Deep Learning Based Noise Reduction for Brain MR Imaging: Tests on Phantoms and Healthy Volunteers

Ultra-short Echo-time MR Angiography Combined with a Modified Signal Targeting Alternating Radio Frequency with Asymmetric Inversion Slabs Technique to Assess Visceral Artery Aneurysm after Coil Embolization

Contrast-enhanced CT and MR angiography are widely used for follow-up of visceral artery aneurysms after coil embolization. However, potential adverse reactions to contrast agents and image deterioration due to susceptibility artifacts from the coils are major drawbacks of these modalities. Herein, we introduced a novel non-contrast-enhanced MR angiography technique using ultra-short TE combined with a modified signal targeting alternating radio frequency with asymmetric inversion slabs, which could provide a serial hemodynamic vascular image with fewer susceptibility artifacts for follow-up after coil embolization.

Deep learning-based magnetic resonance imaging reconstruction for improving the image quality of reduced-field-of-view diffusion-weighted imaging of the pancreas

BACKGROUND

It has been reported that deep learning-based reconstruction (DLR) can reduce image noise and artifacts, thereby improving the signal-to-noise ratio and image sharpness. However, no previous studies have evaluated the efficacy of DLR in improving image quality in reduced-field-of-view (reduced-FOV) diffusion-weighted imaging (DWI) [field-of-view optimized and constrained undistorted single-shot (FOCUS)] of the pancreas. We hypothesized that a combination of these techniques would improve DWI image quality without prolonging the scan time but would influence the apparent diffusion coefficient calculation.

AIM

To evaluate the efficacy of DLR for image quality improvement of FOCUS of the pancreas.

METHODS

This was a retrospective study evaluated 37 patients with pancreatic cystic lesions who underwent magnetic resonance imaging between August 2021 and October 2021. We evaluated three types of FOCUS examinations: FOCUS with DLR (FOCUS-DLR+), FOCUS without DLR (FOCUS-DLR-), and conventional FOCUS (FOCUS-conv). The three types of FOCUS and their apparent diffusion coefficient (ADC) maps were compared qualitatively and quantitatively.

RESULTS

FOCUS-DLR+ (3.62, average score of two radiologists) showed significantly better qualitative scores for image noise than FOCUS-DLR- (2.62) and FOCUS-conv (2.88) (P < 0.05). Furthermore, FOCUS-DLR+ showed the highest contrast ratio (CR) between the pancreatic parenchyma and adjacent fat tissue for b-values of 0 and 600 s/mm2 (0.72 ± 0.08 and 0.68 ± 0.08) and FOCUS-DLR- showed the highest CR between cystic lesions and the pancreatic parenchyma for the b-values of 0 and 600 s/mm2 (0.62 ± 0.21 and 0.62 ± 0.21) (P < 0.05), respectively. FOCUS-DLR+ provided significantly higher ADCs of the pancreas and lesion (1.44 ± 0.24 and 3.00 ± 0.66) compared to FOCUS-DLR- (1.39 ± 0.22 and 2.86 ± 0.61) and significantly lower ADCs compared to FOCUS-conv (1.84 ± 0.45 and 3.32 ± 0.70) (P < 0.05), respectively.

CONCLUSION

This study evaluated the efficacy of DLR for image quality improvement in reduced-FOV DWI of the pancreas. DLR can significantly denoise images without prolonging the scan time or decreasing the spatial resolution. The denoising level of DWI can be controlled to make the images appear more natural to the human eye. However, this study revealed that DLR did not ameliorate pancreatic distortion. Additionally, physicians should pay attention to the interpretation of ADCs after DLR application because ADCs are significantly changed by DLR.

Ultrafast Brain MRI Protocol at 1.5 T Using Deep Learning and Multi-shot EPI

RATIONALE AND OBJECTIVES

To evaluate clinical feasibility and image quality of a comprehensive ultrafast brain MRI protocol with multi-shot echo planar imaging and deep learning-enhanced reconstruction at 1.5T.

MATERIALS AND METHODS

Thirty consecutive patients who underwent clinically indicated MRI at a 1.5 T scanner were prospectively included. A conventional MRI (c-MRI) protocol, including T1-, T2-, T2*-, T2-FLAIR, and diffusion-weighted images (DWI)-weighted sequences were acquired. In addition, ultrafast brain imaging with deep learning-enhanced reconstruction and multi-shot EPI (DLe-MRI) was performed. Subjective image quality was evaluated by three readers using a 4-point Likert scale. To assess interrater agreement, Fleiss' kappa (ϰ) was determined. For objective image analysis, relative signal intensity levels for grey matter, white matter, and cerebrospinal fluid were calculated.

RESULTS

Time of acquisition (TA) of c-MRI protocols added up to 13:55 minutes, whereas the TA of DLe-MRI-based protocol added up to 3:04 minutes, resulting in a time reduction of 78%. All DLe-MRI acquisitions yielded diagnostic image quality with good absolute values for subjective image quality. C-MRI demonstrated slight advantages for DWI in overall subjective image quality (c-MRI: 3.93 [+/- 0.25] vs DLe-MRI: 3.87 [+/- 0.37], P = .04) and diagnostic confidence (c-MRI: 3.93 [+/- 0.25] vs DLe-MRI: 3.83 [+/- 3.83], P = .01). For most evaluated quality scores, moderate interobserver agreement was found. Objective image evaluation revealed comparable results for both techniques.

CONCLUSION

DLe-MRI is feasible and allows for highly accelerated comprehensive brain MRI within 3minutes at 1.5 T with good image quality. This technique may potentially strengthen the role of MRI in neurological emergencies.

Applicability Evaluation of Full-Reference Image Quality Assessment Methods for Computed Tomography Images

Image quality assessments (IQA) are an important task for providing appropriate medical care. Full-reference IQA (FR-IQA) methods, such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), are often used to evaluate imaging conditions, reconstruction conditions, and image processing algorithms, including noise reduction and super-resolution technology. However, these IQA methods may be inapplicable for medical images because they were designed for natural images. Therefore, this study aimed to investigate the correlation between objective assessment by some FR-IQA methods and human subjective assessment for computed tomography (CT) images. For evaluation, 210 distorted images were created from six original images using two types of degradation: noise and blur. We employed nine widely used FR-IQA methods for natural images: PSNR, SSIM, feature similarity (FSIM), information fidelity criterion (IFC), visual information fidelity (VIF), noise quality measure (NQM), visual signal-to-noise ratio (VSNR), multi-scale SSIM (MSSSIM), and information content-weighted SSIM (IWSSIM). Six observers performed subjective assessments using the double stimulus continuous quality scale (DSCQS) method. The performance of IQA methods was quantified using Pearson's linear correlation coefficient (PLCC), Spearman rank order correlation coefficient (SROCC), and root-mean-square error (RMSE). Nine FR-IQA methods developed for natural images were all strongly correlated with the subjective assessment (PLCC and SROCC > 0.8), indicating that these methods can apply to CT images. Particularly, VIF had the best values for all three items, PLCC, SROCC, and RMSE. These results suggest that VIF provides the most accurate alternative measure to subjective assessments for CT images.

Noise reduction by multiple path neural network using Attention mechanisms with an emphasis on robustness against Errors: A pilot study on brain Diffusion-Weighted images

PURPOSE

In deep learning-based noise reduction, larger networks offer advanced and complex functionality by utilizing its greater degree of freedom, but come with increased unpredictability, raising the potential risk of unforeseen errors. Here, we introduce a novel denoising model for diffusion-weighted images that intentionally limits the network output freedom by incorporating multiple pathways with varying degrees of freedom, with the aim of minimizing the chance of unintended alterations to the input. The purpose of this pilot study is to assess the model's ability to perform effective denoising under the constraints.

METHODS

Images from 10 healthy volunteers were used. Key innovations in our model development include: (1) neural network architecture that separated the function for calculating the specific output values from the function for adjusting the calculation for each pixel and (2) training that optimised the network based on both image and secondary obtained diffusion tensor. The generated images were compared with the original ones by measuring the deviation from ground truth images (averaged across eight acquisitions).

RESULTS

The generated images demonstrated closer alignment with the ground truth images, both visually and statistically (Q < 0.05), compared to the original images. Furthermore, the advantage of the generated images over the original images was also found in the secondary obtained quantitative parameter maps with significance (Q < 0.05).

CONCLUSION

The usefulness of the proposed method was suggested because it was successful in improving both the quality of the generated images and accuracy of the major diffusion parameter maps under the given restrictions.

Feasibility study of super-resolution deep learning-based reconstruction using k-space data in brain diffusion-weighted images

PURPOSE

The purpose of this study is to evaluate the influence of super-resolution deep learning-based reconstruction (SR-DLR), which utilizes k-space data, on the quality of images and the quantitation of the apparent diffusion coefficient (ADC) for diffusion-weighted images (DWI) in brain magnetic resonance imaging (MRI).

METHODS

A retrospective analysis was performed on 34 patients who had undergone DWI using a 3 T MRI system with SR-DLR reconstruction based on k-space data in August 2022. DWI was reconstructed with SR-DLR (Matrix = 684 × 684) and without SR-DLR (Matrix = 228 × 228). Measurements were made of the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) in white matter (WM) and grey matter (GM), and the full width at half maximum (FWHM) of the septum pellucidum. Two radiologists assessed image noise, contrast, artifacts, blur, and the overall quality of three image types using a four-point scale. Quantitative and qualitative scores between images with and without SR-DLR were compared using the Wilcoxon signed-rank test.

RESULTS

Images with SR-DLR showed significantly higher SNRs and CNRs than those without SR-DLR (p < 0.001). No statistically significant variances were found in the apparent diffusion coefficients (ADCs) in WM and GM between images with and without SR-DLR (ADC in WM, p = 0.945; ADC in GM, p = 0.235). Moreover, the FWHM without SR-DLR was notably lower compared to that with SR-DLR (p < 0.001).

CONCLUSION

SR-DLR has the potential to augment the quality of DWI in DL MRI scans without significantly impacting ADC quantitation.

Denoising approach with deep learning-based reconstruction for neuromelanin-sensitive MRI: image quality and diagnostic performance

PURPOSE

Neuromelanin-sensitive MRI (NM-MRI) has proven useful for diagnosing Parkinson's disease (PD) by showing reduced signals in the substantia nigra (SN) and locus coeruleus (LC), but requires a long scan time. The aim of this study was to assess the image quality and diagnostic performance of NM-MRI with a shortened scan time using a denoising approach with deep learning-based reconstruction (dDLR).

MATERIALS AND METHODS

We enrolled 22 healthy volunteers, 22 non-PD patients and 22 patients with PD who underwent NM-MRI, and performed manual ROI-based analysis. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) in ten healthy volunteers were compared among images with a number of excitations (NEX) of 1 (NEX1), NEX1 images with dDLR (NEX1 + dDLR) and 5-NEX images (NEX5). Acquisition times for NEX1 and NEX5 were 3 min 12 s and 15 min 58 s, respectively. Diagnostic performances using the contrast ratio (CR) of the SN (CR_SN) and LC (CR_LC) and those by visual assessment for differentiating PD from non-PD were also compared between NEX1 and NEX1 + dDLR.

RESULTS

Image quality analyses revealed that SNRs and CNRs of the SN and LC in NEX1 + dDLR were significantly higher than in NEX1, and comparable to those in NEX5. In diagnostic performance analysis, areas under the receiver operating characteristic curve (AUC) using CR_SN and CR_LC of NEX1 + dDLR were 0.87 and 0.75, respectively, which had no significant difference with those of NEX1. Visual assessment showed improvement of diagnostic performance by applying dDLR.

CONCLUSION

Image quality for NEX1 + dDLR was comparable to that of NEX5. dDLR has the potential to reduce scan time of NM-MRI without degrading image quality. Both 1-NEX NM-MRI with and without dDLR showed high AUCs for diagnosing PD by CR. The results of visual assessment suggest advantages of dDLR. Further tuning of dDLR would be expected to provide clinical merits in diagnosing PD.

Deep learning-based reconstruction and 3D hybrid profile order technique for MRCP at 3T: evaluation of image quality and acquisition time

OBJECTIVES

To evaluate the image quality of the 3D hybrid profile order technique and deep-learning-based reconstruction (DLR) for 3D magnetic resonance cholangiopancreatography (MRCP) within a single breath-hold (BH) at 3 T magnetic resonance imaging (MRI).

METHODS

This retrospective study included 32 patients with biliary and pancreatic disorders. BH images were reconstructed with and without DLR. The signal-to-noise ratio (SNR), contrast, contrast-to-noise ratio (CNR) between the common bile duct (CBD) and periductal tissues, and full width at half maximum (FWHM) of CBD on 3D-MRCP were evaluated quantitatively. Two radiologists scored image noise, contrast, artifacts, blur, and overall image quality of the three image types using a 4-point scale. Quantitative and qualitative scores were compared using the Friedman test and post hoc Nemenyi test.

RESULTS

The SNR and CNR were not significantly different when under respiratory gating- and BH-MRCP without DLR. However, they were significantly higher under BH with DLR than under respiratory gating (SNR, p = 0.013; CNR, p = 0.027). The contrast and FWHM of MRCP under BH with and without DLR were lower than those under respiratory gating (contrast, p < 0.001; FWHM, p = 0.015). Qualitative scores for noise, blur, and overall image quality were higher under BH with DLR than those under respiratory gating (blur, p = 0.003; overall, p = 0.008).

CONCLUSIONS

The combination of the 3D hybrid profile order technique and DLR is useful for MRCP within a single BH and does not lead to the deterioration of image quality and space resolution at 3 T MRI.

CLINICAL RELEVANCE STATEMENT

Considering its advantages, this sequence might become the standard protocol for MRCP in clinical practice, at least at 3.0 T.

KEY POINTS

• The 3D hybrid profile order can achieve MRCP within a single breath-hold without a decrease in spatial resolution. • The DLR significantly improved the CNR and SNR of BH-MRCP. • The 3D hybrid profile order technique with DLR reduces the deterioration of image quality in MRCP within a single breath-hold.

Deep Learning-Based Denoising of CEST MR Data: A Feasibility Study on Applying Synthetic Phantoms in Medical Imaging

Chemical Exchange Saturation Transfer (CEST) magnetic resonance imaging (MRI) provides a novel method for analyzing biomolecule concentrations in tissues without exogenous contrast agents. Despite its potential, achieving a high signal-to-noise ratio (SNR) is imperative for detecting small CEST effects. Traditional metrics such as Magnetization Transfer Ratio Asymmetry (MTRasym) and Lorentzian analyses are vulnerable to image noise, hampering their precision in quantitative concentration estimations. Recent noise-reduction algorithms like principal component analysis (PCA), nonlocal mean filtering (NLM), and block matching combined with 3D filtering (BM3D) have shown promise, as there is a burgeoning interest in the utilization of neural networks (NNs), particularly autoencoders, for imaging denoising. This study uses the Bloch-McConnell equations, which allow for the synthetic generation of CEST images and explores NNs efficacy in denoising these images. Using synthetically generated phantoms, autoencoders were created, and their performance was compared with traditional denoising methods using various datasets. The results underscored the superior performance of NNs, notably the ResUNet architectures, in noise identification and abatement compared to analytical approaches across a wide noise gamut. This superiority was particularly pronounced at elevated noise intensities in the in vitro data. Notably, the neural architectures significantly improved the PSNR values, achieving up to 35.0, while some traditional methods struggled, especially in low-noise reduction scenarios. However, the application to the in vivo data presented challenges due to varying noise profiles. This study accentuates the potential of NNs as robust denoising tools, but their translation to clinical settings warrants further investigation.

Model-based Deep Learning Reconstruction Using a Folded Image Training Strategy for Abdominal 3D T1-weighted Imaging

PURPOSE

To evaluate the feasibility of folded image training strategy (FITS) and the quality of images reconstructed using the improved model-based deep learning (iMoDL) network trained with FITS (FITS-iMoDL) for abdominal MR imaging.

METHODS

This retrospective study included abdominal 3D T1-weighted images of 122 patients. In the experimental analyses, peak SNR (PSNR) and structure similarity index (SSIM) of images reconstructed with FITS-iMoDL were compared with those with the following reconstruction methods: conventional model-based deep learning (conv-MoDL), MoDL trained with FITS (FITS-MoDL), total variation regularized compressed sensing (CS), and parallel imaging (CG-SENSE). In the clinical analysis, SNR and image contrast were measured on the reference, FITS-iMoDL, and CS images. Three radiologists evaluated the image quality using a 5-point scale to determine the mean opinion score (MOS).

RESULTS

The PSNR of FITS-iMoDL was significantly higher than that of FITS-MoDL, conv-MoDL, CS, and CG-SENSE (P < 0.001). The SSIM of FITS-iMoDL was significantly higher than those of the others (P < 0.001), except for FITS-MoDL (P = 0.056). In the clinical analysis, the SNR of FITS-iMoDL was significantly higher than that of the reference and CS (P < 0.0001). Image contrast was equivalent within an equivalence margin of 10% among these three image sets (P < 0.0001). MOS was significantly improved in FITS-iMoDL (P < 0.001) compared with CS images in terms of liver edge and vessels conspicuity, lesion depiction, artifacts, blurring, and overall image quality.

CONCLUSION

The proposed method, FITS-iMoDL, allowed a deeper MoDL reconstruction network without increasing memory consumption and improved image quality on abdominal 3D T1-weighted imaging compared with CS images.

Evaluation of late gadolinium enhancement cardiac MRI using deep learning reconstruction

BACKGROUND

Deep learning (DL)-based methods have been used to improve the imaging quality of magnetic resonance imaging (MRI) by denoising.

PURPOSE

To assess the effects of DL-based MR reconstruction (DLR) method on late gadolinium enhancement (LGE) image quality.

MATERIAL AND METHODS

A total of 85 patients who underwent cardiovascular magnetic resonance (CMR) examination, including LGE imaging using conventional construction and DLR with varying levels of noise reduction (NR) levels, were included. Both magnitude LGE (MLGE) and phase-sensitive LGE (PSLGE) images were reviewed independently by double-blinded observers who used a 5-point Likert scale for multiple measures regarding image quality. Meanwhile, the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and edge sharpness of images were calculated and compared between conventional LGE imaging and DLR LGE imaging.

RESULTS

Both MLGE and PSLGE with DLR at 50% and 75% noise reduction levels received significantly higher scores than conventional imaging for overall imaging quality (all P < 0.01). In addition, the SNR, CNR, and edge sharpness of all DLR LGE imaging are higher than conventional imaging (all P < 0.01). The highest subjective score and best image quality is obtained when the DLR noise reduction level is at 75%.

CONCLUSION

DLR reduced image noise while improving image contrast and sharpness in the cardiovascular LGE imaging.

Deep learning reconstruction for brain diffusion-weighted imaging: efficacy for image quality improvement, apparent diffusion coefficient assessment, and intravoxel incoherent motion evaluation in in vitro and in vivo studies

PURPOSE

Deep learning reconstruction (DLR) to improve imaging quality has already been introduced, but no studies have evaluated the effect of DLR on diffusion-weighted imaging (DWI) or intravoxel incoherent motion (IVIM) in in vitro or in vivo studies. The purpose of this study was to determine the effect of DLR for magnetic resonance imaging (MRI) in terms of image quality improvement, apparent diffusion coefficient (ADC) assessment, and IVIM index evaluation on DWI through in vitro and in vivo studies.

METHODS

For the in vitro study, a phantom recommended by the Quantitative Imaging Biomarkers Alliance was scanned and reconstructed with and without DLR, and 15 patients with brain tumors with normal-appearing gray and white matter examined using IVIM and reconstructed with and without DLR were included in the in vivo study. The ADCs of all phantoms for DWI with and without DLR, as well as the coefficient of variation percentage (CV%), and ADCs and IVIM indexes for each participant, were evaluated based on DWI with and without DLR by means of region-of-interest measurements. For the in vitro study, using the mean ADCs for all phantoms, a t-test was adopted to compare DWI with and without DLR. For the in vivo study, a Wilcoxon signed-rank test was used to compare the CV% between the two types of DWI. In addition, the Wilcoxon signed-rank test was used to compare the ADC, true diffusion coefficient (D), pseudodiffusion coefficient (D*), and percentage of water molecules in micro perfusion within 1 voxel (f) with and without DLR; the limits of agreement of each parameter were determined through a Bland-Altman analysis.

RESULTS

The in vitro study identified no significant differences between the ADC values for DWI with and without DLR (P > 0.05), and the CV% was significantly different for DWI with and without DLR (P < 0.05) when b values ≥250 s/mm2 were used. The in vivo study revealed that D* and f with and without DLR were significantly different (P < 0.001). The limits of agreement of the ADC, D, and D* values for DWI with and without DLR were determined as 0.00 ± 0.51 × 10-3, 0.00 ± 0.06 × 10-3, and 1.13 ± 4.04 × 10-3 mm2/s, respectively. The limits of agreement of the f values for DWI with and without DLR were determined as -0.01 ± 0.07.

CONCLUSION

Deep learning reconstruction for MRI has the potential to significantly improve DWI quality at higher b values. It has some effect on D* and f values in the IVIM index evaluation, but ADC and D values are less affected by DLR.

A Feasibility Study on Deep Learning Reconstruction to Improve Image Quality With PROPELLER Acquisition in the Setting of T2-Weighted Gynecologic Pelvic Magnetic Resonance Imaging

OBJECTIVES

Evaluate deep learning (DL) to improve the image quality of the PROPELLER (Periodically Rotated Overlapping Parallel Lines with Enhanced Reconstruction technique) for 3 T magnetic resonance imaging of the female pelvis.

METHODS

Three radiologists prospectively and independently compared non-DL and DL PROPELLER sequences from 20 patients with a history of gynecologic malignancy. Sequences with different noise reduction factors (DL 25%, DL 50%, and DL 75%) were blindly reviewed and scored based on artifacts, noise, relative sharpness, and overall image quality. The generalized estimating equation method was used to assess the effect of methods on the Likert scales. Quantitatively, the contrast-to-noise ratio and signal-to-noise ratio (SNR) of the iliac muscle were calculated, and pairwise comparisons were performed based on a linear mixed model. P values were adjusted using the Dunnett method. Interobserver agreement was assessed using the κ statistic. P value was considered statistically significant at less than 0.05.

RESULTS

Qualitatively, DL 50 and DL 75 were ranked as the best sequences in 86% of cases. Images generated by the DL method were significantly better than non-DL images ( P < 0.0001). Iliacus muscle SNR on DL 50 and DL 75 was significantly better than non-DL images ( P < 0.0001). There was no difference in contrast-to-noise ratio between the DL and non-DL techniques in the iliac muscle. There was a high percent agreement (97.1%) in terms of DL sequences' superior image quality (97.1%) and sharpness (100%) relative to non-DL images.

CONCLUSION

The utilization of DL reconstruction improves the image quality of PROPELLER sequences with improved SNR quantitatively.

The role of AI in prostate MRI quality and interpretation: Opportunities and challenges

Prostate MRI plays an important role in imaging the prostate gland and surrounding tissues, particularly in the diagnosis and management of prostate cancer. With the widespread adoption of multiparametric magnetic resonance imaging in recent years, the concerns surrounding the variability of imaging quality have garnered increased attention. Several factors contribute to the inconsistency of image quality, such as acquisition parameters, scanner differences and interobserver variabilities. While efforts have been made to standardize image acquisition and interpretation via the development of systems, such as PI-RADS and PI-QUAL, the scoring systems still depend on the subjective experience and acumen of humans. Artificial intelligence (AI) has been increasingly used in many applications, including medical imaging, due to its ability to automate tasks and lower human error rates. These advantages have the potential to standardize the tasks of image interpretation and quality control of prostate MRI. Despite its potential, thorough validation is required before the implementation of AI in clinical practice. In this article, we explore the opportunities and challenges of AI, with a focus on the interpretation and quality of prostate MRI.

MRI Advancements in Musculoskeletal Clinical and Research Practice

Over the past decades, MRI has become increasingly important for diagnosing and longitudinally monitoring musculoskeletal disorders, with ongoing hardware and software improvements aiming to optimize image quality and speed. However, surging demand for musculoskeletal MRI and increased interest to provide more personalized care will necessitate a stronger emphasis on efficiency and specificity. Ongoing hardware developments include more powerful gradients, improvements in wide-bore magnet designs to maintain field homogeneity, and high-channel phased-array coils. There is also interest in low-field-strength magnets with inherently lower magnetic footprints and operational costs to accommodate global demand in middle- and low-income countries. Previous approaches to decrease acquisition times by means of conventional acceleration techniques (eg, parallel imaging or compressed sensing) are now largely overshadowed by deep learning reconstruction algorithms. It is expected that greater emphasis will be placed on improving synthetic MRI and MR fingerprinting approaches to shorten overall acquisition times while also addressing the demand of personalized care by simultaneously capturing microstructural information to provide greater detail of disease severity. Authors also anticipate increased research emphasis on metal artifact reduction techniques, bone imaging, and MR neurography to meet clinical needs.

Deep Learning Convolutional Neural Network Reconstruction and Radial k-Space Acquisition MR Technique for Enhanced Detection of Retropatellar Cartilage Lesions of the Knee Joint

OBJECTIVES

To assess diagnostic performance of standard radial k-space (PROPELLER) MRI sequences and compare with accelerated acquisitions combined with a deep learning-based convolutional neural network (DL-CNN) reconstruction for evaluation of the knee joint.

METHODS

Thirty-five patients undergoing MR imaging of the knee at 1.5 T were prospectively included. Two readers evaluated image quality and diagnostic confidence of standard and DL-CNN accelerated PROPELLER MR sequences using a four-point Likert scale. Pathological findings of bone, cartilage, cruciate and collateral ligaments, menisci, and joint space were analyzed. Inter-reader agreement (IRA) for image quality and diagnostic confidence was assessed using intraclass coefficients (ICC). Cohen's Kappa method was used for evaluation of IRA and consensus between sequences in assessing different structures. In addition, image quality was quantitatively evaluated by signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) measurements.

RESULTS

Mean acquisition time of standard vs. DL-CNN sequences was 10 min 3 s vs. 4 min 45 s. DL-CNN sequences showed significantly superior image quality and diagnostic confidence compared to standard MR sequences. There was moderate and good IRA for assessment of image quality in standard and DL-CNN sequences with ICC of 0.524 and 0.830, respectively. Pathological findings of the knee joint could be equally well detected in both sequences (κ-value of 0.8). Retropatellar cartilage could be significantly better assessed on DL-CNN sequences. SNR and CNR was significantly higher for DL-CNN sequences (both p < 0.05).

CONCLUSIONS

In MR imaging of the knee, DL-CNN sequences showed significantly higher image quality and diagnostic confidence compared to standard PROPELLER sequences, while reducing acquisition time substantially. Both sequences perform comparably in the detection of knee-joint pathologies, while DL-CNN sequences are superior for evaluation of retropatellar cartilage lesions.

Commercially Available Deep-learning-reconstruction of MR Imaging of the Knee at 1.5T Has Higher Image Quality Than Conventionally-reconstructed Imaging at 3T: A Normal Volunteer Study

PURPOSE

This study aimed to evaluate whether the image quality of 1.5T magnetic resonance imaging (MRI) of the knee is equal to or higher than that of 3T MRI by applying deep learning reconstruction (DLR).

METHODS

Proton density-weighted images of the right knee of 27 healthy volunteers were obtained by 3T and 1.5T MRI scanners using similar imaging parameters (21 for high resolution image and 6 for normal resolution image). Commercially available DLR was applied to the 1.5T images to obtain 1.5T/DLR images. The 3T and 1.5T/DLR images were compared subjectively for visibility of structures, image noise, artifacts, and overall diagnostic acceptability and objectively. One-way ANOVA and Friedman tests were used for the statistical analyses.

RESULTS

For the high resolution images, all of the anatomical structures, except for bone, were depicted significantly better on the 1.5T/DLR compared with 3T images. Image noise scored statistically lower and overall diagnostic acceptability scored higher on the 1.5T/DLR images. The contrast between lateral meniscus and articular cartilage of the 1.5T/DLR images was significantly higher (5.89 ± 1.30 vs. 4.34 ± 0.87, P < 0.001), and also the contrast between medial meniscus and articular cartilage of the 1.5T/DLR images was significantly higher (5.12 ± 0.93 vs. 3.87 ± 0.56, P < 0.001). Similar image quality improvement by DLR was observed for the normal resolution images.

CONCLUSION

The 1.5T/DLR images can achieve less noise, more precise visualization of the meniscus and ligaments, and higher overall image quality compared with the 3T images acquired using a similar protocol.

Focused Abbreviated Survey MRI Protocols for Brain and Spine Imaging

There has been extensive growth in both the technical development and the clinical applications of MRI, establishing this modality as one of the most powerful diagnostic imaging tools. However, long examination and image interpretation times still limit the application of MRI, especially in emergent clinical settings. Rapid and abbreviated MRI protocols have been developed as alternatives to standard MRI, with reduced imaging times, and in some cases limited numbers of sequences, to more efficiently answer specific clinical questions. A group of rapid MRI protocols used at the authors' institution, referred to as FAST (focused abbreviated survey techniques), are designed to include or exclude emergent or urgent conditions or screen for specific entities. These FAST protocols provide adequate diagnostic image quality with use of accelerated approaches to produce imaging studies faster than traditional methods. FAST protocols have become critical diagnostic screening tools at the authors' institution, allowing confident and efficient confirmation or exclusion of actionable findings. The techniques commonly used to reduce imaging times, the imaging protocols used at the authors' institution, and future directions in FAST imaging are reviewed to provide a practical and comprehensive overview of FAST MRI for practicing neuroradiologists. ©RSNA, 2023 Quiz questions for this article are available in the supplemental material.

Clinical Impact of Deep Learning Reconstruction in MRI

Deep learning has been recognized as a paradigm-shifting tool in radiology. Deep learning reconstruction (DLR) has recently emerged as a technology used in the image reconstruction process of MRI, which is an essential procedure in generating MR images. Denoising, which is the first DLR application to be realized in commercial MRI scanners, improves signal-to-noise ratio. When applied to lower magnetic field-strength scanners, the signal-to-noise ratio can be increased without extending the imaging time, and image quality is comparable to that of higher-field-strength scanners. Shorter imaging times decrease patient discomfort and reduce MRI scanner running costs. The incorporation of DLR into accelerated acquisition imaging techniques, such as parallel imaging or compressed sensing, shortens the reconstruction time. DLR is based on supervised learning using convolutional layers and is divided into the following three categories: image domain, k-space learning, and direct mapping types. Various studies have reported other derivatives of DLR, and several have shown the feasibility of DLR in clinical practice. Although DLR efficiently reduces Gaussian noise from MR images, denoising makes image artifacts more prominent, and a solution to this problem is desired. Depending on the training of the convolutional neural network, DLR may change the imaging features of lesions and obscure small lesions. Therefore, radiologists may need to adopt the habit of questioning whether any information has been lost on images that appear clean. ^©RSNA, 2023 Quiz questions for this article are available in the supplemental material.

High-Resolution Single-Shot Fast Spin-Echo MR Imaging with Deep Learning Reconstruction Algorithm Can Improve Repeatability and Reproducibility of Follicle Counting

OBJECTIVE

To investigate the diagnostic performance of high-resolution single-shot fast spin-echo (SSFSE) imaging with deep learning (DL) reconstruction algorithm on follicle counting and compare it with original SSFSE images and conventional fast spin-echo (FSE) images.

METHODS

This study included 20 participants (40 ovaries) with clinically confirmed polycystic ovary syndrome (PCOS) who underwent high-resolution ovary MRI, including three-plane T2-weighted FSE sequences and slice-matched T2-weighted SSFSE sequences. A DL reconstruction algorithm was applied to the SSFSE sequences to generate SSFSE-DL images, and the original SSFSE images were also saved. Subjective evaluations such as the blurring artifacts, subjective noise, and clarity of the follicles on the SSFSE-DL, SSFSE, and conventional FSE images were independently conducted by two observers. Intra-class correlation coefficients and Bland-Altman plots were used to present the repeatability and reproducibility of the follicle number per ovary (FNPO) based on the three types of images.

RESULTS

SSFSE-DL images showed less blurring artifact, subjective noise, and better clarity of the follicles than SSFSE and FSE (p < 0.05). For the repeatability of the FNPO, SSFSE-DL showed the highest intra-observer (ICC = 0.930; 95% CI: 0.878-0.962) and inter-observer (ICC = 0.914; 95% CI: 0.843-0.953) agreements. The inter-observer 95% limits of agreement (LOA) for SSFSE-DL, SSFSE, and FSE ranged from -3.7 to 4.5, -4.4 to 7.0, and -7.1 to 7.6, respectively. The intra-observer 95% LOA for SSFSE-DL, SSFSE, and FSE ranged from -3.5 to 4.0, -5.1 to 6.1, and -5.7 to 4.2, respectively. The absolute values of intra-observer and inter-observer differences for SSFSE-DL were significantly lower than those for SSFSE and FSE (p < 0.05).

CONCLUSIONS

Compared with the original SSFSE images and the conventional FSE images, high-resolution SSFSE images with DL reconstruction algorithm can better display follicles, thus improving FNPO assessment.

Update on the Use of Artificial Intelligence in Hepatobiliary MR Imaging

The application of machine learning (ML) and deep learning (DL) in radiology has expanded exponentially. In recent years, an extremely large number of studies have reported about the hepatobiliary domain. Its applications range from differential diagnosis to the diagnosis of tumor invasion and prediction of treatment response and prognosis. Moreover, it has been utilized to improve the image quality of DL reconstruction. However, most clinicians are not familiar with ML and DL, and previous studies about these concepts are relatively challenging to understand. In this review article, we aimed to explain the concepts behind ML and DL and to summarize recent achievements in their use in the hepatobiliary region.

MR-self Noise2Noise: self-supervised deep learning-based image quality improvement of submillimeter resolution 3D MR images

OBJECTIVES

The study aimed to develop a deep neural network (DNN)-based noise reduction and image quality improvement by only using routine clinical scans and evaluate its performance in 3D high-resolution MRI.

METHODS

This retrospective study included T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) images from 185 clinical scans: 135 for DNN training, 11 for DNN validation, 20 for qualitative evaluation, and 19 for quantitative evaluation. Additionally, 18 vessel wall imaging (VWI) data were included to evaluate generalization. In each scan of the DNN training set, two noise-independent images were generated from the k-space data, resulting in an input-label pair. 2.5D U-net architecture was utilized for the DNN model. Qualitative evaluation between conventional MP-RAGE and DNN-based MP-RAGE was performed by two radiologists in image quality, fine structure delineation, and lesion conspicuity. Quantitative evaluation was performed with full sampled data as a reference by measuring quantitative error metrics and volumetry at 7 different simulated noise levels. DNN application on VWI was evaluated by two radiologists in image quality.

RESULTS

Our DNN-based MP-RAGE outperformed conventional MP-RAGE in all image quality parameters (average scores = 3.7 vs. 4.9, p < 0.001). In the quantitative evaluation, DNN showed better error metrics (p < 0.001) and comparable (p > 0.09) or better (p < 0.02) volumetry results than conventional MP-RAGE. DNN application to VWI also revealed improved image quality (3.5 vs. 4.6, p < 0.001).

CONCLUSION

The proposed DNN model successfully denoises 3D MR image and improves its image quality by using routine clinical scans only.

KEY POINTS

• Our deep learning framework successfully improved conventional 3D high-resolution MRI in all image quality parameters, fine structure delineation, and lesion conspicuity. • Compared to conventional MRI, the proposed deep neural network-based MRI revealed better quantitative error metrics and comparable or better volumetry results. • Deep neural network application to 3D MRI whose pulse sequences and parameters were different from the training data showed improvement in image quality, revealing the potential to generalize on various clinical MRI.

Combination Use of Compressed Sensing and Deep Learning for Shoulder Magnetic Resonance Imaging With Various Sequences

OBJECTIVE

For compressed sensing (CS) to become widely used in routine magnetic resonance imaging (MRI), it is essential to improve image quality. This study aimed to evaluate the usefulness of combining CS and deep learning-based reconstruction (DLR) for various sequences in shoulder MRI.

METHODS

This retrospective study included 37 consecutive patients who underwent undersampled shoulder MRIs, including T1-weighted (T1WI), T2-weighted (T2WI), and fat-saturation T2-weighted (FS-T2WI) images. Images were reconstructed using the conventional wavelet-based denoising method (wavelet method) and a combination of wavelet and DLR-based denoising methods (hybrid-DLR method) for each sequence. The signal-to-noise ratio and contrast-to-noise ratio of the bone, muscle, and fat and the full width at half maximum of the shoulder joint were compared between the 2 image types. In addition, 2 board-certified radiologists scored the image noise, contrast, sharpness, artifacts, and overall image quality of the 2 image types on a 4-point scale.

RESULTS

The signal-to-noise ratios and contrast-to-noise ratios of the bone, muscle, and fat in T1WI, T2WI, and FS-T2WI obtained from the hybrid-DLR method were significantly higher than those of the conventional wavelet method (P < 0.001). However, there were no significant differences in the full width at half maximum of the shoulder joint in any of the sequences (P > 0.05). Furthermore, in all sequences, the mean scores of the image noise, sharpness, artifacts, and overall image quality were significantly higher in the hybrid-DLR method than in the wavelet method (P < 0.001), but there were no significant differences in contrast among the sequences (P > 0.05).

CONCLUSIONS

The DLR denoising method can improve the image quality of CS in T1-weighted images, T2-weighted images, and fat-saturation T2-weighted images of the shoulder compared with the wavelet denoising method alone.

Deep Learning for Image Enhancement and Correction in Magnetic Resonance Imaging-State-of-the-Art and Challenges

Magnetic resonance imaging (MRI) provides excellent soft-tissue contrast for clinical diagnoses and research which underpin many recent breakthroughs in medicine and biology. The post-processing of reconstructed MR images is often automated for incorporation into MRI scanners by the manufacturers and increasingly plays a critical role in the final image quality for clinical reporting and interpretation. For image enhancement and correction, the post-processing steps include noise reduction, image artefact correction, and image resolution improvements. With the recent success of deep learning in many research fields, there is great potential to apply deep learning for MR image enhancement, and recent publications have demonstrated promising results. Motivated by the rapidly growing literature in this area, in this review paper, we provide a comprehensive overview of deep learning-based methods for post-processing MR images to enhance image quality and correct image artefacts. We aim to provide researchers in MRI or other research fields, including computer vision and image processing, a literature survey of deep learning approaches for MR image enhancement. We discuss the current limitations of the application of artificial intelligence in MRI and highlight possible directions for future developments. In the era of deep learning, we highlight the importance of a critical appraisal of the explanatory information provided and the generalizability of deep learning algorithms in medical imaging.

Acceleration of knee magnetic resonance imaging using a combination of compressed sensing and commercially available deep learning reconstruction: a preliminary study

PURPOSE

To evaluate whether deep learning reconstruction (DLR) accelerates the acquisition of 1.5-T magnetic resonance imaging (MRI) knee data without image deterioration.

MATERIALS AND METHODS

Twenty-one healthy volunteers underwent MRI of the right knee on a 1.5-T MRI scanner. Proton-density-weighted images with one or four numbers of signal averages (NSAs) were obtained via compressed sensing, and DLR was applied to the images with 1 NSA to obtain 1NSA-DLR images. The 1NSA-DLR and 4NSA images were compared objectively (by deriving the signal-to-noise ratios of the lateral and the medial menisci and the contrast-to-noise ratios of the lateral and the medial menisci and articular cartilages) and subjectively (in terms of the visibility of the anterior cruciate ligament, the medial collateral ligament, the medial and lateral menisci, and bone) and in terms of image noise, artifacts, and overall diagnostic acceptability. The paired t-test and Wilcoxon signed-rank test were used for statistical analyses.

RESULTS

The 1NSA-DLR images were obtained within 100 s. The signal-to-noise ratios (lateral: 3.27 ± 0.30 vs. 1.90 ± 0.13, medial: 2.71 ± 0.24 vs. 1.80 ± 0.15, both p < 0.001) and contrast-to-noise ratios (lateral: 2.61 ± 0.51 vs. 2.18 ± 0.58, medial 2.19 ± 0.32 vs. 1.97 ± 0.36, both p < 0.001) were significantly higher for 1NSA-DLR than 4NSA images. Subjectively, all anatomical structures (except bone) were significantly clearer on the 1NSA-DLR than on the 4NSA images. Also, in the former images, the noise was lower, and the overall diagnostic acceptability was higher.

CONCLUSION

Compared with the 4NSA images, the 1NSA-DLR images exhibited less noise, higher overall image quality, and allowed more precise visualization of the menisci and ligaments.

The mechanisms of arterial signal intensity profile in non-contrast coronary MRA (NC-MRCA): a 3D printed phantom investigation and clinical translations

Signal intensity (SI) drop has been proposed as an indirect stenosis assessment in non-contrast coronary MRA (NC-MRCA) but it uses unproven assumptions. We aimed to clarify the mechanisms that govern the SI in vitro and develop a stenosis detection method in vivo. Flow phantom tubes with/without stenosis were scanned under two spatial resolutions (0.5/1.0 mm³) on a 3.0 T MRI. Thirty-two coronary arteries from 11 volunteers were prospectively scanned with an EKG- and respiratory-gated 3D NC-MRCA with a resolution of 1.0 mm³, with coronary computed tomography angiography (CTA) as reference. The normalized SI along the centerline of the tubes or the coronary arteries was assessed against the distance from the orifice using a linear regression model. Its coefficient (SI decay slope) and goodness-of-fit (R2) were extracted to assess the effect of flow velocity and stenosis on the SI profile curve. The R2 was utilized for the stenosis detection. Phantom study: A slow flow velocity caused a steep SI decay slope. The SI drop revealed only at the inlet and outlet of stenosis due to the flow turbulence/vortex and yielded low R2, in which shape changed by the resolution. Clinical study: The R2 cutoff to detect ≥ 50% stenosis for the left and right coronary arteries were 0.64 and 0.20 with a sensitivity/specificity of 71.5/71.5 and 66.7/100 (%), respectively. The SI drop did not reflect the actual stenosis position and not suitable for the stenosis localization. The R2 cutoff represents an alternative method to detect stenoses on NC-MRCA at vessel level.Trial registration: ClinicalTrials.gov; NCT03768999, registered on December 7, 2018.

Deep learning-based image reconstruction improves radiologic evaluation of pituitary axis and cavernous sinus invasion in pituitary adenoma

PURPOSE

To compare performance of 1-mm deep learning reconstruction (DLR) with 3-mm routine MRI imaging for the delineation of pituitary axis and identification of cavernous sinus invasion for pituitary macroadenoma.

METHOD

This retrospective study included 104 patients (59.4 ± 13.1 years; 46 women) who underwent an MRI protocol including 1-mm deep learning-reconstructed and 3-mm routine images for evaluating pituitary adenoma between August 2019 and October 2020. Five readers (24, 9, 2 years, and <1 year of experience) assessed the delineation of pituitary axis (gland and stalk) and the presence of cavernous sinus invasion for using a pairwise design. The signal-to-noise ratio (SNR) was measured. Diagnostic performance as well as image preference data were analysed and compared according to the readers' experience using the McNemar test.

RESULTS

For delineation of normal pituitary axis, all readers preferred thin 1-mm DLR MRI over 3-mm MRI (overall superiority, 55.8 %, P <.001), with this preference being greater in the less experienced readers (92.3 % vs. 55.8 % [expert], P <.001). The readers showed higher diagnostic performance for cavernous sinus invasion on 1-mm (AUC, 0.91 and 0.92) than on 3-mm imaging (AUC, 0.87 and 0.88). The SNR of the 1-mm DLR was 1.21-fold higher than that of the routine 3-mm imaging.

CONCLUSION

Deep learning reconstruction-based 1-mm imaging demonstrates improved image quality and better delineation of microstructure in the sellar fossa and is preferred by both radiologists and non-radiologist physicians, especially in less experienced readers.

Application of deep learning-based image reconstruction in MR imaging of the shoulder joint to improve image quality and reduce scan time

OBJECTIVES

To compare the image quality and diagnostic performance of conventional motion-corrected periodically rotated overlapping parallel line with enhanced reconstruction (PROPELLER) MRI sequences with post-processed PROPELLER MRI sequences using deep learning-based (DL) reconstructions.

METHODS

In this prospective study of 30 patients, conventional (19 min 18 s) and accelerated MRI sequences (7 min 16 s) using the PROPELLER technique were acquired. Accelerated sequences were post-processed using DL. The image quality and diagnostic confidence were qualitatively assessed by 2 readers using a 5-point Likert scale. Analysis of the pathological findings of cartilage, rotator cuff tendons and muscles, glenoid labrum and subacromial bursa was performed. Inter-reader agreement was calculated using Cohen's kappa statistic. Quantitative evaluation of image quality was measured using the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR).

RESULTS

Mean image quality and diagnostic confidence in evaluation of all shoulder structures were higher in DL sequences (p value = 0.01). Inter-reader agreement ranged between kappa values of 0.155 (assessment of the bursa) and 0.947 (assessment of the rotator cuff muscles). In 17 cases, thickening of the subacromial bursa of more than 2 mm was only visible in DL sequences. The pathologies of the other structures could be properly evaluated by conventional and DL sequences. Mean SNR (p value = 0.01) and CNR (p value = 0.02) were significantly higher for DL sequences.

CONCLUSIONS

The accelerated PROPELLER sequences with DL post-processing showed superior image quality and higher diagnostic confidence compared to the conventional PROPELLER sequences. Subacromial bursa can be thoroughly assessed in DL sequences, while the other structures of the shoulder joint can be assessed in conventional and DL sequences with a good agreement between sequences.

KEY POINTS

• MRI of the shoulder requires long scan times and can be hampered by motion artifacts. • Deep learning-based convolutional neural networks are used to reduce image noise and scan time while maintaining optimal image quality. The radial k-space acquisition technique (PROPELLER) can reduce the scan time and has potential to reduce motion artifacts. • DL sequences show a higher diagnostic confidence than conventional sequences and therefore are preferred for assessment of the subacromial bursa, while conventional and DL sequences show comparable performance in the evaluation of the shoulder joint.

Usefulness of deep learning-based noise reduction for 1.5 T MRI brain images

AIM

To evaluate 1.5 T magnetic resonance imaging (MRI) brain images with denoising procedures using deep learning-based reconstruction (dDLR) relative to the original 1.5 and 3 T images.

MATERIALS AND METHODS

Eleven volunteers underwent MRI at 3 and 1.5 T. Two-dimensional fast spin-echo T2-weighted imaging (T2WI), fluid-attenuated inversion recovery (FLAIR) imaging and diffusion-weighted imaging (DWI) sequences were performed. The dDLR method was applied to the 1.5 T data (dDLR-1.5 T), then the image quality of the dDLR-1.5 T data relative to the original 1.5 T and 3 T data was qualitatively and quantitatively assessed based on the structure similarity (SSIM) index; the signal-to-noise ratios (SNRs) of the grey matter (GM) and white matter (WM); and the contrast-to-noise ratios (CNRs) between the GM and WM (CNRgm-wm) and between the striatum (ST) and WM (CNRst-wm).

RESULTS

The perceived image quality, and SNRs and CNRs were significantly higher for the dDLR-1.5 T images versus the 1.5 T images for all sequences and almost comparable or even superior to those of the 3 T images. For DWI, the SNRs and CNRst-wm were significantly higher for the dDLR-1.5 T images versus the 3 T images.

CONCLUSION

The dDLR technique improved the image quality of 1.5 T brain MRI images. With respect to qualitative and quantitative measurements, the denoised 1.5 T brain images were almost equivalent or even superior to the 3 T brain images.

Ultrafast cervcial spine MRI protocol using deep learning-based reconstruction: Diagnostic equivalence to a conventional protocol

PURPOSE

A major drawback of magnetic resonance imaging (MRI) is its limited imaging speed. This study proposed an ultrafast cervical spine MRI protocol (2 min 57 s) using deep learning-based reconstruction (DLR) and compared the diagnostic results to those of conventional MRI protocols (12 min 54 s).

METHODS

Fifty patients who underwent cervical spine MRI using both conventional and ultrafast protocols, including sagittal T1-weighted, T2-weighted, short-TI inversion recovery, and axial T2*-weighted imaging were included in this study. The ultrafast protocol shortened the acquisition time to approximately-one-fourth of that of the conventional protocol by reducing the phase matrix, oversampling rate, and number of excitations, and by applying compressed sensing. To compensate for the decreased signal-to-noise ratio caused by acceleration, noise reduction using DLR was performed. For image interpretation, three neuroradiologists graded or classified degenerative changes, including central canal stenosis, foraminal stenosis, endplate degeneration, disc degeneration, and disc hernia. The presence of other pathologies was also recorded. Given the absence of a reference standard, we tested the interchangeability of the two protocols by calculating the 95% confidence interval (CI) of the individual equivalence index. We also assessed the inter-protocol intra-reader agreement using kappa statistics.

RESULTS

Except for endplate degeneration, the 95 % CI of the individual equivalence index for all variables did not exceed 5 %, indicating interchangeability between the two protocols. The kappa values ranged from 0.600 to 0.977, indicating substantial to almost perfect agreement.

CONCLUSIONS

The proposed ultrafast MRI protocol yielded almost equivalent diagnostic results compared as the conventional protocol.

Application value of T2 fluid-attenuated inversion recovery sequence based on deep learning in static lacunar infarction

Background

Regular monitoring of static lacunar infarction (SLI) lesions plays an important role in preventing disease development and managing prognosis. Magnetic resonance imaging is one method used to monitor SLI lesions.

Purpose

To evaluate the image quality of the T2 fluid-attenuated inversion recovery (T2-FLAIR) sequence using artificial intelligence-assisted compressed sensing (ACS) in detecting SLI lesions and assess its clinical applicability.

Methods

A total of 42 patients were prospectively enrolled and scanned by T2-FLAIR. Two independent readers reviewed the images acquired with accelerated modes 1D (acceleration factor 2) and ACS (acceleration factors 2, 3, and 4). The overall image quality and lesion image quality were analyzed, as were signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and number of lesions between groups.

Results

The subjective assessment of overall brain image quality and lesion image quality was consistent between the two readers. The lesion display quality and the overall image quality were better with the traditional 1D acceleration method than with the ACS accelerated method. There was no significant difference in the SNR of the lacunar infarction in the images between the groups. The CNR of the images with the 1D acceleration mode was significantly lower than that of images with the ACS acceleration mode. Images with the 1D, ACS2, and ACS3 acceleration modes showed no significant differences in terms of detecting lesions but scan time can be reduced by 40% (1D vs. ACS3).

Conclusion

ACS acceleration mode can greatly reduce the scan time. In addition, the images have good SNR, high CNR, and strong SLI lesion detection ability.

Denoising Using Noise2Void for Low-Field Magnetic Resonance Imaging: A Phantom Study

To reduce noise for low-field magnetic resonance imaging (MRI) using Noise2Void (N2V) and to demonstrate the N2V validity. N2V is one of the denoising convolutional neural network methods that allows the training of a model without a noiseless clean image. In this study, a kiwi fruit was scanned using a 0.35 Tesla MRI system, and the image qualities at pre- and postdenoising were evaluated. Structural similarity (SSIM), signal-to-noise ratio (SNR), and contrast ratio (CR) were measured, and visual assessment of noise and sharpness was observed. Both SSIM and SNR were significantly improved using N2V (P < 0.05). CR was unchanged between pre- and postdenoising images. The results of visual assessment for noise revealed higher scores in postdenoising images than that in predenoising images. The sharpness scores of postdenoising images were high when SNR was low. N2V provides effective noise reduction and is a useful denoising technique in low-field MRI.

Denoising diffusion weighted imaging data using convolutional neural networks

Diffusion weighted imaging (DWI) with multiple, high b-values is critical for extracting tissue microstructure measurements; however, high b-value DWI images contain high noise levels that can overwhelm the signal of interest and bias microstructural measurements. Here, we propose a simple denoising method that can be applied to any dataset, provided a low-noise, single-subject dataset is acquired using the same DWI sequence. The denoising method uses a one-dimensional convolutional neural network (1D-CNN) and deep learning to learn from a low-noise dataset, voxel-by-voxel. The trained model can then be applied to high-noise datasets from other subjects. We validated the 1D-CNN denoising method by first demonstrating that 1D-CNN denoising resulted in DWI images that were more similar to the noise-free ground truth than comparable denoising methods, e.g., MP-PCA, using simulated DWI data. Using the same DWI acquisition but reconstructed with two common reconstruction methods, i.e. SENSE1 and sum-of-square, to generate a pair of low-noise and high-noise datasets, we then demonstrated that 1D-CNN denoising of high-noise DWI data collected from human subjects showed promising results in three domains: DWI images, diffusion metrics, and tractography. In particular, the denoised images were very similar to a low-noise reference image of that subject, more than the similarity between repeated low-noise images (i.e. computational reproducibility). Finally, we demonstrated the use of the 1D-CNN method in two practical examples to reduce noise from parallel imaging and simultaneous multi-slice acquisition. We conclude that the 1D-CNN denoising method is a simple, effective denoising method for DWI images that overcomes some of the limitations of current state-of-the-art denoising methods, such as the need for a large number of training subjects and the need to account for the rectified noise floor.

Improving Image Quality and Reducing Scan Time for Synthetic MRI of Breast by Using Deep Learning Reconstruction

Objectives. To investigate a deep learning reconstruction algorithm to reduce the time of synthetic MRI (SynMRI) scanning on the breast and improve the image quality. Materials and Methods. A total of 192 healthy female volunteers (mean age: 48.1 years) underwent the breast MR examination at 3.0 T from September 2020 to June 2021. Standard SynMRI and fast SynMRI scans were collected simultaneously on the same volunteer. Deep learning technology with a generative adversarial network (GAN) was used to generate high-quality fast SynMRI images by end-to-end training. Peak signal-to-noise ratio (PSNR), mean squared error (MSE), and structural similarity index measure (SSIM) were used to compare the image quality of generated images from fast SynMRI by deep learning algorithms. Results. Fast SynMRI acquisition time is half of the standard SynMRI scan, and the generated images of the GAN model show that PSNR and SSIM are improved and MSE is reduced. Conclusion. The application of deep learning algorithms with GAN model in breast MAGiC MRI improves the image quality and reduces the scanning time.

Deep learning reconstruction for the evaluation of neuroforaminal stenosis using 1.5T cervical spine MRI: comparison with 3T MRI without deep learning reconstruction

PURPOSE

To compare image quality and interobserver agreement in evaluations of neuroforaminal stenosis between 1.5T cervical spine magnetic resonance imaging (MRI) with deep learning reconstruction (DLR) and 3T MRI without DLR.

METHODS

In this prospective study, 21 volunteers (mean age: 42.4 ± 11.9 years; 17 males) underwent cervical spine T2-weighted sagittal 1.5T and 3T MRI on the same day. The 1.5T and 3T MRI data were used to reconstruct images with (1.5T-DLR) and without (3T-nonDLR) DLR, respectively. Regions of interest were marked on the spinal cord to calculate non-uniformity (NU; standard deviation/signal intensity × 100), as an indicator of image noise. Two blinded radiologists evaluated the images in terms of the depiction of structures, artifacts, noise, overall image quality, and neuroforaminal stenosis. The NU value and the subjective image quality scores were compared between 1.5T-DLR and 3T-nonDLR using the Wilcoxon signed-rank test. Interobserver agreement in evaluations of neuroforaminal stenosis for 1.5T-DLR and 3T-nonDLR was evaluated using Cohen's weighted kappa analysis.

RESULTS

The NU value for 1.5T-DLR was 8.4, which was significantly better than that for 3T-nonDLR (10.3; p < 0.001). Subjective image scores were significantly better for 1.5T-DLR than 3T-nonDLR images (p < 0.037). Interobserver agreement (95% confidence intervals) in the evaluations of neuroforaminal stenosis was significantly superior for 1.5T-DLR (0.920 [0.916-0.924]) than 3T-nonDLR (0.894 [0.889-0.898]).

CONCLUSION

By using DLR, image quality and interobserver agreement in evaluations of neuroforaminal stenosis on 1.5T cervical spine MRI could be improved compared to 3T MRI without DLR.

Validation of a Denoising Method Using Deep Learning-Based Reconstruction to Quantify Multiple Sclerosis Lesion Load on Fast FLAIR Imaging

BACKGROUND AND PURPOSE

Accurate quantification of WM lesion load is essential for the care of patients with multiple sclerosis. We tested whether the combination of accelerated 3D-FLAIR and denoising using deep learning-based reconstruction could provide a relevant strategy while shortening the imaging examination.

MATERIALS AND METHODS

Twenty-eight patients with multiple sclerosis were prospectively examined using 4 implementations of 3D-FLAIR with decreasing scan times (4 minutes 54 seconds, 2 minutes 35 seconds, 1 minute 40 seconds, and 1 minute 15 seconds). Each FLAIR sequence was reconstructed without and with denoising using deep learning-based reconstruction, resulting in 8 FLAIR sequences per patient. Image quality was assessed with the Likert scale, apparent SNR, and contrast-to-noise ratio. Manual and automatic lesion segmentations, performed randomly and blindly, were quantitatively evaluated against ground truth using the absolute volume difference, true-positive rate, positive predictive value, Dice similarity coefficient, Hausdorff distance, and F1 score based on the lesion count. The Wilcoxon signed-rank test and 2-way ANOVA were performed.

RESULTS

Both image-quality evaluation and the various metrics showed deterioration when the FLAIR scan time was accelerated. However, denoising using deep learning-based reconstruction significantly improved subjective image quality and quantitative performance metrics, particularly for manual segmentation. Overall, denoising using deep learning-based reconstruction helped to recover contours closer to those from the criterion standard and to capture individual lesions otherwise overlooked. The Dice similarity coefficient was equivalent between the 2-minutes-35-seconds-long FLAIR with denoising using deep learning-based reconstruction and the 4-minutes-54-seconds-long reference FLAIR sequence.

CONCLUSIONS

Denoising using deep learning-based reconstruction helps to recognize multiple sclerosis lesions buried in the noise of accelerated FLAIR acquisitions, a possibly useful strategy to efficiently shorten the scan time in clinical practice.

Brain Tumor Diagnosis Using Machine Learning, Convolutional Neural Networks, Capsule Neural Networks and Vision Transformers, Applied to MRI: A Survey

Management of brain tumors is based on clinical and radiological information with presumed grade dictating treatment. Hence, a non-invasive assessment of tumor grade is of paramount importance to choose the best treatment plan. Convolutional Neural Networks (CNNs) represent one of the effective Deep Learning (DL)-based techniques that have been used for brain tumor diagnosis. However, they are unable to handle input modifications effectively. Capsule neural networks (CapsNets) are a novel type of machine learning (ML) architecture that was recently developed to address the drawbacks of CNNs. CapsNets are resistant to rotations and affine translations, which is beneficial when processing medical imaging datasets. Moreover, Vision Transformers (ViT)-based solutions have been very recently proposed to address the issue of long-range dependency in CNNs. This survey provides a comprehensive overview of brain tumor classification and segmentation techniques, with a focus on ML-based, CNN-based, CapsNet-based, and ViT-based techniques. The survey highlights the fundamental contributions of recent studies and the performance of state-of-the-art techniques. Moreover, we present an in-depth discussion of crucial issues and open challenges. We also identify some key limitations and promising future research directions. We envisage that this survey shall serve as a good springboard for further study.

Feasibility of accelerated whole-body diffusion-weighted imaging using a deep learning-based noise-reduction technique in patients with prostate cancer

PURPOSE

To assess the possibility of reducing the image acquisition time for diffusion-weighted whole-body imaging with background body signal suppression (DWIBS) by denoising with deep learning-based reconstruction (dDLR).

METHODS

Seventeen patients with prostate cancer who underwent DWIBS by 1.5 T magnetic resonance imaging with a number of excitations of 2 (NEX2) and 8 (NEX8) were prospectively enrolled. The NEX2 image data were processed by dDLR (dDLR-NEX2), and the NEX2, dDLR-NEX2, and NEX8 image data were analyzed. In qualitative analysis, two radiologists rated the perceived coarseness, conspicuity of metastatic lesions (lymph nodes and bone), and overall image quality. The contrast-to-noise ratios (CNRs), contrast ratios, and mean apparent diffusion coefficients (ADCs) of metastatic lesions were calculated in a quantitative analysis.

RESULTS

The image acquisition time of NEX2 was 2.8 times shorter than that of NEX8 (3 min 30 s vs 9 min 48 s). The perceived coarseness and overall image quality scores reported by both readers were significantly higher for dDLR-NEX2 than for NEX2 (P = 0.005-0.040). There was no significant difference between dDLR-NEX2 and NEX8 in the qualitative analysis. The CNR of bone metastasis was significantly greater for dDLR-NEX2 than for NEX2 and NEX8 (P = 0.012 for both comparisons). The contrast ratios and mean ADCs were not significantly different among the three image types.

CONCLUSIONS

dDLR improved the image quality of DWIBS with NEX2. In the context of lymph node and bone metastasis evaluation with DWIBS in patients with prostate cancer, dDLR-NEX2 has potential to be an alternative to NEX8 and reduce the image acquisition time.

Diagnostic advantage of thin slice 2D MRI and multiplanar reconstruction of the knee joint using deep learning based denoising approach

The purpose of this study is to evaluate whether thin-slice high-resolution 2D fat-suppressed proton density-weighted image of the knee joint using denoising approach with deep learning-based reconstruction (dDLR) with MPR is more useful than 3D FS-PD multi planar voxel image. Twelve patients who underwent MRI of the knee at 3T and 13 knees were enrolled. Denoising effect was quantitatively evaluated by comparing the coefficient of variation (CV) before and after dDLR. For the qualitative assessment, two radiologists evaluated image quality, artifacts, anatomical structures, and abnormal findings using a 5-point Likert scale between 2D and 3D. All of them were statistically analyzed. Gwet's agreement coefficients were also calculated. For the scores of abnormal findings, we calculated the percentages of the cases with agreement with high confidence. The CV after dDLR was significantly lower than the one before dDLR (p < 0.05). As for image quality, artifacts and anatomical structure, no significant differences were found except for flow artifact (p < 0.05). The agreement was significantly higher in 2D than in 3D in abnormal findings (p < 0.05). In abnormal findings, the percentage with high confidence was higher in 2D than in 3D (p < 0.05). By applying dDLR to 2D, almost equivalent image quality to 3D could be obtained. Furthermore, abnormal findings could be depicted with greater confidence and consistency, indicating that 2D with dDLR can be a promising imaging method for the knee joint disease evaluation.

Comparison of utility of deep learning reconstruction on 3D MRCPs obtained with three different k-space data acquisitions in patients with IPMN

OBJECTIVE

To compare the utility of deep learning reconstruction (DLR) for improving acquisition time, image quality, and intraductal papillary mucinous neoplasm (IPMN) evaluation for 3D MRCP obtained with parallel imaging (PI), multiple k-space data acquisition for each repetition time (TR) technique (Fast 3D mode multiple: Fast 3Dm) and compressed sensing (CS) with PI.

MATERIALS AND METHODS

A total of 32 IPMN patients who had undergone 3D MRCPs obtained with PI, Fast 3Dm, and CS with PI and reconstructed with and without DLR were retrospectively included in this study. Acquisition time, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) obtained with all protocols were compared using Tukey's HSD test. Results of endoscopic ultrasound, ERCP, surgery, or pathological examination were determined as standard reference, and distribution classifications were compared among all 3D MRCP protocols by McNemar's test.

RESULTS

Acquisition times of Fast 3Dm and CS with PI with and without DLR were significantly shorter than those of PI with and without DLR (p < 0.05). Each MRCP sequence with DLR showed significantly higher SNRs and CNRs than those without DLR (p < 0.05). IPMN distribution accuracy of PI with and without DLR and Fast 3Dm with DLR was significantly higher than that of Fast 3Dm without DLR and CS with PI without DLR (p < 0.05).

CONCLUSION

DLR is useful for improving image quality and IPMN evaluation capability on 3D MRCP obtained with PI, Fast 3Dm, or CS with PI. Moreover, Fast 3Dm and CS with PI may play as substitution to PI for MRCP in patients with IPMN.

KEY POINTS

• Mean examination times of multiple k-space data acquisitions for each TR and compressed sensing with parallel imaging were significantly shorter than that of parallel imaging (p < 0.0001). • When comparing image quality of 3D MRCPs with and without deep learning reconstruction, deep learning reconstruction significantly improved signal-to-noise ratio and contrast-to-noise ratio (p < 0.05). • IPMN distribution accuracies of parallel imaging with and without deep learning reconstruction (with vs. without: 88.0% vs. 88.0%) and multiple k-space data acquisitions for each TR with deep learning reconstruction (86.0%) were significantly higher than those of others (p < 0.05).

Magnetic Resonance Imaging Evaluation of Hemangioma Resection for Encephalofacial Angiomatosis (Sturge-Weber Syndrome) in Children under Intelligent Algorithm

This study was aimed to evaluate the clinical efficacy of hemangioma resection in the treatment of infantile encephalofacial angiomatosis (Sturge–Weber syndrome, SWS) through magnetic resonance imaging (MRI) images, and intelligent algorithms were employed to process MRI images. A retrospective study of 45 children diagnosed with facial hemangioma admitted to hospital was conducted. Then, MRS images were acquired, and a mathematical model for MRI image denoising and reconstruction was constructed based on nonlocal similar block low-rank prior algorithms. The processing effect was assessed regarding the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Finally, MRI images were collected to analyze the difference between the metabolites of N-acetylaspartic acid (NAA), creatine (Cr), choline (Cho), and their ratios in the lesions of the children before and after treatment. The improvement rate was analyzed through a twelve-month follow-up. The algorithm test results showed that compared with the classic K-singular value decomposition (K-SVD) denoising algorithm and the Sparse MRI reconstruction algorithm, the proposed algorithm processed MRI images more clearly and had more detailed information. The quantitative results showed that the PSNR and SSIM in the image processed by the algorithm proposed were remarkably large. The clinical treatment results showed that compared with those before treatment, the nCho level after treatment, the ratio of Cho/Cr and Cho/NAA were remarkably reduced, and the difference was remarkable (P < 0.05). The follow-up results showed that the considerable improvement rate was 88.89%, the postoperative organ remodeling rate was 17.78%, and the probability of reoperation was only 6.67%. In summary, the introduction of intelligent algorithms for denoising and reconstruction of MRI images can remarkably improve image quality and help doctors use image information to diagnose diseases and evaluate treatment effects. The hemangioma resection for the treatment of pediatric SWS had a high treatment improvement rate and was worthy of clinical adoption.

Evaluation of deep learning reconstructed high-resolution 3D lumbar spine MRI

OBJECTIVES

To compare interobserver agreement and image quality of 3D T2-weighted fast spin echo (T2w-FSE) L-spine MRI images processed with a deep learning reconstruction (DLRecon) against standard-of-care (SOC) reconstruction, as well as against 2D T2w-FSE images. The hypothesis was that DLRecon 3D T2w-FSE would afford improved image quality and similar interobserver agreement compared to both SOC 3D and 2D T2w-FSE.

METHODS

Under IRB approval, patients who underwent routine 3-T lumbar spine (L-spine) MRI from August 17 to September 17, 2020, with both isotropic 3D and 2D T2w-FSE sequences, were retrospectively included. A DLRecon algorithm, with denoising and sharpening properties was applied to SOC 3D k-space to generate 3D DLRecon images. Four musculoskeletal radiologists blinded to reconstruction status evaluated randomized images for motion artifact, image quality, central/foraminal stenosis, disc degeneration, annular fissure, disc herniation, and presence of facet joint cysts. Inter-rater agreement for each graded variable was evaluated using Conger's kappa (κ).

RESULTS

Thirty-five patients (mean age 58 ± 19, 26 female) were evaluated. 3D DLRecon demonstrated statistically significant higher median image quality score (2.0/2) when compared to SOC 3D (1.0/2, p < 0.001), 2D axial (1.0/2, p < 0.001), and 2D sagittal sequences (1.0/2, p value < 0.001). κ ranges (and 95% CI) for foraminal stenosis were 0.55-0.76 (0.32-0.86) for 3D DLRecon, 0.56-0.73 (0.35-0.84) for SOC 3D, and 0.58-0.71 (0.33-0.84) for 2D. Mean κ (and 95% CI) for central stenosis at L4-5 were 0.98 (0.96-0.99), 0.97 (0.95-0.99), and 0.98 (0.96-0.99) for 3D DLRecon, 3D SOC and 2D, respectively.

CONCLUSIONS

DLRecon 3D T2w-FSE L-spine MRI demonstrated higher image quality and similar interobserver agreement for graded variables of interest when compared to 3D SOC and 2D imaging.

KEY POINTS

• 3D DLRecon T2w-FSE isotropic lumbar spine MRI provides improved image quality when compared to 2D MRI, with similar interobserver agreement for clinical evaluation of pathology. • 3D DLRecon images demonstrated better image quality score (2.0/2) when compared to standard-of-care (SOC) 3D (1.0/2), p value < 0.001; 2D axial (1.0/2), p value < 0.001; and 2D sagittal sequences (1.0/2), p value < 0.001. • Interobserver agreement for major variables of interest was similar among all sequences and reconstruction types. For foraminal stenosis, κ ranged from 0.55 to 0.76 (95% CI 0.32-0.86) for 3D DLRecon, 0.56-0.73 (95% CI 0.35-0.84) for standard-of-care (SOC) 3D, and 0.58-0.71 (95% CI 0.33-0.84) for 2D.

Highly accelerated 3D MPRAGE using deep neural network-based reconstruction for brain imaging in children and young adults

OBJECTIVES

This study aimed to accelerate the 3D magnetization-prepared rapid gradient-echo (MPRAGE) sequence for brain imaging through the deep neural network (DNN).

METHODS

This retrospective study used the k-space data of 240 scans (160 for the training set, mean ± standard deviation age, 93 ± 80 months, 94 males; 80 for the test set, 106 ± 83 months, 44 males) of conventional MPRAGE (C-MPRAGE) and 102 scans (77 ± 74 months, 52 males) of both C-MPRAGE and accelerated MPRAGE. All scans were acquired with 3T scanners. DNN was developed with simulated-acceleration data generated by under-sampling. Quantitative error metrics were compared between images reconstructed with DNN, GRAPPA, and E-SPIRIT using the paired t-test. Qualitative image quality was compared between C-MPRAGE and accelerated MPRAGE reconstructed with DNN (DNN-MPRAGE) by two readers. Lesions were segmented and the agreement between C-MPRAGE and DNN-MPRAGE was assessed using linear regression.

RESULTS

Accelerated MPRAGE reduced scan times by 38% compared to C-MPRAGE (142 s vs. 320 s). For quantitative error metrics, DNN showed better performance than GRAPPA and E-SPIRIT (p < 0.001). For qualitative evaluation, overall image quality of DNN-MPRAGE was comparable (p > 0.999) or better (p = 0.025) than C-MPRAGE, depending on the reader. Pixelation was reduced in DNN-MPRAGE (p < 0.001). Other qualitative parameters were comparable (p > 0.05). Lesions in C-MPRAGE and DNN-MPRAGE showed good agreement for the dice similarity coefficient (= 0.68) and linear regression (R² = 0.97; p < 0.001).

CONCLUSIONS

DNN-MPRAGE reduced acquisition time by 38% and revealed comparable image quality to C-MPRAGE.

KEY POINTS

• DNN-MPRAGE reduced acquisition times by 38%. • DNN-MPRAGE outperformed conventional reconstruction on accelerated scans (SSIM of DNN-MPRAGE = 0.96, GRAPPA = 0.43, E-SPIRIT = 0.88; p < 0.001). • Compared to C-MPRAGE scans, DNN-MPRAGE showed improved mean scores for overall image quality (2.46 vs. 2.52; p < 0.001) or comparable perceived SNR (2.56 vs. 2.58; p = 0.08).