Prospective Deployment of Deep Learning in MRI: A Framework for Important Considerations, Challenges, and Recommendations for Best Practices

Deep Learning Reconstruction of Prospectively Accelerated MRI of the Pancreas: Clinical Evaluation of Shortened Breath-Hold Examinations With Dixon Fat Suppression

OBJECTIVE

Deep learning (DL)-enabled magnetic resonance imaging (MRI) reconstructions can enable shortening of breath-hold examinations and improve image quality by reducing motion artifacts. Prospective studies with DL reconstructions of accelerated MRI of the upper abdomen in the context of pancreatic pathologies are lacking. In a clinical setting, the purpose of this study is to investigate the performance of a novel DL-based reconstruction algorithm in T1-weighted volumetric interpolated breath-hold examinations with partial Fourier sampling and Dixon fat suppression (hereafter, VIBE-Dixon DL ). The objective is to analyze its impact on acquisition time, image sharpness and quality, diagnostic confidence, pancreatic lesion conspicuity, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR).

METHODS

This prospective single-center study included participants with various pancreatic pathologies who gave written consent from January 2023 to September 2023. During the same session, each participant underwent 2 MRI acquisitions using a 1.5 T scanner: conventional precontrast and postcontrast T1-weighted VIBE acquisitions with Dixon fat suppression (VIBE-Dixon, reference standard) using 4-fold parallel imaging acceleration and 6-fold accelerated VIBE-Dixon acquisitions with partial Fourier sampling utilizing a novel DL reconstruction tailored to the acquisition. A qualitative image analysis was performed by 4 readers. Acquisition time, image sharpness, overall image quality, image noise and artifacts, diagnostic confidence, as well as pancreatic lesion conspicuity and size were compared. Furthermore, a quantitative analysis of SNR and CNR was performed.

RESULTS

Thirty-two participants were evaluated (mean age ± SD, 62 ± 19 years; 20 men). The VIBE-Dixon DL method enabled up to 52% reduction in average breath-hold time (7 seconds for VIBE-Dixon DL vs 15 seconds for VIBE-Dixon, P < 0.001). A significant improvement of image sharpness, overall image quality, diagnostic confidence, and pancreatic lesion conspicuity was observed in the images recorded using VIBE-Dixon DL ( P < 0.001). Furthermore, a significant reduction of image noise and motion artifacts was noted in the images recorded using the VIBE-Dixon DL technique ( P < 0.001). In addition, for all readers, there was no evidence of a difference in lesion size measurement between VIBE-Dixon and VIBE-Dixon DL . Interreader agreement between VIBE-Dixon and VIBE-Dixon DL regarding lesion size was excellent (intraclass correlation coefficient, >90). Finally, a statistically significant increase of pancreatic SNR in VIBE-DIXON DL was observed in both the precontrast ( P = 0.025) and postcontrast images ( P < 0.001). Also, an increase of splenic SNR in VIBE-DIXON DL was observed in both the precontrast and postcontrast images, but only reaching statistical significance in the postcontrast images ( P = 0.34 and P = 0.003, respectively). Similarly, an increase of pancreas CNR in VIBE-DIXON DL was observed in both the precontrast and postcontrast images, but only reaching statistical significance in the postcontrast images ( P = 0.557 and P = 0.026, respectively).

CONCLUSIONS

The prospectively accelerated, DL-enhanced VIBE with Dixon fat suppression was clinically feasible. It enabled a 52% reduction in breath-hold time and provided superior image quality, diagnostic confidence, and pancreatic lesion conspicuity. This technique might be especially useful for patients with limited breath-hold capacity.

Prospective Deployment of Deep Learning Reconstruction Facilitates Highly Accelerated Upper Abdominal MRI

RATIONALE AND OBJECTIVES

To compare a conventional T1 volumetric interpolated breath-hold examination (VIBE) with SPectral Attenuated Inversion Recovery (SPAIR) fat saturation and a deep learning (DL)-reconstructed accelerated VIBE sequence with SPAIR fat saturation achieving a 50 % reduction in breath-hold duration (hereafter, VIBE-SPAIRDL) in terms of image quality and diagnostic confidence.

MATERIALS AND METHODS

This prospective study enrolled consecutive patients referred for upper abdominal MRI from November 2023 to December 2023 at a single tertiary center. Patients underwent upper abdominal MRI with acquisition of non-contrast and gadobutrol-enhanced conventional VIBE-SPAIR (fourfold acceleration, acquisition time 16 s) and VIBE-SPAIRDL (sixfold acceleration, acquisition time 8 s) on a 1.5 T scanner. Image analysis was performed by four readers, evaluating homogeneity of fat suppression, perceived signal-to-noise ratio (SNR), edge sharpness, artifact level, lesion detectability and diagnostic confidence. A statistical power analysis for patient sample size estimation was performed. Image quality parameters were compared by a repeated measures analysis of variance, and interreader agreement was assessed using Fleiss' κ.

RESULTS

Among 450 consecutive patients, 45 patients were evaluated (mean age, 60 years ± 15 [SD]; 27 men, 18 women). VIBE-SPAIRDL acquisition demonstrated superior SNR (P < 0.001), edge sharpness (P < 0.001), and reduced artifacts (P < 0.001) with substantial to almost perfect interreader agreement for non-contrast (κ: 0.70-0.91) and gadobutrol-enhanced MRI (κ: 0.68-0.87). No evidence of a difference was found between conventional VIBE-SPAIR and VIBE-SPAIRDL regarding homogeneity of fat suppression, lesion detectability, or diagnostic confidence (all P > 0.05).

CONCLUSION

Deep learning reconstruction of VIBE-SPAIR facilitated a reduction of breath-hold duration by half, while reducing artifacts and improving image quality.

SUMMARY

Deep learning reconstruction of prospectively accelerated T1 volumetric interpolated breath-hold examination for upper abdominal MRI enabled a 50 % reduction in breath-hold time with superior image quality.

KEY RESULTS

1) In a prospective analysis of 45 patients referred for upper abdominal MRI, accelerated deep learning (DL)-reconstructed VIBE images with spectral fat saturation (SPAIR) showed better overall image quality, with better perceived signal-to-noise ratio and less artifacts (all P < 0.001), despite a 50 % reduction in acquisition time compared to conventional VIBE. 2) No evidence of a difference was found between conventional VIBE-SPAIR and accelerated VIBE-SPAIRDL regarding lesion detectability or diagnostic confidence.

Improved Resolution and Image Quality of Musculoskeletal Magnetic Resonance Imaging using Deep Learning-based Denoising Reconstruction: A Prospective Clinical Study

OBJECTIVE

To prospectively evaluate a deep learning-based denoising reconstruction (DLR) for improved resolution and image quality in musculoskeletal (MSK) magnetic resonance imaging (MRI).

METHODS

Images from 137 contrast-weighted sequences in 40 MSK patients were evaluated. Each sequence was performed twice, first with the routine parameters and reconstructed with a routine reconstruction filter (REF), then with higher resolution and reconstructed with DLR, and with three conventional reconstruction filters (NL2, GA43, GA53). The five reconstructions (REF, DLR, NL2, GA43, and GA53) were de-identified, randomized, and blindly reviewed by three MSK radiologists using eight scoring criteria and a forced ranking. Quantitative SNR, CNR, and structure's full width at half maximum (FWHM) for resolution assessment were measured and compared. To account for repeated measures, Generalized Estimating Equations (GEE) with Bonferroni adjustment was used to compare the reader's scores, SNR, CNR, and FWHM between DLR vs. NL2, GA43, GA53, and REF.

RESULTS

Compared to the routine REF images, the resolution was improved by 47.61% with DLR from 0.39 ± 0.15 mm2 to 0.20 ± 0.06 mm2 (p < 0.001). Per-sequence average scan time was shortened by 7.93% with DLR from 165.58 ± 21.86 s to 152.45 ± 25.65 s (p < 0.001). Based on the average scores, DLR images were rated significantly higher in all image quality criteria and the forced ranking (p < 0.001).

CONCLUSION

This prospective clinical evaluation demonstrated that DLR allows approximately two times finer resolution and improved image quality compared to the standard-of-care images.

Accelerated High-Resolution Deep Learning Reconstruction Turbo Spin Echo MRI of the Knee at 7 T

OBJECTIVES

The aim of this study was to compare the image quality of 7 T turbo spin echo (TSE) knee images acquired with varying factors of parallel-imaging acceleration reconstructed with deep learning (DL)-based and conventional algorithms.

MATERIALS AND METHODS

This was a prospective single-center study. Twenty-three healthy volunteers underwent 7 T knee magnetic resonance imaging. Two-, 3-, and 4-fold accelerated high-resolution fat-signal-suppressing proton density (PD-fs) and T1-weighted coronal 2D TSE acquisitions with an encoded voxel volume of 0.31 × 0.31 × 1.5 mm 3 were acquired. Each set of raw data was reconstructed with a DL-based and a conventional Generalized Autocalibrating Partially Parallel Acquisition (GRAPPA) algorithm. Three readers rated image contrast, sharpness, artifacts, noise, and overall quality. Friedman analysis of variance and the Wilcoxon signed rank test were used for comparison of image quality criteria.

RESULTS

The mean age of the participants was 32.0 ± 8.1 years (15 male, 8 female). Acquisition times at 4-fold acceleration were 4 minutes 15 seconds (PD-fs, Supplemental Video is available at http://links.lww.com/RLI/A938 ) and 3 minutes 9 seconds (T1, Supplemental Video available at http://links.lww.com/RLI/A939 ). At 4-fold acceleration, image contrast, sharpness, noise, and overall quality of images reconstructed with the DL-based algorithm were significantly better rated than the corresponding GRAPPA reconstructions ( P < 0.001). Four-fold accelerated DL-reconstructed images scored significantly better than 2- to 3-fold GRAPPA-reconstructed images with regards to image contrast, sharpness, noise, and overall quality ( P ≤ 0.031). Image contrast of PD-fs images at 2-fold acceleration ( P = 0.087), image noise of T1-weighted images at 2-fold acceleration ( P = 0.180), and image artifacts for both sequences at 2- and 3-fold acceleration ( P ≥ 0.102) of GRAPPA reconstructions were not rated differently than those of 4-fold accelerated DL-reconstructed images. Furthermore, no significant difference was observed for all image quality measures among 2-fold, 3-fold, and 4-fold accelerated DL reconstructions ( P ≥ 0.082).

CONCLUSIONS

This study explored the technical potential of DL-based image reconstruction in accelerated 2D TSE acquisitions of the knee at 7 T. DL reconstruction significantly improved a variety of image quality measures of high-resolution TSE images acquired with a 4-fold parallel-imaging acceleration compared with a conventional reconstruction algorithm.

Accelerated spine MRI with deep learning based image reconstruction: a prospective comparison with standard MRI

RATIONALE AND OBJECTIVES

To evaluate the performance of deep learning (DL) reconstructed MRI in terms of image acquisition time, overall image quality and diagnostic interchangeability compared to standard-of-care (SOC) MRI.

MATERIALS AND METHODS

This prospective study recruited participants between July 2023 and August 2023 who had spinal discomfort. All participants underwent two separate MRI examinations (Standard and accelerated scanning). Signal-to-noise ratios (SNR), contrast-to-noise ratios (CNR) and similarity metrics were calculated for quantitative evaluation. Four radiologists performed subjective quality and lesion characteristic assessment. Wilcoxon test was used to assess the differences of SNR, CNR and subjective image quality between DL and SOC. Various lesions of spine were also tested for interchangeability using individual equivalence index. Interreader and intrareader agreement and concordance (κ and Kendall τ and W statistics) were computed and McNemar tests were performed for comprehensive evaluation.

RESULTS

200 participants (107 male patients, mean age 46.56 ± 17.07 years) were included. Compared with SOC, DL enabled scan time reduced by approximately 40%. The SNR and CNR of DL were significantly higher than those of SOC (P < 0.001). DL showed varying degrees of improvement (0-0.35) in each of similarity metrics. All absolute individual equivalence indexes were less than 4%, indicating interchangeability between SOC and DL. Kappa and Kendall showed a good to near-perfect agreement in range of 0.72-0.98. There is no difference between SOC and DL regarding subjective scoring and frequency of lesion detection.

CONCLUSION

Compared to SOC, DL provided high-quality image for diagnosis and reduced examination time for patients. DL was found to be interchangeable with SOC in detecting various spinal abnormalities.

Accelerated Musculoskeletal Magnetic Resonance Imaging

With a substantial growth in the use of musculoskeletal MRI, there has been a growing need to improve MRI workflow, and faster imaging has been suggested as one of the solutions for a more efficient examination process. Consequently, there have been considerable advances in accelerated MRI scanning methods. This article aims to review the basic principles and applications of accelerated musculoskeletal MRI techniques including widely used conventional acceleration methods, more advanced deep learning-based techniques, and new approaches to reduce scan time. Specifically, conventional accelerated MRI techniques, including parallel imaging, compressed sensing, and simultaneous multislice imaging, and deep learning-based accelerated MRI techniques, including undersampled MR image reconstruction, super-resolution imaging, artifact correction, and generation of unacquired contrast images, are discussed. Finally, new approaches to reduce scan time, including synthetic MRI, novel sequences, and new coil setups and designs, are also reviewed. We believe that a deep understanding of these fast MRI techniques and proper use of combined acceleration methods will synergistically improve scan time and MRI workflow in daily practice. EVIDENCE LEVEL: 3 TECHNICAL EFFICACY: Stage 1.

Comparison of image quality and lesion conspicuity between conventional and deep learning reconstruction in gadoxetic acid-enhanced liver MRI

OBJECTIVE

To compare the image quality and lesion conspicuity of conventional vs deep learning (DL)-based reconstructed three-dimensional T1-weighted images in gadoxetic acid-enhanced liver magnetic resonance imaging (MRI).

METHODS

This prospective study (NCT05182099) enrolled participants scheduled for gadoxetic acid-enhanced liver MRI due to suspected focal liver lesions (FLLs) who provided signed informed consent. A liver MRI was conducted using a 3-T scanner. T1-weighted images were reconstructed using both conventional and DL-based (AIRTM Recon DL 3D) reconstruction algorithms. Three radiologists independently reviewed the image quality and lesion conspicuity on a 5-point scale.

RESULTS

Fifty participants (male = 36, mean age 62 ± 11 years) were included for image analysis. The DL-based reconstruction showed significantly higher image quality than conventional images in all phases (3.71-4.40 vs 3.37-3.99, p < 0.001 for all), as well as significantly less noise and ringing artifacts than conventional images (p < 0.05 for all), while also showing significantly altered image texture (p < 0.001 for all). Lesion conspicuity was significantly higher in DL-reconstructed images than in conventional images in the arterial phase (2.15 [95% confidence interval: 1.78, 2.52] vs 2.03 [1.65, 2.40], p = 0.036), but no significant difference was observed in the portal venous phase and hepatobiliary phase (p > 0.05 for all). There was no significant difference in the figure-of-merit (0.728 in DL vs 0.709 in conventional image, p = 0.474).

CONCLUSION

DL reconstruction provided higher-quality three-dimensional T1-weighted imaging than conventional reconstruction in gadoxetic acid-enhanced liver MRI.

CRITICAL RELEVANCE STATEMENT

DL reconstruction of 3D T1-weighted images improves image quality and arterial phase lesion conspicuity in gadoxetic acid-enhanced liver MRI compared to conventional reconstruction.

KEY POINTS

DL reconstruction is feasible for 3D T1-weighted images across different spatial resolutions and phases. DL reconstruction showed superior image quality with reduced noise and ringing artifacts. Hepatic anatomic structures were more conspicuous on DL-reconstructed images.

Intraoperative near infrared functional imaging of rectal cancer using artificial intelligence methods - now and near future state of the art

Colorectal cancer remains a major cause of cancer death and morbidity worldwide. Surgery is a major treatment modality for primary and, increasingly, secondary curative therapy. However, with more patients being diagnosed with early stage and premalignant disease manifesting as large polyps, greater accuracy in diagnostic and therapeutic precision is needed right from the time of first endoscopic encounter. Rapid advancements in the field of artificial intelligence (AI), coupled with widespread availability of near infrared imaging (currently based around indocyanine green (ICG)) can enable colonoscopic tissue classification and prognostic stratification for significant polyps, in a similar manner to contemporary dynamic radiological perfusion imaging but with the advantage of being able to do so directly within interventional procedural time frames. It can provide an explainable method for immediate digital biopsies that could guide or even replace traditional forceps biopsies and provide guidance re margins (both areas where current practice is only approximately 80% accurate prior to definitive excision). Here, we discuss the concept and practice of AI enhanced ICG perfusion analysis for rectal cancer surgery while highlighting recent and essential near-future advancements. These include breakthrough developments in computer vision and time series analysis that allow for real-time quantification and classification of fluorescent perfusion signals of rectal cancer tissue intraoperatively that accurately distinguish between normal, benign, and malignant tissues in situ endoscopically, which are now undergoing international prospective validation (the Horizon Europe CLASSICA study). Next stage advancements may include detailed digital characterisation of small rectal malignancy based on intraoperative assessment of specific intratumoral fluorescent signal pattern. This could include T staging and intratumoral molecular process profiling (e.g. regarding angiogenesis, differentiation, inflammatory component, and tumour to stroma ratio) with the potential to accurately predict the microscopic local response to nonsurgical treatment enabling personalised therapy via decision support tools. Such advancements are also applicable to the next generation fluorophores and imaging agents currently emerging from clinical trials. In addition, by providing an understandable, applicable method for detailed tissue characterisation visually, such technology paves the way for acceptance of other AI methodology during surgery including, potentially, deep learning methods based on whole screen/video detailing.

Deep learning for accelerated and robust MRI reconstruction

Deep learning (DL) has recently emerged as a pivotal technology for enhancing magnetic resonance imaging (MRI), a critical tool in diagnostic radiology. This review paper provides a comprehensive overview of recent advances in DL for MRI reconstruction, and focuses on various DL approaches and architectures designed to improve image quality, accelerate scans, and address data-related challenges. It explores end-to-end neural networks, pre-trained and generative models, and self-supervised methods, and highlights their contributions to overcoming traditional MRI limitations. It also discusses the role of DL in optimizing acquisition protocols, enhancing robustness against distribution shifts, and tackling biases. Drawing on the extensive literature and practical insights, it outlines current successes, limitations, and future directions for leveraging DL in MRI reconstruction, while emphasizing the potential of DL to significantly impact clinical imaging practices.

Improvement of image quality in diffusion-weighted imaging with model-based deep learning reconstruction for evaluations of the head and neck

OBJECTIVES

To investigate the utility of deep learning (DL)-based image reconstruction using a model-based approach in head and neck diffusion-weighted imaging (DWI).

MATERIALS AND METHODS

We retrospectively analyzed the cases of 41 patients who underwent head/neck DWI. The DWI in 25 patients demonstrated an untreated lesion. We performed qualitative and quantitative assessments in the DWI analyses with both deep learning (DL)- and conventional parallel imaging (PI)-based reconstructions. For the qualitative assessment, we visually evaluated the overall image quality, soft tissue conspicuity, degree of artifact(s), and lesion conspicuity based on a five-point system. In the quantitative assessment, we measured the signal-to-noise ratio (SNR) of the bilateral parotid glands, submandibular gland, the posterior muscle, and the lesion. We then calculated the contrast-to-noise ratio (CNR) between the lesion and the adjacent muscle.

RESULTS

Significant differences were observed in the qualitative analysis between the DWI with PI-based and DL-based reconstructions for all of the evaluation items (p < 0.001). In the quantitative analysis, significant differences in the SNR and CNR between the DWI with PI-based and DL-based reconstructions were observed for all of the evaluation items (p = 0.002 ~ p < 0.001).

DISCUSSION

DL-based image reconstruction with the model-based technique effectively provided sufficient image quality in head/neck DWI.

Speeding Up and Improving Image Quality in Glioblastoma MRI Protocol by Deep Learning Image Reconstruction

A fully diagnostic MRI glioma protocol is key to monitoring therapy assessment but is time-consuming and especially challenging in critically ill and uncooperative patients. Artificial intelligence demonstrated promise in reducing scan time and improving image quality simultaneously. The purpose of this study was to investigate the diagnostic performance, the impact on acquisition acceleration, and the image quality of a deep learning optimized glioma protocol of the brain. Thirty-three patients with histologically confirmed glioblastoma underwent standardized brain tumor imaging according to the glioma consensus recommendations on a 3-Tesla MRI scanner. Conventional and deep learning-reconstructed (DLR) fluid-attenuated inversion recovery, and T2- and T1-weighted contrast-enhanced Turbo spin echo images with an improved in-plane resolution, i.e., super-resolution, were acquired. Two experienced neuroradiologists independently evaluated the image datasets for subjective image quality, diagnostic confidence, tumor conspicuity, noise levels, artifacts, and sharpness. In addition, the tumor volume was measured in the image datasets according to Response Assessment in Neuro-Oncology (RANO) 2.0, as well as compared between both imaging techniques, and various clinical-pathological parameters were determined. The average time saving of DLR sequences was 30% per MRI sequence. Simultaneously, DLR sequences showed superior overall image quality (all p < 0.001), improved tumor conspicuity and image sharpness (all p < 0.001, respectively), and less image noise (all p < 0.001), while maintaining diagnostic confidence (all p > 0.05), compared to conventional images. Regarding RANO 2.0, the volume of non-enhancing non-target lesions (p = 0.963), enhancing target lesions (p = 0.993), and enhancing non-target lesions (p = 0.951) did not differ between reconstruction types. The feasibility of the deep learning-optimized glioma protocol was demonstrated with a 30% reduction in acquisition time on average and an increased in-plane resolution. The evaluated DLR sequences improved subjective image quality and maintained diagnostic accuracy in tumor detection and tumor classification according to RANO 2.0.

Scalable Swin Transformer network for brain tumor segmentation from incomplete MRI modalities

BACKGROUND

Deep learning methods have shown great potential in processing multi-modal Magnetic Resonance Imaging (MRI) data, enabling improved accuracy in brain tumor segmentation. However, the performance of these methods can suffer when dealing with incomplete modalities, which is a common issue in clinical practice. Existing solutions, such as missing modality synthesis, knowledge distillation, and architecture-based methods, suffer from drawbacks such as long training times, high model complexity, and poor scalability.

METHOD

This paper proposes IMS2Trans, a novel lightweight scalable Swin Transformer network by utilizing a single encoder to extract latent feature maps from all available modalities. This unified feature extraction process enables efficient information sharing and fusion among the modalities, resulting in efficiency without compromising segmentation performance even in the presence of missing modalities.

RESULTS

Two datasets, BraTS 2018 and BraTS 2020, containing incomplete modalities for brain tumor segmentation are evaluated against popular benchmarks. On the BraTS 2018 dataset, our model achieved higher average Dice similarity coefficient (DSC) scores for the whole tumor, tumor core, and enhancing tumor regions (86.57, 75.67, and 58.28, respectively), in comparison with a state-of-the-art model, i.e. mmFormer (86.45, 75.51, and 57.79, respectively). Similarly, on the BraTS 2020 dataset, our model scored higher DSC scores in these three brain tumor regions (87.33, 79.09, and 62.11, respectively) compared to mmFormer (86.17, 78.34, and 60.36, respectively). We also conducted a Wilcoxon test on the experimental results, and the generated p-value confirmed that our model's performance was statistically significant. Moreover, our model exhibits significantly reduced complexity with only 4.47 M parameters, 121.89G FLOPs, and a model size of 77.13 MB, whereas mmFormer comprises 34.96 M parameters, 265.79 G FLOPs, and a model size of 559.74 MB. These indicate our model, being light-weighted with significantly reduced parameters, is still able to achieve better performance than a state-of-the-art model.

CONCLUSION

By leveraging a single encoder for processing the available modalities, IMS2Trans offers notable scalability advantages over methods that rely on multiple encoders. This streamlined approach eliminates the need for maintaining separate encoders for each modality, resulting in a lightweight and scalable network architecture. The source code of IMS2Trans and the associated weights are both publicly available at https://github.com/hudscomdz/IMS2Trans.

Deep learning-accelerated image reconstruction in back pain-MRI imaging: reduction of acquisition time and improvement of image quality

INTRODUCTION

Low back pain is a global health issue causing disability and missed work days. Commonly used MRI scans including T1-weighted and T2-weighted images provide detailed information of the spine and surrounding tissues. Artificial intelligence showed promise in improving image quality and simultaneously reducing scan time. This study evaluates the performance of deep learning (DL)-based T2 turbo spin-echo (TSE, T2DLR) and T1 TSE (T1DLR) in lumbar spine imaging regarding acquisition time, image quality, artifact resistance, and diagnostic confidence.

MATERIAL AND METHODS

This retrospective monocentric study included 60 patients with lower back pain who underwent lumbar spinal MRI between February and April 2023. MRI parameters and DL reconstruction (DLR) techniques were utilized to acquire images. Two neuroradiologists independently evaluated image datasets based on various parameters using a 4-point Likert scale.

RESULTS

Accelerated imaging showed significantly less image noise and artifacts, as well as better image sharpness, compared to standard imaging. Overall image quality and diagnostic confidence were higher in accelerated imaging. Relevant disk herniations and spinal fractures were detected in both DLR and conventional images. Both readers favored accelerated imaging in the majority of examinations. The lumbar spine examination time was cut by 61% in accelerated imaging compared to standard imaging.

CONCLUSION

In conclusion, the utilization of deep learning-based image reconstruction techniques in lumbar spinal imaging resulted in significant time savings of up to 61% compared to standard imaging, while also improving image quality and diagnostic confidence. These findings highlight the potential of these techniques to enhance efficiency and accuracy in clinical practice for patients with lower back pain.

Deep learning-accelerated image reconstruction in MRI of the orbit to shorten acquisition time and enhance image quality

BACKGROUND AND PURPOSE

This study explores the use of deep learning (DL) techniques in MRI of the orbit to enhance imaging. Standard protocols, although detailed, have lengthy acquisition times. We investigate DL-based methods for T2-weighted and T1-weighted, fat-saturated, contrast-enhanced turbo spin echo (TSE) sequences, aiming to improve image quality, reduce acquisition time, minimize artifacts, and enhance diagnostic confidence in orbital imaging.

METHODS

In a 3-Tesla MRI study of 50 patients evaluating orbital diseases from March to July 2023, conventional (TSES ) and DL TSE sequences (TSEDL ) were used. Two neuroradiologists independently assessed the image datasets for image quality, diagnostic confidence, noise levels, artifacts, and image sharpness using a randomized and blinded 4-point Likert scale.

RESULTS

TSEDL significantly reduced image noise and artifacts, enhanced image sharpness, and decreased scan time, outperforming TSES (p < .05). TSEDL showed superior overall image quality and diagnostic confidence, with relevant findings effectively detected in both DL-based and conventional images. In 94% of cases, readers preferred accelerated imaging.

CONCLUSION

The study proved that using DL for MRI image reconstruction in orbital scans significantly cut acquisition time by 69%. This approach also enhanced image quality, reduced image noise, sharpened images, and boosted diagnostic confidence.

Ultrafast Brain MRI with Deep Learning Reconstruction for Suspected Acute Ischemic Stroke

Background Deep learning (DL)-accelerated MRI can substantially reduce examination times. However, studies prospectively evaluating the diagnostic performance of DL-accelerated MRI reconstructions in acute suspected stroke are lacking. Purpose To investigate the interchangeability of DL-accelerated MRI with conventional MRI in patients with suspected acute ischemic stroke at 1.5 T. Materials and Methods In this prospective study, 211 participants with suspected acute stroke underwent clinically indicated MRI at 1.5 T between June 2022 and March 2023. For each participant, conventional MRI (including T1-weighted, T2-weighted, T2*-weighted, T2 fluid-attenuated inversion-recovery, and diffusion-weighted imaging; 14 minutes 18 seconds) and DL-accelerated MRI (same sequences; 3 minutes 4 seconds) were performed. The primary end point was the interchangeability between conventional and DL-accelerated MRI for acute ischemic infarction detection. Secondary end points were interchangeability regarding the affected vascular territory and clinically relevant secondary findings (eg, microbleeds, neoplasm). Three readers evaluated the overall occurrence of acute ischemic stroke, affected vascular territory, clinically relevant secondary findings, overall image quality, and diagnostic confidence. For acute ischemic lesions, size and signal intensities were assessed. The margin for interchangeability was chosen as 5%. For interrater agreement analysis and interrater reliability analysis, multirater Fleiss κ and the intraclass correlation coefficient, respectively, was determined. Results The study sample consisted of 211 participants (mean age, 65 years ± 16 [SD]); 123 male and 88 female). Acute ischemic stroke was confirmed in 79 participants. Interchangeability was demonstrated for all primary and secondary end points. No individual equivalence indexes (IEIs) exceeded the interchangeability margin of 5% (IEI, -0.002 [90% CI: -0.007, 0.004]). Almost perfect interrater agreement was observed (P > .91). DL-accelerated MRI provided higher overall image quality (P < .001) and diagnostic confidence (P < .001). The signal properties of acute ischemic infarctions were similar in both techniques and demonstrated good to excellent interrater reliability (intraclass correlation coefficient, ≥0.8). Conclusion Despite being four times faster, DL-accelerated brain MRI was interchangeable with conventional MRI for acute ischemic lesion detection. © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Haller in this issue.

Developing and deploying deep learning models in brain magnetic resonance imaging: A review

Magnetic resonance imaging (MRI) of the brain has benefited from deep learning (DL) to alleviate the burden on radiologists and MR technologists, and improve throughput. The easy accessibility of DL tools has resulted in a rapid increase of DL models and subsequent peer-reviewed publications. However, the rate of deployment in clinical settings is low. Therefore, this review attempts to bring together the ideas from data collection to deployment in the clinic, building on the guidelines and principles that accreditation agencies have espoused. We introduce the need for and the role of DL to deliver accessible MRI. This is followed by a brief review of DL examples in the context of neuropathologies. Based on these studies and others, we collate the prerequisites to develop and deploy DL models for brain MRI. We then delve into the guiding principles to develop good machine learning practices in the context of neuroimaging, with a focus on explainability. A checklist based on the United States Food and Drug Administration's good machine learning practices is provided as a summary of these guidelines. Finally, we review the current challenges and future opportunities in DL for brain MRI.

Parallel-stream fusion of scan-specific and scan-general priors for learning deep MRI reconstruction in low-data regimes

Magnetic resonance imaging (MRI) is an essential diagnostic tool that suffers from prolonged scan times. Reconstruction methods can alleviate this limitation by recovering clinically usable images from accelerated acquisitions. In particular, learning-based methods promise performance leaps by employing deep neural networks as data-driven priors. A powerful approach uses scan-specific (SS) priors that leverage information regarding the underlying physical signal model for reconstruction. SS priors are learned on each individual test scan without the need for a training dataset, albeit they suffer from computationally burdening inference with nonlinear networks. An alternative approach uses scan-general (SG) priors that instead leverage information regarding the latent features of MRI images for reconstruction. SG priors are frozen at test time for efficiency, albeit they require learning from a large training dataset. Here, we introduce a novel parallel-stream fusion model (PSFNet) that synergistically fuses SS and SG priors for performant MRI reconstruction in low-data regimes, while maintaining competitive inference times to SG methods. PSFNet implements its SG prior based on a nonlinear network, yet it forms its SS prior based on a linear network to maintain efficiency. A pervasive framework for combining multiple priors in MRI reconstruction is algorithmic unrolling that uses serially alternated projections, causing error propagation under low-data regimes. To alleviate error propagation, PSFNet combines its SS and SG priors via a novel parallel-stream architecture with learnable fusion parameters. Demonstrations are performed on multi-coil brain MRI for varying amounts of training data. PSFNet outperforms SG methods in low-data regimes, and surpasses SS methods with few tens of training samples. On average across tasks, PSFNet achieves 3.1 dB higher PSNR, 2.8% higher SSIM, and 0.3 × lower RMSE than baselines. Furthermore, in both supervised and unsupervised setups, PSFNet requires an order of magnitude lower samples compared to SG methods, and enables an order of magnitude faster inference compared to SS methods. Thus, the proposed model improves deep MRI reconstruction with elevated learning and computational efficiency.

Noise2Recon: Enabling SNR-robust MRI reconstruction with semi-supervised and self-supervised learning

PURPOSE

To develop a method for building MRI reconstruction neural networks robust to changes in signal-to-noise ratio (SNR) and trainable with a limited number of fully sampled scans.

METHODS

We propose Noise2Recon, a consistency training method for SNR-robust accelerated MRI reconstruction that can use both fully sampled (labeled) and undersampled (unlabeled) scans. Noise2Recon uses unlabeled data by enforcing consistency between model reconstructions of undersampled scans and their noise-augmented counterparts. Noise2Recon was compared to compressed sensing and both supervised and self-supervised deep learning baselines. Experiments were conducted using retrospectively accelerated data from the mridata three-dimensional fast-spin-echo knee and two-dimensional fastMRI brain datasets. All methods were evaluated in label-limited settings and among out-of-distribution (OOD) shifts, including changes in SNR, acceleration factors, and datasets. An extensive ablation study was conducted to characterize the sensitivity of Noise2Recon to hyperparameter choices.

RESULTS

In label-limited settings, Noise2Recon achieved better structural similarity, peak signal-to-noise ratio, and normalized-RMS error than all baselines and matched performance of supervised models, which were trained with14 × $$ 14\times $$ more fully sampled scans. Noise2Recon outperformed all baselines, including state-of-the-art fine-tuning and augmentation techniques, among low-SNR scans and when generalizing to OOD acceleration factors. Augmentation extent and loss weighting hyperparameters had negligible impact on Noise2Recon compared to supervised methods, which may indicate increased training stability.

CONCLUSION

Noise2Recon is a label-efficient reconstruction method that is robust to distribution shifts, such as changes in SNR, acceleration factors, and others, with limited or no fully sampled training data.

Optimal MRI undersampling patterns for ultimate benefit of medical vision tasks

Compressed sensing is commonly concerned with optimizing the image quality after a partial undersampling of the measurable k-space to accelerate MRI. In this article, we propose to change the focus from the quality of the reconstructed image to the quality of the downstream image analysis outcome. Specifically, we propose to optimize the patterns according to how well a sought-after pathology could be detected or localized in the reconstructed images. We find the optimal undersampling patterns in k-space that maximize target value functions of interest in commonplace medical vision problems (reconstruction, segmentation, and classification) and propose a new iterative gradient sampling routine universally suitable for these tasks. We validate the proposed MRI acceleration paradigm on three classical medical datasets, demonstrating a noticeable improvement of the target metrics at the high acceleration factors (for the segmentation problem at ×16 acceleration, we report up to 12% improvement in Dice score over the other undersampling patterns).

Ocular condition prognosis in Keratoconus patients after corneal ring implantation using artificial neural networks

The common disorder, Keratoconus (KC), is distinguished by cumulative corneal slimming and steepening. The corneal ring implantation has become a successful surgical procedure to correct the KC patient's vision. The determination of suitable patients for the surgery alternative is among the paramount concerns of ophthalmologists. To reduce the burden on them and enhance the treatment, this research aims to previse the ocular condition of KC patients after the corneal ring implantation. It focuses on predicting post-surgical corneal topographic indices and visual characteristics. This study applied an efficacious artificial neural network approach to foretell the aforementioned ocular features of KC subjects 6 and 12 months after implanting KeraRing and MyoRing based on the accumulated data. The datasets are composed of sufficient numbers of corneal topographic maps and visual characteristics recorded from KC patients before and after implanting the rings. The visual characteristics under study are uncorrected visual acuity (UCVA), sphere (SPH), astigmatism (Ast), astigmatism orientation (Axe), and best corrected visual acuity (BCVA). In addition, the statistical data of multiple KC subjects were registered, including three effective indices of corneal topography (i.e., Ast, K-reading, and pachymetry) pre- and post-ring embedding. The outcomes represent the contribution of practical training of the introduced models to the estimation of ocular features of KC subjects following the implantation. The corneal topographic indices and visual characteristics were estimated with mean errors of 7.29% and 8.60%, respectively. Further, the errors of 6.82% and 7.65% were respectively realized for the visual characteristics and corneal topographic indices while assessing the predictions by the leave-one-out cross-validation (LOOCV) procedure. The results confirm the great potential of neural networks to guide ophthalmologists in choosing appropriate surgical candidates and their specific intracorneal rings by predicting post-implantation ocular features.

Contrast Agents of Magnetic Resonance Imaging and Future Perspective

In recent times, magnetic resonance imaging (MRI) has emerged as a highly promising modality for diagnosing severe diseases. Its exceptional spatiotemporal resolution and ease of use have established it as an indispensable clinical diagnostic tool. Nevertheless, there are instances where MRI encounters challenges related to low contrast, necessitating the use of contrast agents (CAs). Significant efforts have been made by scientists to enhance the precision of observing diseased body parts by leveraging the synergistic potential of MRI in conjunction with other imaging techniques and thereby modifying the CAs. In this work, our focus is on elucidating the rational designing approach of CAs and optimizing their compatibility for multimodal imaging and other intelligent applications. Additionally, we emphasize the importance of incorporating various artificial intelligence tools, such as machine learning and deep learning, to explore the future prospects of disease diagnosis using MRI. We also address the limitations associated with these techniques and propose reasonable remedies, with the aim of advancing MRI as a cutting-edge diagnostic tool for the future.

Two-Stage Self-Supervised Cycle-Consistency Transformer Network for Reducing Slice Gap in MR Images

Magnetic resonance (MR) images are usually acquired with large slice gap in clinical practice, i.e., low resolution (LR) along the through-plane direction. It is feasible to reduce the slice gap and reconstruct high-resolution (HR) images with the deep learning (DL) methods. To this end, the paired LR and HR images are generally required to train a DL model in a popular fully supervised manner. However, since the HR images are hardly acquired in clinical routine, it is difficult to get sufficient paired samples to train a robust model. Moreover, the widely used convolutional Neural Network (CNN) still cannot capture long-range image dependencies to combine useful information of similar contents, which are often spatially far away from each other across neighboring slices. To this end, a Two-stage Self-supervised Cycle-consistency Transformer Network (TSCTNet) is proposed to reduce the slice gap for MR images in this work. A novel self-supervised learning (SSL) strategy is designed with two stages respectively for robust network pre-training and specialized network refinement based on a cycle-consistency constraint. A hybrid Transformer and CNN structure is utilized to build an interpolation model, which explores both local and global slice representations. The experimental results on two public MR image datasets indicate that TSCTNet achieves superior performance over other compared SSL-based algorithms.

Generalizability of Deep Learning Segmentation Algorithms for Automated Assessment of Cartilage Morphology and MRI Relaxometry

BACKGROUND

Deep learning (DL)-based automatic segmentation models can expedite manual segmentation yet require resource-intensive fine-tuning before deployment on new datasets. The generalizability of DL methods to new datasets without fine-tuning is not well characterized.

PURPOSE

Evaluate the generalizability of DL-based models by deploying pretrained models on independent datasets varying by MR scanner, acquisition parameters, and subject population.

STUDY TYPE

Retrospective based on prospectively acquired data.

POPULATION

Overall test dataset: 59 subjects (26 females); Study 1: 5 healthy subjects (zero females), Study 2: 8 healthy subjects (eight females), Study 3: 10 subjects with osteoarthritis (eight females), Study 4: 36 subjects with various knee pathology (10 females).

FIELD STRENGTH/SEQUENCE

A 3-T, quantitative double-echo steady state (qDESS).

ASSESSMENT

Four annotators manually segmented knee cartilage. Each reader segmented one of four qDESS datasets in the test dataset. Two DL models, one trained on qDESS data and another on Osteoarthritis Initiative (OAI)-DESS data, were assessed. Manual and automatic segmentations were compared by quantifying variations in segmentation accuracy, volume, and T2 relaxation times for superficial and deep cartilage.

STATISTICAL TESTS

Dice similarity coefficient (DSC) for segmentation accuracy. Lin's concordance correlation coefficient (CCC), Wilcoxon rank-sum tests, root-mean-squared error-coefficient-of-variation to quantify manual vs. automatic T2 and volume variations. Bland-Altman plots for manual vs. automatic T2 agreement. A P value < 0.05 was considered statistically significant.

RESULTS

DSCs for the qDESS-trained model, 0.79-0.93, were higher than those for the OAI-DESS-trained model, 0.59-0.79. T2 and volume CCCs for the qDESS-trained model, 0.75-0.98 and 0.47-0.95, were higher than respective CCCs for the OAI-DESS-trained model, 0.35-0.90 and 0.13-0.84. Bland-Altman 95% limits of agreement for superficial and deep cartilage T2 were lower for the qDESS-trained model, ±2.4 msec and ±4.0 msec, than the OAI-DESS-trained model, ±4.4 msec and ±5.2 msec.

DATA CONCLUSION

The qDESS-trained model may generalize well to independent qDESS datasets regardless of MR scanner, acquisition parameters, and subject population.

EVIDENCE LEVEL

1 TECHNICAL EFFICACY: Stage 1.

Deep learning-assisted diagnosis of femoral trochlear dysplasia based on magnetic resonance imaging measurements

BACKGROUND

Femoral trochlear dysplasia (FTD) is an important risk factor for patellar instability. Dejour classification is widely used at present and relies on standard lateral X-rays, which are not common in clinical work. Therefore, magnetic resonance imaging (MRI) has become the first choice for the diagnosis of FTD. However, manually measuring is tedious, time-consuming, and easily produces great variability.

AIM

To use artificial intelligence (AI) to assist diagnosing FTD on MRI images and to evaluate its reliability.

METHODS

We searched 464 knee MRI cases between January 2019 and December 2020, including FTD (n = 202) and normal trochlea (n = 252). This paper adopts the heatmap regression method to detect the key points network. For the final evaluation, several metrics (accuracy, sensitivity, specificity, etc.) were calculated.

RESULTS

The accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the AI model ranged from 0.74-0.96. All values were superior to junior doctors and intermediate doctors, similar to senior doctors. However, diagnostic time was much lower than that of junior doctors and intermediate doctors.

CONCLUSION

The diagnosis of FTD on knee MRI can be aided by AI and can be achieved with a high level of accuracy.

Deep Learning Reconstruction for Accelerated Spine MRI: Prospective Analysis of Interchangeability

Background Deep learning (DL)-based MRI reconstructions can reduce examination times for turbo spin-echo (TSE) acquisitions. Studies that prospectively employ DL-based reconstructions of rapidly acquired, undersampled spine MRI are needed. Purpose To investigate the diagnostic interchangeability of an unrolled DL-reconstructed TSE (hereafter, TSE_DL) T1- and T2-weighted acquisition method with standard TSE and to test their impact on acquisition time, image quality, and diagnostic confidence. Materials and Methods This prospective single-center study included participants with various spinal abnormalities who gave written consent from November 2020 to July 2021. Each participant underwent two MRI examinations: standard fully sampled T1- and T2-weighted TSE acquisitions (reference standard) and prospectively undersampled TSE_DL acquisitions with threefold and fourfold acceleration. Image evaluation was performed by five readers. Interchangeability analysis and an image quality-based analysis were used to compare the TSE and TSE_DL images. Acquisition time and diagnostic confidence were also compared. Interchangeability was tested using the individual equivalence index regarding various degenerative and nondegenerative entities, which were analyzed on each vertebra and defined as discordant clinical judgments of less than 5%. Interreader and intrareader agreement and concordance (κ and Kendall τ and W statistics) were computed and Wilcoxon and McNemar tests were used. Results Overall, 50 participants were evaluated (mean age, 46 years ± 18 [SD]; 26 men). The TSE_DL method enabled up to a 70% reduction in total acquisition time (100 seconds for TSE_DL vs 328 seconds for TSE, P < .001). All individual equivalence indexes were less than 4%. TSE_DL acquisition was rated as having superior image noise by all readers (P < .001). No evidence of a difference was found between standard TSE and TSE_DL regarding frequency of major findings, overall image quality, or diagnostic confidence. Conclusion The deep learning (DL)-reconstructed turbo spin-echo (TSE) method was found to be interchangeable with standard TSE for detecting various abnormalities of the spine at MRI. DL-reconstructed TSE acquisition provided excellent image quality, with a 70% reduction in examination time. German Clinical Trials Register no. DRKS00023278 © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Hallinan in this issue.

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.

The KNee OsteoArthritis Prediction (KNOAP2020) challenge: An image analysis challenge to predict incident symptomatic radiographic knee osteoarthritis from MRI and X-ray images

OBJECTIVES

The KNee OsteoArthritis Prediction (KNOAP2020) challenge was organized to objectively compare methods for the prediction of incident symptomatic radiographic knee osteoarthritis within 78 months on a test set with blinded ground truth.

DESIGN

The challenge participants were free to use any available data sources to train their models. A test set of 423 knees from the Prevention of Knee Osteoarthritis in Overweight Females (PROOF) study consisting of magnetic resonance imaging (MRI) and X-ray image data along with clinical risk factors at baseline was made available to all challenge participants. The ground truth outcomes, i.e., which knees developed incident symptomatic radiographic knee osteoarthritis (according to the combined ACR criteria) within 78 months, were not provided to the participants. To assess the performance of the submitted models, we used the area under the receiver operating characteristic curve (ROCAUC) and balanced accuracy (BACC).

RESULTS

Seven teams submitted 23 entries in total. A majority of the algorithms were trained on data from the Osteoarthritis Initiative. The model with the highest ROCAUC (0.64 (95% confidence interval (CI): 0.57-0.70)) used deep learning to extract information from X-ray images combined with clinical variables. The model with the highest BACC (0.59 (95% CI: 0.52-0.65)) ensembled three different models that used automatically extracted X-ray and MRI features along with clinical variables.

CONCLUSION

The KNOAP2020 challenge established a benchmark for predicting incident symptomatic radiographic knee osteoarthritis. Accurate prediction of incident symptomatic radiographic knee osteoarthritis is a complex and still unsolved problem requiring additional investigation.

Evaluating artificial intelligence algorithms for use in veterinary radiology

Artificial intelligence is increasingly being used for applications in veterinary radiology, including detection of abnormalities and automated measurements. Unlike human radiology, there is no formal regulation or validation of AI algorithms for veterinary medicine and both general practitioner and specialist veterinarians must rely on their own judgment when deciding whether or not to incorporate AI algorithms to aid their clinical decision-making. The benefits and challenges to developing clinically useful and diagnostically accurate AI algorithms are discussed. Considerations for the development of AI research projects are also addressed. A framework is suggested to help veterinarians, in both research and clinical practice contexts, assess AI algorithms for veterinary radiology.

Multilevel Modeling of Joint Damage in Rheumatoid Arthritis

While most deep learning approaches are developed for single images, in real world applications, images are often obtained as a series to inform decision making. Due to hardware (memory) and software (algorithm) limitations, few methods have been developed to integrate multiple images so far. In this study, we present an approach that seamlessly integrates deep learning and traditional machine learning models, to study multiple images and score joint damages in rheumatoid arthritis. This method allows the quantification of joining space narrowing to approach the clinical upper limit. Beyond predictive performance, we integrate the multilevel interconnections across joints and damage types into the machine learning model and reveal the cross-regulation map of joint damages in rheumatoid arthritis.

The requirements for performing artificial-intelligence-related research and model development

Artificial intelligence research in health care has undergone tremendous growth in the last several years thanks to the explosion of digital health care data and systems that can leverage large amounts of data to learn patterns that can be applied to clinical tasks. In addition, given broad acceleration in machine learning across industries like transportation, media and commerce, there has been a significant growth in demand for machine-learning practitioners such as engineers and data scientists, who have skill sets that can be applied to health care use cases but who simultaneously lack important health care domain expertise. The purpose of this paper is to discuss the requirements of building an artificial-intelligence research enterprise including the research team, technical software/hardware, and procurement and curation of health care data.

Combined Deep Learning-based Super-Resolution and Partial Fourier Reconstruction for Gradient Echo Sequences in Abdominal MRI at 3 Tesla: Shortening Breath-Hold Time and Improving Image Sharpness and Lesion Conspicuity

RATIONALE AND OBJECTIVES

To investigate the impact of a prototypical deep learning-based super-resolution reconstruction algorithm tailored to partial Fourier acquisitions on acquisition time and image quality for abdominal T1-weighted volume-interpolated breath-hold examination (VIBE_SR) at 3 Tesla. The standard T1-weighted images were used as the reference standard (VIBE_SD).

MATERIALS AND METHODS

Patients with diverse abdominal pathologies, who underwent a clinically indicated contrast-enhanced abdominal VIBE magnetic resonance imaging at 3T between March and June 2021 were retrospectively included. Following the acquisition of the standard VIBE_SD sequences, additional images for the non-contrast, dynamic contrast-enhanced and post-contrast T1-weighted VIBE acquisition were retrospectively reconstructed using the same raw data and employing a prototypical deep learning-based super-resolution reconstruction algorithm. The algorithm was designed to enhance edge sharpness by avoiding conventional k-space filtering and to perform a partial Fourier reconstruction in the slice phase-encoding direction for a predefined asymmetric sampling ratio. In the retrospective reconstruction, the asymmetric sampling was realized by omitting acquired samples at the end of the acquisition and therefore corresponding to a shorter acquisition. Four radiologists independently analyzed the image datasets (VIBE_SR and VIBE_SD) in a blinded manner. Outcome measures were: sharpness of abdominal organs, sharpness of vessels, image contrast, noise, hepatic lesion conspicuity and size, overall image quality and diagnostic confidence. These parameters were statistically compared and interrater reliability was computed using Fleiss' Kappa and intraclass correlation coefficient (ICC). Finally, the rate of detection of hepatic lesions was documented and was statistically compared using the paired Wilcoxon test.

RESULTS

A total of 32 patients aged 59 ± 16 years (23 men (72%), 9 women (28%)) were included. For VIBE_SR, breath-hold time was significantly reduced by approximately 13.6% (VIBE_SR 11.9 ± 1.2 seconds vs. VIBE_SD: 13.9 ± 1.4 seconds, p < 0.001). All readers rated sharpness of abdominal organs, sharpness of vessels to be superior in images with VIBE_SR (p values ranged between p = 0.005 and p < 0.001). Despite reduction of acquisition time, image contrast, noise, overall image quality and diagnostic confidence were not compromised, as there was no evidence of a difference between VIBE_SR and VIBE_SD (p > 0.05). The inter-reader agreement was substantial with a Fleiss' Kappa of >0.7 in all contrast phases. A total of 13 hepatic lesions were analyzed. The four readers observed a superior lesion conspicuity in VIBE_SR than in VIBE_SD (p values ranged between p = 0.046 and p < 0.001). In terms of lesion size, there was no significant difference between VIBE_SD and VIBE_SR for all readers. Finally, there was an excellent inter-reader agreement regarding lesion size (ICC > 0.9). For all readers, no statistically significant difference was observed regarding detection of hepatic lesions between VIBE_SD and VIBE_SR.

CONCLUSION

The deep learning-based super-resolution reconstruction with partial Fourier in the slice phase-encoding direction enabled a reduction of breath-hold time and improved image sharpness and lesion conspicuity in T1-weighted gradient echo sequences in abdominal magnetic resonance imaging at 3 Tesla. Faster acquisition time without compromising image quality or diagnostic confidence was possible by using this deep learning-based reconstruction technique.

Towards autonomous analysis of chemical exchange saturation transfer experiments using deep neural networks

Macromolecules often exchange between functional states on timescales that can be accessed with NMR spectroscopy and many NMR tools have been developed to characterise the kinetics and thermodynamics of the exchange processes, as well as the structure of the conformers that are involved. However, analysis of the NMR data that report on exchanging macromolecules often hinges on complex least-squares fitting procedures as well as human experience and intuition, which, in some cases, limits the widespread use of the methods. The applications of deep neural networks (DNNs) and artificial intelligence have increased significantly in the sciences, and recently, specifically, within the field of biomolecular NMR, where DNNs are now available for tasks such as the reconstruction of sparsely sampled spectra, peak picking, and virtual decoupling. Here we present a DNN for the analysis of chemical exchange saturation transfer (CEST) data reporting on two- or three-site chemical exchange involving sparse state lifetimes of between approximately 3-60 ms, the range most frequently observed via experiment. The work presented here focuses on the ¹H CEST class of methods that are further complicated, in relation to applications to other nuclei, by anti-phase features. The developed DNNs accurately predict the chemical shifts of nuclei in the exchanging species directly from anti-phase ¹H^N CEST profiles, along with an uncertainty associated with the predictions. The performance of the DNN was quantitatively assessed using both synthetic and experimental anti-phase CEST profiles. The assessments show that the DNN accurately determines chemical shifts and their associated uncertainties. The DNNs developed here do not contain any parameters for the end-user to adjust and the method therefore allows for autonomous analysis of complex NMR data that report on conformational exchange.

Assessing the Economic Value of Clinical Artificial Intelligence: Challenges and Opportunities

OBJECTIVES

Clinical artificial intelligence (AI) is a novel technology, and few economic evaluations have focused on it to date. Before its wider implementation, it is important to highlight the aspects of AI that challenge traditional health technology assessment methods.

METHODS

We used an existing broad value framework to assess potential ways AI can provide good value for money. We also developed a rubric of how economic evaluations of AI should vary depending on the case of its use.

RESULTS

We found that the measurement of core elements of value-health outcomes and cost-are complicated by AI because its generalizability across different populations is often unclear and because its use may necessitate reconfigured clinical processes. Clinicians' productivity may improve when AI is used. If poorly implemented though, AI may also cause clinicians' workload to increase. Some AI has been found to exacerbate health disparities. Nevertheless, AI may promote equity by expanding access to medical care and, when properly trained, providing unbiased diagnoses and prognoses. The approach to assessment of AI should vary based on its use case: AI that creates new clinical possibilities can improve outcomes, but regulation and evidence collection may be difficult; AI that extends clinical expertise can reduce disparities and lower costs but may result in overuse; and AI that automates clinicians' work can improve productivity but may reduce skills.

CONCLUSIONS

The potential uses of clinical AI create challenges for health technology assessment methods originally developed for pharmaceuticals and medical devices. Health economists should be prepared to examine data collection and methods used to train AI, as these may impact its future value.

Introduction to Deep Learning in Clinical Neuroscience

The use of deep learning (DL) is rapidly increasing in clinical neuroscience. The term denotes models with multiple sequential layers of learning algorithms, architecturally similar to neural networks of the brain. We provide examples of DL in analyzing MRI data and discuss potential applications and methodological caveats.Important aspects are data pre-processing, volumetric segmentation, and specific task-performing DL methods, such as CNNs and AEs. Additionally, GAN-expansion and domain mapping are useful DL techniques for generating artificial data and combining several smaller datasets.We present results of DL-based segmentation and accuracy in predicting glioma subtypes based on MRI features. Dice scores range from 0.77 to 0.89. In mixed glioma cohorts, IDH mutation can be predicted with a sensitivity of 0.98 and specificity of 0.97. Results in test cohorts have shown improvements of 5-7% in accuracy, following GAN-expansion of data and domain mapping of smaller datasets.The provided DL examples are promising, although not yet in clinical practice. DL has demonstrated usefulness in data augmentation and for overcoming data variability. DL methods should be further studied, developed, and validated for broader clinical use. Ultimately, DL models can serve as effective decision support systems, and are especially well-suited for time-consuming, detail-focused, and data-ample tasks.

Upstream Machine Learning in Radiology

Machine learning (ML) and Artificial intelligence (AI) has the potential to dramatically improve radiology practice at multiple stages of the imaging pipeline. Most of the attention has been garnered by applications focused on improving the end of the pipeline: image interpretation. However, this article reviews how AI/ML can be applied to improve upstream components of the imaging pipeline, including exam modality selection, hardware design, exam protocol selection, data acquisition, image reconstruction, and image processing. A breadth of applications and their potential for impact is shown across multiple imaging modalities, including ultrasound, computed tomography, and MRI.

Low-count whole-body PET with deep learning in a multicenter and externally validated study

More widespread use of positron emission tomography (PET) imaging is limited by its high cost and radiation dose. Reductions in PET scan time or radiotracer dosage typically degrade diagnostic image quality (DIQ). Deep-learning-based reconstruction may improve DIQ, but such methods have not been clinically evaluated in a realistic multicenter, multivendor environment. In this study, we evaluated the performance and generalizability of a deep-learning-based image-quality enhancement algorithm applied to fourfold reduced-count whole-body PET in a realistic clinical oncologic imaging environment with multiple blinded readers, institutions, and scanner types. We demonstrate that the low-count-enhanced scans were noninferior to the standard scans in DIQ (p < 0.05) and overall diagnostic confidence (p < 0.001) independent of the underlying PET scanner used. Lesion detection for the low-count-enhanced scans had a high patient-level sensitivity of 0.94 (0.83-0.99) and specificity of 0.98 (0.95-0.99). Interscan kappa agreement of 0.85 was comparable to intrareader (0.88) and pairwise inter-reader agreements (maximum of 0.72). SUV quantification was comparable in the reference regions and lesions (lowest p-value=0.59) and had high correlation (lowest CCC = 0.94). Thus, we demonstrated that deep learning can be used to restore diagnostic image quality and maintain SUV accuracy for fourfold reduced-count PET scans, with interscan variations in lesion depiction, lower than intra- and interreader variations. This method generalized to an external validation set of clinical patients from multiple institutions and scanner types. Overall, this method may enable either dose or exam-duration reduction, increasing safety and lowering the cost of PET imaging.

Diagnostic Accuracy of Quantitative Multicontrast 5-Minute Knee MRI Using Prospective Artificial Intelligence Image Quality Enhancement

BACKGROUND. Potential approaches for abbreviated knee MRI, including prospective acceleration with deep learning, have achieved limited clinical implementation. OBJECTIVE. The objective of this study was to evaluate the interreader agreement between conventional knee MRI and a 5-minute 3D quantitative double-echo steady-state (qDESS) sequence with automatic T2 mapping and deep learning super-resolutionaugmentation and to compare the diagnostic performance of the two methods regarding findings from arthroscopic surgery. METHODS. Fifty-one patients with knee pain underwent knee MRI that included an additional 3D qDESS sequence with automatic T2 mapping. Fourier interpolation was followed by prospective deep learning super resolution to enhance qDESS slice resolution twofold. A musculoskeletal radiologist and a radiology resident performed independent retrospective evaluations of articular cartilage, menisci, ligaments, bones, extensor mechanism, and synovium using conventional MRI. Following a 2-month washout period, readers reviewed qDESS images alone followed by qDESS with the automatic T2 maps. Interreader agreement between conventional MRI and qDESS was computed using percentage agreement and Cohen kappa. The sensitivity and specificity of conventional MRI, qDESS alone, and qDESS plus T2 mapping were compared with arthroscopic findings using exact McNemar tests. RESULTS. Conventional MRI and qDESS showed 92% agreement in evaluating all tissues. Kappa was 0.79 (95% CI, 0.76-0.81) across all imaging findings. In 43 patients who underwent arthroscopy, sensitivity and specificity were not significantly different (p = .23 to > .99) between conventional MRI (sensitivity, 58-93%; specificity, 27-87%) and qDESS alone (sensitivity, 54-90%; specificity, 23-91%) for cartilage, menisci, ligaments, and synovium. For grade 1 cartilage lesions, sensitivity and specificity were 33% and 56%, respectively, for conventional MRI; 23% and 53% for qDESS (p = .81); and 46% and 39% for qDESS with T2 mapping (p = .80). For grade 2A lesions, values were 27% and 53% for conventional MRI, 26% and 52% for qDESS (p = .02), and 58% and 40% for qDESS with T2 mapping (p < .001). CONCLUSION. The qDESS method prospectively augmented with deep learning showed strong interreader agreement with conventional knee MRI and near-equivalent diagnostic performance regarding arthroscopy. The ability of qDESS to automatically generate T2 maps increases sensitivity for cartilage abnormalities. CLINICAL IMPACT. Using prospective artificial intelligence to enhance qDESS image quality may facilitate an abbreviated knee MRI protocol while generating quantitative T2 maps.

Update on brain MRI for the diagnosis and follow-up of MS patients

Over the past decades, MRI has become a major tool in the diagnosis and the follow-up of patients with multiple sclerosis (MS), especially for monitoring the effectiveness of therapy. The recent international recommendations issued for the standardization of neurological and radiological clinical practices converge on many points. In this setting, recommendations made by the "Observatoire français de la sclérose en plaques", the French MS registry, can be distinguished by its interdisciplinary complementarity, its longevity, its size, and its positions in direct connection with the clinic. Hence, after suspicions of gadolinium deposition in the brain, with multiple warning from the American and European health authorities, a national consultation took place and resulted in limitation to useful injections. The precautionary principle prevailing, the patient receives a limited quantity of contrast product even if no clinically harmful manifestation has been detected to date. The result of this round table bringing together neurologists and neuroradiologists from specialized centers was published in the form of a recommendation in early 2020. The interest of this project also lies in the constant improvement of the management of patients with MS and the possibility of developing advanced techniques to assist the clinician. The aim of this review is to explain to the neurologist, the interest of following this imaging protocol both in his/her clinical practice and in the possibilities that this opens up.

The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset

PURPOSE

To organize a multi-institute knee MRI segmentation challenge for characterizing the semantic and clinical efficacy of automatic segmentation methods relevant for monitoring osteoarthritis progression.

MATERIALS AND METHODS

A dataset partition consisting of three-dimensional knee MRI from 88 retrospective patients at two time points (baseline and 1-year follow-up) with ground truth articular (femoral, tibial, and patellar) cartilage and meniscus segmentations was standardized. Challenge submissions and a majority-vote ensemble were evaluated against ground truth segmentations using Dice score, average symmetric surface distance, volumetric overlap error, and coefficient of variation on a holdout test set. Similarities in automated segmentations were measured using pairwise Dice coefficient correlations. Articular cartilage thickness was computed longitudinally and with scans. Correlation between thickness error and segmentation metrics was measured using the Pearson correlation coefficient. Two empirical upper bounds for ensemble performance were computed using combinations of model outputs that consolidated true positives and true negatives.

RESULTS

Six teams (T 1-T 6) submitted entries for the challenge. No differences were observed across any segmentation metrics for any tissues (P = .99) among the four top-performing networks (T 2, T 3, T 4, T 6). Dice coefficient correlations between network pairs were high (> 0.85). Per-scan thickness errors were negligible among networks T 1-T 4 (P = .99), and longitudinal changes showed minimal bias (< 0.03 mm). Low correlations (ρ < 0.41) were observed between segmentation metrics and thickness error. The majority-vote ensemble was comparable to top-performing networks (P = .99). Empirical upper-bound performances were similar for both combinations (P = .99).

CONCLUSION

Diverse networks learned to segment the knee similarly, where high segmentation accuracy did not correlate with cartilage thickness accuracy and voting ensembles did not exceed individual network performance.See also the commentary by Elhalawani and Mak in this issue.Keywords: Cartilage, Knee, MR-Imaging, Segmentation © RSNA, 2020Supplemental material is available for this article.

Identification of Tumor-Specific MRI Biomarkers Using Machine Learning (ML)

The identification of reliable and non-invasive oncology biomarkers remains a main priority in healthcare. There are only a few biomarkers that have been approved as diagnostic for cancer. The most frequently used cancer biomarkers are derived from either biological materials or imaging data. Most cancer biomarkers suffer from a lack of high specificity. However, the latest advancements in machine learning (ML) and artificial intelligence (AI) have enabled the identification of highly predictive, disease-specific biomarkers. Such biomarkers can be used to diagnose cancer patients, to predict cancer prognosis, or even to predict treatment efficacy. Herein, we provide a summary of the current status of developing and applying Magnetic resonance imaging (MRI) biomarkers in cancer care. We focus on all aspects of MRI biomarkers, starting from MRI data collection, preprocessing and machine learning methods, and ending with summarizing the types of existing biomarkers and their clinical applications in different cancer types.

MRSaiFE: An AI-based Approach Towards the Real-Time Prediction of Specific Absorption Rate

The purpose of this study is to investigate feasibility of estimating the specific absorption rate (SAR) in MRI in real time. To this goal, SAR maps are predicted from 3T- and 7T-simulated magnetic resonance (MR) images in 10 realistic human body models via a convolutional neural network. Two-dimensional (2-D) U-Net architectures with varying contraction layers and different convolutional filters were designed to estimate the SAR distribution in realistic body models. Sim4Life (ZMT, Switzerland) was used to create simulated anatomical images and SAR maps at 3T and 7T imaging frequencies for Duke, Ella, Charlie, and Pregnant Women (at 3, 7, and 9 month gestational stages) body models. Mean squared error (MSE) was used as the cost function and the structural similarity index (SSIM) was reported. A 2-D U-Net with 4 contracting (and 4 expanding) layers and 64 convolutional filters at the initial stage showed the best compromise to estimate SAR distributions. Adam optimizer outperformed stochastic gradient descent (SGD) for all cases with an average SSIM of 90.5∓3.6 % and an average MSE of 0.7∓0.6% for head images at 7T, and an SSIM of >85.1∓6.2 % and an MSE of 0.4∓0.4% for 3T body imaging. Algorithms estimated the SAR maps for 224×224 slices under 30 ms. The proposed methodology shows promise to predict real-time SAR in clinical imaging settings without using extra mapping techniques or patient-specific calibrations.

Artificial Intelligence and Big Data

Medical images play an important role in ophthalmology and radiology. Medical image analysis has greatly benefited from the application of "deep learning" techniques in clinical and experimental radiology. Clinical applications and their relevance for radiological imaging in ophthalmology are presented.