS curve and the average spatial frequency (f_av ) of the NPS curve were measured. The detectability index (d') was computed to model the detectability of either a large mass in the liver or lung, or a small calcification or high contrast tissue boundaries.

RESULTS

The noise magnitude was lower with both AiCE versions than with AIDR 3De. The noise magnitude was lower with AiCE V10 than with AiCE V8 (-4 ± 6% for Mild, -13 ± 3% for Standard, and -48 ± 0% for Strong levels). f_av and TTF_50% values for both inserts shifted towards higher frequencies with AiCE than with AIDR 3De. Compared to AiCE V08, f_av shifted towards higher frequencies with AiCE V10 (45 ± 4%, 36 ± 3%, and 5 ± 4% for all levels, respectively). The TTF_50% values shifted towards higher frequencies with AiCE V10 as compared with AiCE V8 for both inserts, except for the Strong level for the acrylic insert. Whatever the dose and AiCE levels, d' values were higher with AiCE V10 than with AiCE V8 for the small object/calcification and for the large object/lesion.

CONCLUSION

As compared to AIDR 3De, lower noise magnitude and higher spatial resolution and detectability index were found with both versions of AiCE. As compared to AiCE V8, AiCE V10 reduced noise and improved spatial resolution and detectability without changing the noise texture in a simple geometric phantom, except for the Strong level. AiCE V10 seems to have a greater potential for dose reduction than AiCE V8.

Dose reduction potential of vendor-agnostic deep learning model in comparison with deep learning-based image reconstruction algorithm on CT: a phantom study

Objectives

To compare the dose reduction potential (DRP) of a vendor-agnostic deep learning model (DLM, ClariCT.AI) with that of a vendor-specific deep learning–based image reconstruction algorithm (DLR, TrueFidelity™).

Methods

Computed tomography (CT) images of a multi-sized image quality phantom (Mercury v4.0) were acquired under six radiation dose levels (0.48/0.97/1.93/3.87/7.74/15.47 mGy) and were reconstructed using filtered back projection (FBP) and three strength levels of the DLR (low/medium/high). The FBP images were denoised using the DLM. For all DLM and DLR images, the detectability index (d′) (a task-based detection performance metric) was obtained, under various combinations of three target sizes (10/5/1 mm), five inlets (CT value difference with the background; −895/50/90/335/1000 HU), five phantom diameters (36/31/26/21/16 cm), and six radiation dose levels. Dose reduction potential (DRP) measures the dose reduction made by using DLM or DLR, while yielding d′ equivalent to that of FBP at full dose.

Results

The DRPs of the DLM, DLR-low, DLR-medium, and DLR-high were 86% (81–88%), 60% (46–67%), 76% (60–81%), and 87% (78–92%), respectively. For 10-mm targets, the DRP of the DLM (87%) was higher than that of all DLR algorithms (58–86%). However, for smaller targets (5 mm/1 mm), the DRPs of the DLR-high (89/88%) were greater than those of the DLM (87/84%).

Conclusion

The dose reduction potential of the vendor-agnostic DLM was shown to be comparable to that of the vendor-specific DLR at high strength and superior to those of the DLRs at medium and low strengths.

Key Points

• DRP of the vendor-agnostic model was comparable to that of high-strength vendor-specific model and superior to those of medium- and low-strength models.

• Under various radiation dose levels, the deep learning model shows higher detectability indexes compared to FBP.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00330-021-08199-9.

Diagnostic performance and image quality of deep learning image reconstruction (DLIR) on unenhanced low-dose abdominal CT for urolithiasis

BACKGROUND

Patients with urolithiasis undergo radiation overexposure from computed tomography (CT) scans. Improvement of image reconstruction is necessary for radiation dose reduction.

PURPOSE

To evaluate a deep learning-based reconstruction algorithm for CT (DLIR) in the detection of urolithiasis at low-dose non-enhanced abdominopelvic CT.

MATERIAL AND METHODS

A total of 75 patients who underwent low-dose abdominopelvic CT for urolithiasis were retrospectively included. Each examination included three reconstructions: DLIR; filtered back projection (FBP); and hybrid iterative reconstruction (IR; ASiR-V 70%). Image quality was subjectively and objectively assessed using attenuation and noise measurements in order to calculate the signal-to-noise ratio (SNR), absolute contrast, and contrast-to-noise ratio (CNR). Attenuation of the largest stones were also compared. Detectability of urinary stones was assessed by two observers.

RESULTS

Image noise was significantly reduced with DLIR: 7.2 versus 17 and 22 for ASiR-V 70% and FBP, respectively. Similarly, SNR and CNR were also higher compared to the standard reconstructions. When the structures had close attenuation values, contrast was lower with DLIR compared to ASiR-V. Attenuation of stones was also lowered in the DLIR series. Subjective image quality was significantly higher with DLIR. The detectability of all stones and stones >3 mm was excellent with DLIR for the two observers (intraclass correlation [ICC] = 0.93 vs. 0.96 and 0.95 vs. 0.99). For smaller stones (<3 mm), results were different (ICC = 0.77 vs. 0.86).

CONCLUSION

For low-dose abdominopelvic CT, DLIR reconstruction exhibited image quality superior to ASiR-V and FBP as well as an excellent detection of urinary stones.

Effect of deep learning image reconstruction in the prediction of resectability of pancreatic cancer: Diagnostic performance and reader confidence

OBJECTIVE

To assess the diagnostic performance and reader confidence in determining the resectability of pancreatic cancer at computed tomography (CT) using a new deep learning image reconstruction (DLIR) algorithm.

METHODS

A retrospective review was conduct of on forty-seven patients with pathologically confirmed pancreatic cancers who underwent baseline multiphasic contrast-enhanced CT scan. Image data sets were reconstructed using filtered back projection (FBP), hybrid model-based adaptive statistical iterative reconstruction (ASiR-V) 60 %, and DLIR "TrueFidelity" at low(L), medium(M), and high strength levels(H). Four board-certified abdominal radiologists reviewed the CT images and classified cancers as resectable, borderline resectable, or unresectable. Diagnostic performance and reader confidence for categorizing the resectability of pancreatic cancer were evaluated based on the reference standards, and the interreader agreement was assessed using Fleiss k statistics.

RESULTS

For prediction of margin-negative resections(ie, R0), the average area under the receiver operating characteristic curve was significantly higher with DLIR-H (0.91; 95 % confidence interval [CI]: 0.79, 0.98) than FBP (0.75; 95 % CI:0.60, 0.86) and ASiR-V (0.81; 95 % CI:0.67, 0.91) (p = 0.030 and 0.023 respectively). Reader confidence scores were significantly better using DLIR compared to FBP and ASiR-V 60 % and increased linearly with the increase of DLIR strength level (all p < 0.001). Among the image reconstructions, DLIR-H showed the highest interreader agreement in the resectability classification and lowest subject variability in the reader confidence.

CONCLUSIONS

The DLIR-H algorithm may improve the diagnostic performance and reader confidence in the CT assignment of the local resectability of pancreatic cancer while reducing the interreader variability.

Image texture, low contrast liver lesion detectability and impact on dose: Deep learning algorithm compared to partial model-based iterative reconstruction

OBJECTIVES

To compare deep learning (True Fidelity, TF) and partial model based Iterative Reconstruction (ASiR-V) algorithm for image texture, low contrast lesion detectability and potential dose reduction.

METHODS

Anthropomorphic phantoms (mimicking non-overweight and overweight patient), containing lesions of 6 mm in diameter with 20HU contrast, were scanned at five different dose levels (2,6,10,15,20 mGy) on a CT system, using clinical routine protocols for liver lesion detection. Images were reconstructed using ASiR-V 0% (surrogate for FBP), 60 % and TF at low, medium and high strength. Noise texture was characterized by computing a normalized Noise Power Spectrum filtered by an eye filter. The similarity against FBP texture was evaluated using peak frequency difference (PFD) and root mean square deviation (RMSD). Low contrast detectability was assessed using a channelized Hotelling observer and the area under the ROC curve (AUC) was used as figure of merit. Potential dose reduction was calculated to obtain the same AUC for TF and ASiR-V.

RESULTS

FBP-like noise texture was more preserved with TF (PFD from -0.043mm-1 to -0.09mm-1, RMSD from 0.12mm-1 to 0.21mm-1) than with ASiR-V (PFD equal to 0.12 mm-1, RMSD equal to 0.53mm-1), resulting in a sharper image. AUC was always higher with TF than ASIR-V. In average, TF compared to ASiR-V, enabled a radiation dose reduction potential of 7%, 25 % and 33 % for low, medium and high strength respectively.

CONCLUSION

Compared to ASIR-V, TF at high strength does not impact noise texture and maintains low contrast liver lesions detectability at significant lower dose.

Advanced CT techniques for assessing hepatocellular carcinoma

Hepatocellular carcinoma (HCC) is the sixth-most common cancer in the world, and hepatic dynamic CT studies are routinely performed for its evaluation. Ongoing studies are examining advanced imaging techniques that may yield better findings than are obtained with conventional hepatic dynamic CT scanning. Dual-energy CT-, perfusion CT-, and artificial intelligence-based methods can be used for the precise characterization of liver tumors, the quantification of treatment responses, and for predicting the overall survival rate of patients. In this review, the advantages and disadvantages of conventional hepatic dynamic CT imaging are reviewed and the general principles of dual-energy- and perfusion CT, and the clinical applications and limitations of these technologies are discussed with respect to HCC. Finally, we address the utility of artificial intelligence-based methods for diagnosing HCC.

Sinogram-based deep learning image reconstruction technique in abdominal CT: image quality considerations

OBJECTIVES

To investigate the image quality and perception of a sinogram-based deep learning image reconstruction (DLIR) algorithm for single-energy abdominal CT compared to standard-of-care strength of ASIR-V.

METHODS

In this retrospective study, 50 patients (62% F; 56.74 ± 17.05 years) underwent portal venous phase. Four reconstructions (ASIR-V at 40%, and DLIR at three strengths: low (DLIR-L), medium (DLIR-M), and high (DLIR-H)) were generated. Qualitative and quantitative image quality analysis was performed on the 200 image datasets. Qualitative scores were obtained for image noise, contrast, small structure visibility, sharpness, and artifact by three blinded radiologists on a 5-point scale (1, excellent; 5, very poor). Radiologists also indicated image preference on a 3-point scale (1, most preferred; 3, least preferred). Quantitative assessment was performed by measuring image noise and contrast-to-noise ratio (CNR).

RESULTS

DLIR had better image quality scores compared to ASIR-V. Scores on DLIR-H for noise (1.40 ± 0.53), contrast (1.41 ± 0.55), small structure visibility (1.51 ± 0.61), and sharpness (1.60 ± 0.54) were the best (p < 0.05) followed by DLIR-M (1.85 ± 0.52, 1.66 ± 0.57, 1.69 ± 0.59, 1.68 ± 0.46), DLIR-L (2.29 ± 0.58, 1.96 ± 0.61, 1.90 ± 0.65, 1.86 ± 0.46), and ASIR-V (2.86 ± 0.67, 2.55 ± 0.58, 2.34 ± 0.66, 2.01 ± 0.36). Ratings for artifacts were similar for all reconstructions (p > 0.05). DLIRs did not influence subjective textural perceptions and were preferred over ASIR-V from the beginning. All DLIRs had a higher CNR (26.38-102.30%) and lower noise (20.64-48.77%) than ASIR-V. DLIR-H had the best objective scores.

CONCLUSION

Sinogram-based deep learning image reconstructions were preferred over iterative reconstruction subjectively and objectively due to improved image quality and lower noise, even in large patients. Use in clinical routine may allow for radiation dose reduction.

KEY POINTS

• Deep learning image reconstructions (DLIRs) have a higher contrast-to-noise ratio compared to medium-strength hybrid iterative reconstruction techniques. • DLIR may be advantageous in patients with large body habitus due to a lower image noise. • DLIR can enable further optimization of radiation doses used in abdominal CT.

Improved image quality in abdominal computed tomography reconstructed with a novel Deep Learning Image Reconstruction technique - initial clinical experience

BACKGROUND

A novel Deep Learning Image Reconstruction (DLIR) technique for computed tomography has recently received clinical approval.

PURPOSE

To assess image quality in abdominal computed tomography reconstructed with DLIR, and compare with standardly applied iterative reconstruction.

MATERIAL AND METHODS

Ten abdominal computed tomography scans were reconstructed with iterative reconstruction and DLIR of medium and high strength, with 0.625 mm and 2.5 mm slice thickness. Image quality was assessed using eight visual grading criteria in a side-by-side comparative setting. All series were presented twice to evaluate intraobserver agreement. Reader scores were compared using univariate logistic regression. Image noise and contrast-to-noise ratio were calculated for quantitative analyses.

RESULTS

For 2.5 mm slice thickness, DLIR images were more frequently perceived as equal or better than iterative reconstruction across all visual grading criteria (for both DLIR of medium and high strength, p < 0.001). Correspondingly, DLIR images were more frequently perceived as better (as opposed to equal or in favor of iterative reconstruction) for visual reproduction of liver parenchyma, intrahepatic vascular structures as well as overall impression of image noise and texture (p < 0.001). This improved image quality was also observed for 0.625 mm slice images reconstructed with DLIR of high strength when directly comparing to traditional iterative reconstruction in 2.5 mm slices. Image noise was significantly lower and contrast-to-noise ratio measurements significantly higher for images reconstructed with DLIR compared to iterative reconstruction (p < 0.01).

CONCLUSIONS

Abdominal computed tomography images reconstructed using a DLIR technique shows improved image quality when compared to standardly applied iterative reconstruction across a variety of clinical image quality criteria.