Artificial intelligence–powered magnetic resonance imaging reconstruction: methods, challenges, and emerging trends—a narrative review

Introduction

Magnetic resonance imaging (MRI) is an indispensable diagnostic tool that provides exceptional soft-tissue contrast without ionizing radiation exposure, making it essential for neurological, cardiac, and musculoskeletal imaging (1,2). Despite its diagnostic advantages, MRI remains limited by inherently slow data acquisition due to sequential k-space sampling, which results in long scan times, motion artifacts, and patient discomfort (3,4). Accelerating MRI acquisition while maintaining image quality has therefore been a major focus of research for more than two decades.

Early acceleration strategies such as parallel imaging (PI) and compressed sensing (CS) introduced foundational improvements by exploiting coil sensitivity encoding and sparsity priors (2,5-7). However, these traditional methods are constrained by their hand-crafted regularizers and fixed mathematical models, often leading to residual artifacts, poor generalization, and reduced performance at high acceleration factors. A comparative summary of conventional and AI-based reconstruction techniques is provided in Table 1 (7).

The evolution of deep learning (DL) and artificial intelligence (AI) has redefined MRI reconstruction paradigms by enabling data-driven priors that learn complex mappings from undersampled k-space data to fully sampled images. Convolutional neural networks (CNNs), generative adversarial networks (GANs), transformers, and model-driven unrolled architectures have achieved substantial gains in reconstruction fidelity, artifact suppression, and inference speed (11-13). In particular, physics-guided networks that embed data-consistency layers have bridged the gap between traditional optimization and purely data-driven learning, offering interpretable and stable reconstructions suitable for clinical translation (14).

As illustrated in Figures 1-4, the field has progressed from conventional CS/PI pipelines to hybrid AI architectures that incorporate physical modeling, multi-coil encoding, and attention-based global context learning. These developments have led to state-of-the-art (SOTA) performance benchmarks on public datasets such as fastMRI, Calgary-Campinas, and IXI (15). The fastMRI dataset is a large-scale, multi-coil benchmark released by NYU and Facebook AI Research that has become a de facto standard for evaluating accelerated MRI reconstruction. The Calgary-Campinas dataset provides multi-vendor brain MRI data that enables robustness assessment across scanner platforms, while the IXI dataset offers publicly available multi-contrast brain MRI scans frequently used for cross-domain validation studies. Nevertheless, the community continues to face challenges involving generalization across scanners, limited data availability, and interpretability of learned reconstructions (16,17).

Figure 1 Conventional versus AI-based optimization frameworks in medical imaging. Schematic comparison between a conventional pipeline relying on handcrafted feature extraction, mathematical optimization, and rule-based decision making, and an AI-based optimized pipeline that employs data-driven feature learning, deep neural network models, and end-to-end learning frameworks. The AI-based approach enhances adaptability and enables automated optimization, supporting improved image quality and robustness in medical imaging applications (illustrative). AI, artificial intelligence.

Figure 2 Conventional CS/PI versus AI/deep learning pipelines. Conventional CS/PI uses hand-crafted priors and iterative solvers and may retain residual artifacts under aggressive undersampling. AI/deep learning maps measurements to images using learned priors—including physics-guided networks—targeting higher fidelity at comparable acceleration (illustrative). AI, artificial intelligence; CS, compressed sensing; PI, parallel imaging.

Figure 3 Taxonomy of AI-based MRI reconstruction methods. Representative families are CNN-based, GAN-based, transformer-based, and unrolled/model-driven approaches with embedded DC updates. Each exhibits distinct trade-offs in reconstruction fidelity, global-context modeling, computational cost, and artifact suppression. Schematic representation. AI, artificial intelligence; CNN, convolutional neural network; DC, data consistency; GAN, generative adversarial network; MRI, magnetic resonance imaging.

Figure 4 Timeline of key milestones in MRI reconstruction and AI. Selected landmarks include SENSE (1999), GRAPPA (2002), compressed sensing for MRI (2007), U-Net and variational networks (2015–2018), fastMRI release (2018), transformer-based models (from 2021), hybrid CNN-transformer methods (from 2023), and privacy-preserving/federated learning initiatives (2024–2025). Years are approximate and intended as an orientation guide. AI, artificial intelligence; CNN, convolutional neural network; GRAPPA, Generalized Autocalibrating Partially Parallel Acquisition; MRI, magnetic resonance imaging; SENSE, Sensitivity Encoding.

The present narrative review aims to provide a consolidated overview of the technical foundations, taxonomy of AI models, and evolving trends in physics-informed and federated MRI reconstruction. It also discusses critical open issues—including dataset bias, reproducibility, and clinical validation—that must be addressed before AI-driven MRI reconstruction can achieve widespread adoption in radiological workflows. We present this article in accordance with the Narrative Review reporting checklist (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-240/rc).

Methods

We conducted a focused literature search across PubMed, Scopus, IEEE Xplore, and arXiv. The overall coverage spanned 1999–2025, with a primary focus on January 2015–October 2025 to capture contemporary deep learning developments; earlier seminal works were included for foundational context. Search terms included magnetic resonance imaging reconstruction, deep learning, artificial intelligence, compressed sensing, parallel imaging, unrolled networks, transformers, and federated learning. We included English-language peer-reviewed articles and benchmark reports; preprints were included selectively to capture emerging trends when peer-reviewed alternatives were not yet available.

A summary of the search strategy, databases, timeframe, inclusion/exclusion criteria, and selection process is provided in Table 2 (search strategy summary). We extracted model family, dataset/domain setting (e.g., vendor, field strength when reported), sampling/acceleration factors, and evaluation metrics [peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), normalized mean squared error (NMSE)], and summarized representative models, datasets, and metrics in Table S1 for transparency.

Technical foundations

MRI reconstruction transforms raw k-space data into spatial domain images using mathematical operations based on Fourier transforms (1,2). The primary goal is to reconstruct high-quality anatomical information from undersampled datasets while minimizing acquisition time and maintaining diagnostic accuracy. Conventional MRI reconstruction relies on sequential k-space sampling, resulting in prolonged acquisition times and motion artifacts (3,4).

Acceleration techniques such as PI and CS have addressed these constraints. PI techniques like SENSE and GRAPPA leverage coil sensitivity encoding to reconstruct missing k-space data, reducing scan time by exploiting spatial redundancy (5,6). However, their performance declines at higher acceleration factors due to noise amplification and conditioning limitations.

CS introduced a sparsity-driven framework that reconstructs images from sub-Nyquist samples using nonlinear optimization. The general CS formulation can be represented as:

where A is the undersampled Fourier encoding operator (including coil sensitivities in multi-coil settings), y is the acquired k-space data, λ is a regularization parameter, and is a sparsity-promoting prior (e.g., total variation or transform sparsity) (7). Although CS improved reconstruction fidelity and acceleration, it suffers from computational complexity and manual regularization tuning. Figure 5 depicts the general workflow for conventional MRI reconstruction, and Table 3 summarizes comparative aspects of PI and CS techniques. Despite their utility, these approaches reveal a trade-off between reconstruction speed, image quality, and noise resilience, motivating AI-based solutions (11,18).

Figure 5 K-space undersampling and compressed-sensing reconstruction. Variable-density undersampling masks at acceleration R produce incoherent aliasing in the image domain. CS enforces DC and sparsity using iterative solvers (e.g., ISTA/FISTA/ADMM) to recover images. Typical accelerations of ~2–4× are shown for illustration; achievable factors depend on sequence, anatomy, and sampling pattern. ADMM, alternating direction method of multipliers; CS, compressed sensing; DC, data consistency; FISTA, fast ISTA; ISTA, iterative shrinkage-thresholding algorithm.

AI model families for MRI reconstruction

DL-based frameworks reformulate the MRI reconstruction problem by directly learning mappings from undersampled k-space inputs to fully sampled outputs (8-10,19). CNNs represent the earliest and most widely used architecture family. Models such as U-Net, DeepCascade, and DC-CNN exploit hierarchical feature extraction to suppress aliasing artifacts and restore structural details (8-10,20,21). Figure 6 illustrates the qualitative trade-off between inference speed and reconstruction fidelity across major AI model families.

Figure 6 Illustrative speed-fidelity trade-offs across AI model families for MRI reconstruction. Qualitative comparison of CNN-based, GAN-based, transformer-based, and unrolled/model-driven approaches along axes of inference speed and reconstruction fidelity. Relative positioning is conceptual and may vary with dataset characteristics, undersampling patterns, network architecture, and deployment setting. AI, artificial intelligence; CNN, convolutional neural network; GAN, generative adversarial network; MRI, magnetic resonance imaging.

Generative models such as GANs and autoencoders have introduced perceptual consistency through adversarial training, ensuring high-fidelity and visually realistic reconstructions (8-10,19). However, these approaches may also induce hallucinated features, making quantitative validation and uncertainty estimation critical for clinical translation (8-10,19). Transformer architectures, including ReconFormer and SwinMRI, leverage attention mechanisms to capture long-range dependencies and global context, outperforming convolutional counterparts in generalization and interpretability (8,9,13).

Unrolled architectures combine physical forward models with learnable regularization modules. Each network layer emulates an optimization iteration, ensuring data-consistency and stability. Notable examples include MoDL, VarNet, and E2E-VarNet, achieving state-of-the-art performance with interpretable training dynamics (11,12,22,23). A concise comparison of these model families—CNN, GAN, Transformer, and unrolled networks—is presented in Table 4.

Model-driven, physics-informed, and hybrid approaches

Model-driven and physics-informed networks integrate MR physics constraints into the learning process, enhancing data consistency and interpretability (25-28). These frameworks embed MR acquisition operators and data-consistency constraints into trainable architectures, enforcing fidelity to imaging physics. Physics-informed neural networks (PINNs) further incorporate differential equations and acquisition priors into the loss function, ensuring stability under varying sampling trajectories (26-28).

Hybrid models combine CNN or Transformer backbones with unrolled or PINN-inspired modules, achieving a balance between learning capacity and domain knowledge. For instance, MoDL-PINN and Dual-Domain HybridNet architectures simultaneously process k-space and image-domain features, reducing residual artifacts and improving signal-to-noise ratio across contrasts (19,25,29-32). Figure 7 illustrates representative unrolled and hybrid frameworks, respectively (20,25,29). Tables 5,6 summarize comparative results highlighting reconstruction quality and computational efficiency.

Figure 7 Unrolled and hybrid MRI reconstruction frameworks. Unified schematic showing (A) unrolled DC-CNN cascades alternating data-consistency and regularization blocks, and (B) physics-informed hybrid models integrating MR forward models and learned priors for physically consistent reconstruction. CNN, convolutional neural network; DC, data consistency; MRI, magnetic resonance imaging; PINN, physics-informed neural network.

Despite promising results, model-driven and hybrid strategies face challenges related to scalability, cross-domain generalization, and reproducibility. Limited access to paired datasets and high GPU memory demand remains key constraints. Adaptive loss weighting, domain-transfer learning, and dynamic k-space sampling have been proposed to address these issues (27-31).

Federated and distributed training for MRI reconstruction

Federated learning (FL) facilitates collaborative model training across distributed medical centers without direct data sharing (37-40). Each institution trains local models on-site, transmitting only weight updates to a central aggregator. This approach preserves patient privacy while leveraging large-scale data diversity (37,39,40). Figure 8 depicts the typical FL workflow for multi-institutional MRI reconstruction. Compared with centralized learning, FL ensures compliance with data-protection regulations such as GDPR and HIPAA (37).

Figure 8 Federated-learning workflow for MRI reconstruction. Each site trains locally on private data and shares encrypted model updates for secure aggregation; the global model is redistributed to all participants. No raw data leave the institution. MRI, magnetic resonance imaging.

Advanced variants such as FedProx and FedBN improve model convergence under heterogeneous scanner conditions and non-IID data (38,40). Figure 9 shows an edge-cloud pipeline combining FL with distributed inference, enabling real-time reconstruction on edge devices while maintaining cloud-level accuracy (41-45). Tables 7,8 compare centralized and federated configurations in terms of communication latency, accuracy, and scalability.

Figure 9 Edge-cloud integration for AI-assisted MRI reconstruction. The scanner (edge) performs lightweight preprocessing and transmits intermediate data for high-capacity cloud inference; the reconstructed image is returned to the console. Security and latency considerations apply. AI, artificial intelligence; MRI, magnetic resonance imaging.

Despite progress, FL frameworks encounter challenges in synchronization, communication bottlenecks, and privacy-utility trade-offs. Techniques including differential privacy, homomorphic encryption, and blockchain-based secure aggregation have been explored to enhance trustworthiness and transparency (43-46). These federated and distributed paradigms mark a pivotal shift toward scalable, privacy-preserving AI deployment for MRI reconstruction.

Discussion

This section synthesizes the comparative evaluation, clinical implications, and open challenges of AI approaches in MRI reconstruction. Integrating the analyses from previous chapters, it highlights the strengths and limitations of CNNs, GANs, transformer architectures, and unrolled or physics-informed hybrid models. Benchmark results, trade-off analyses, and clinical mappings are consolidated through Tables 9-11 and Figure 10, supported by foundational studies (11-18), with additional developments reported in (8-10,19-23,25,26), with further validation reported in studies (27-32,34-37), with additional supporting evidence from studies (38-46,55), with further corroboration provided by studies (24,33,47-53,56), and with recent findings reported in studies (54,57-69).

Figure 10 Qualitative trade-offs across AI model families for MRI reconstruction. CNNs provide fast inference with moderate fidelity, GANs enhance perceptual realism, transformers capture long-range context for superior fidelity, and unrolled architectures ensure data-consistent, interpretable reconstructions. AI, artificial intelligence; CNN, convolutional neural network; GAN, generative adversarial network; MRI, magnetic resonance imaging.

Comparative analysis of AI architectures

The comparative value of major AI architecture families is best understood through their practical trade-offs in reconstruction fidelity, robustness to domain shift, computational cost, and clinical risk. CNN-based methods remain strong baselines due to fast inference and stable optimization, but their local receptive fields can limit robustness to global anatomical/contextual variation at high acceleration. GAN-based methods can enhance perceptual sharpness, yet they require careful validation because perceptual realism can mask subtle errors and increase hallucination risk under aggressive undersampling. Transformer-based approaches improve global consistency through attention mechanisms and have shown stronger cross-domain behavior in several benchmarks, though they are typically more compute- and data-hungry. Unrolled/model-driven networks often achieve the best balance between fidelity and stability by explicitly enforcing data consistency with the MR forward model, at the cost of higher memory and runtime requirements.

Across widely used benchmarks (Tables 9,10), transformer and unrolled frameworks frequently report high SSIM and PSNR values under common acceleration settings; however, absolute numbers vary with sampling masks, preprocessing, coil configuration, and evaluation protocols. To improve interpretability of reported performance, we now highlight that multi-domain testing often reveals measurable generalization gaps on external data (e.g., typical PSNR reductions on the order of 1–3 dB and SSIM decreases of ~2–5% under vendor/protocol shifts, depending on anatomy and acceleration). We also emphasize the importance of transparent reporting of code availability, training data composition, and benchmark settings (e.g., fastMRI challenge rules) to support reproducible comparisons across studies. These architectural trade-offs motivate the growing interest in physics-informed and federated approaches discussed next.

Clinical mapping and applicability

The clinical translation of AI-based MRI reconstruction depends not only on image fidelity but also on diagnostic reliability. Clinical reader studies demonstrate that AI-accelerated MRI can reduce scan times by 30–50% without compromising lesion detection or grading accuracy (10,15,24,52-54). As summarized in Table 11, CNN and GAN models are predominantly adopted in musculoskeletal and diffusion-weighted imaging due to their texture-preserving characteristics, whereas transformer and unrolled models dominate cardiac and neuroimaging applications for their robustness to motion and noise (10,34,35,52-54). Integration with Picture Archiving and Communication Systems (PACS) and DICOM workflows is emerging, facilitating direct clinical validation of AI reconstructions.

Emerging challenges and open issues

Generalization and domain shift

A persistent challenge in MRI reconstruction is the generalization of trained models across different scanner vendors, field strengths, coil configurations, and acquisition protocols. Models trained on limited or homogeneous datasets can degrade when applied to external cohorts, leading to domain-shift artifacts and potential diagnostic risk (49,56-58,60,70). Representative multi-domain evaluations quantify this effect, often reporting decreases in PSNR (approximately 1–3 dB) and SSIM (approximately 2–5%) on unseen scanners or protocols, although the magnitude depends on anatomy, sampling pattern, and acceleration. Domain adaptation and meta-learning approaches have been explored to mitigate these variations. FL (Figure 8) has emerged as a promising strategy, allowing multiple institutions to collaboratively train models by exchanging weight updates rather than raw data, thereby enhancing robustness while preserving privacy (37-40,46).

Interpretability and trust

Despite substantial accuracy gains, most DL frameworks operate as black boxes, limiting clinician confidence. Explainable AI (XAI) techniques, such as saliency maps, gradient-based relevance propagation, and uncertainty quantification, partially address this issue (61-63). However, these methods remain descriptive rather than diagnostic, lacking the quantitative interpretability required for regulatory validation. Future models should embed MRI physics awareness directly into network architectures to ensure traceable reconstruction logic.

Data scarcity and benchmarking

The availability of large, annotated MRI datasets is limited due to privacy regulations and acquisition costs. Public datasets like fastMRI, CHAOS, and Calgary-Campinas partially address this gap but often lack pathological diversity and standardized evaluation (15,50,53,54,59). Synthetic data augmentation using GANs and simulation-based pipelines helps increase sample diversity but risks bias propagation. Unified benchmarking frameworks with consistent sampling masks and metric definitions (PSNR, SSIM, NMSE) are critical for reproducibility (15,24,51,59). Recent strategies also attempt to train deep reconstruction models directly from undersampled or synthetic data when fully sampled ground truth is unavailable (65).

Regulatory readiness and clinical translation

Regulatory approval for AI-driven MRI reconstruction tools requires adherence to Good Machine Learning Practice (GMLP) and continuous post-market surveillance (64,67-69). Frameworks from the U.S. FDA, EMA, and CDSCO emphasize bias auditing, version tracking, and human-in-the-loop validation. Despite technical maturity, few reconstruction systems have undergone prospective multi-center clinical trials. Integrating AI tools into PACS and DICOM-compliant systems, along with explainability documentation, will be essential for regulatory endorsement and clinician acceptance.

Computational cost and deployment constraints

Advanced architectures such as transformers and unrolled networks require significant computational resources, often exceeding 12–16 GB GPU memory, limiting real-time scanner integration (41-44,55). Lightweight designs, including MobileNet-based CNNs, pruning, and mixed-precision inference, reduce complexity while maintaining quality. Edge–cloud hybrid architectures (Figure 9) provide a viable compromise by enabling inference on distributed hardware while maintaining high-fidelity reconstruction (41).

While a formal risk-of-bias scoring framework was not applied due to the narrative design of this review, we qualitatively assessed transparency and reproducibility indicators across studies, including code and weight availability, dataset accessibility, benchmark participation, and completeness of reporting. We observed variability across model families: benchmark-driven CNN and unrolled/model-based approaches generally exhibit more standardized evaluation practices, whereas emerging transformer-based and FL frameworks demonstrate greater heterogeneity in reporting and reproducibility protocols. These differences highlight the need for harmonized benchmarking and regulatory-aligned reporting standards.

In summary, the comparative evaluation underscores the diversity and maturity of AI frameworks in MRI reconstruction. While transformer and unrolled networks demonstrate superior fidelity, CNN and GAN models retain advantages in interpretability and speed. Addressing generalization, interpretability, and regulatory compliance will be pivotal for clinical-scale deployment. Collaborative benchmarking, FL, and edge-integrated inference strategies represent the next steps toward sustainable, trustworthy, and clinically translatable MRI reconstruction. Open repositories such as the ISMRM MR-Hub and MR-Pub initiatives further promote transparent code sharing and reproducible benchmarking (66).

Future directions and conclusions

Emerging directions

Future work should prioritize specific, testable gaps that currently limit clinical translation. Key needs include: (I) prospective multi-center clinical reader studies that evaluate downstream diagnostic endpoints (not only PSNR/SSIM); (II) standardized regulatory-ready benchmarking protocols that report data provenance, scanner/vendor context, and uncertainty; and (III) robust uncertainty quantification and failure-mode detection to mitigate hallucination risk under aggressive acceleration. Within this agenda, FL remains a practical pathway to train on diverse multi-institutional data without transferring raw images, and can improve domain robustness when combined with heterogeneity-aware strategies (e.g., FedProx/FedBN) (37-40,45,46).

In parallel, deployment research should focus on scanner-integrated inference under realistic latency and resource constraints. Edge–cloud designs can reduce on-scanner compute burden by performing lightweight preprocessing locally while using higher-capacity inference resources when permitted by institutional policy. Studies should report end-to-end latency, memory footprints, and workflow integration details (PACS/DICOM compatibility) to support reproducible deployment comparisons (41-44,55).

Finally, trustworthy operation requires interpretability and transparency aligned with clinical and regulatory expectations. Beyond post-hoc saliency maps, practical approaches include uncertainty maps, calibration reporting, and attention visualization to flag low-confidence regions and guide radiologist review (62,63). Transparent documentation of training data composition, bias audits, and model versioning will be essential for Good Machine Learning Practice (GMLP) and sustained post-deployment monitoring.

Clinical readiness and ethical compliance

Achieving clinical translation extends beyond algorithmic accuracy. Ethical AI deployment demands transparent documentation of training datasets, bias audits, and governance aligned with FAIR/FATE principles. Collaborative benchmarking under standardized datasets such as fastMRI, OCMR, and CMRxRecon will remain vital for reproducibility and trust. Integrating federated and edge–cloud paradigms into regulatory sandboxes may help institutions validate reconstruction pipelines under controlled conditions before large-scale deployment.

Clinically, AI-accelerated MRI will influence several domains: emergency and interventional MRI, where rapid reconstructions reduce decision latency; low-field and portable systems, where robust generalization compensates for hardware limitations; and quantitative imaging, where physics-informed models yield reproducible biomarkers. Ethical frameworks must ensure that automation enhances, rather than replaces, radiologist judgment, maintaining human oversight in the diagnostic loop. Maintaining realistic expectations of AI’s clinical impact remains vital, as emphasized by Ranschaert et al. in their analysis of AI in radiology’s promise versus hype (71).

Strengths and limitations of this review

This review provides a comprehensive synthesis of AI-based MRI reconstruction spanning data-driven, model-based, and federated paradigms, highlighting both technological advances and translational barriers. However, as a narrative review, it is limited by the selective inclusion of studies and lack of quantitative meta-analysis; future systematic evaluations may further validate the observed trends.

Conclusions

In summary, the translational success of AI-assisted MRI reconstruction will depend not only on architectural sophistication but also on privacy-preserving design, interpretability, and compliance with evolving medical-AI standards. As highlighted throughout this review, integrating advances in CS, DL, and hybrid model-based approaches with clinically grounded validation remains essential. Building this ecosystem will move the field closer to real-time, transparent, and equitable imaging for every patient.

Acknowledgments

The authors used ChatGPT (OpenAI, San Francisco, CA, USA) only for language editing (clarity, grammar, and phrasing). The tool was not used to generate scientific content, perform data analysis/interpretation, or select references. All factual statements, interpretations, and conclusions were developed and verified by the authors, who accept full responsibility for the integrity and accuracy of the manuscript (ChatGPT version: GPT-5; sessions during July–October 2025).

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-240/rc

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-240/prf

Funding: None.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-1-240/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Haacke EMB, Brown RW, Thompson MR, et al. Magnetic resonance imaging: physical principles and sequence design. 2nd ed. J. Wiley & Sons; 2014.
2. Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magn Reson Med 2007;58:1182-95. [Crossref] [PubMed]
3. Weller DS, Wang L, Mugler JP 3rd, et al. Motion-compensated reconstruction of magnetic resonance images from undersampled data. Magn Reson Imaging 2019;55:36-45. [Crossref] [PubMed]
4. Knoll F, Hammernik K, Zhang C, et al. Deep-Learning Methods for Parallel Magnetic Resonance Imaging Reconstruction: A Survey of the Current Approaches, Trends, and Issues. IEEE Signal Process Mag 2020;37:128-40. [Crossref] [PubMed]
5. Pruessmann KP, Weiger M, Scheidegger MB, et al. SENSE: sensitivity encoding for fast MRI. Magn Reson Med 1999;42:952-62.
6. Griswold MA, Jakob PM, Heidemann RM, et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 2002;47:1202-10. [Crossref] [PubMed]
7. Feng L, Benkert T, Block KT, et al. Compressed sensing for body MRI. J Magn Reson Imaging 2017;45:966-87. [Crossref] [PubMed]
8. Huang J, Wu Y, Wang F, et al. Data- and Physics-Driven Deep Learning Based Reconstruction for Fast MRI: Fundamentals and Methodologies. IEEE Rev Biomed Eng 2025;18:152-171. [Crossref] [PubMed]
9. Heckel R, Jacob M, Chaudhari A, et al. Deep learning for accelerated and robust MRI reconstruction. MAGMA 2024;37:335-68. [Crossref] [PubMed]
10. Singh D, Monga A, de Moura HL, et al. Emerging Trends in Fast MRI Using Deep-Learning Reconstruction on Undersampled k-Space Data: A Systematic Review. Bioengineering (Basel) 2023;10:1012. [Crossref] [PubMed]
11. Hammernik K, Klatzer T, Kobler E, et al. Learning a variational network for reconstruction of accelerated MRI data. Magn Reson Med 2018;79:3055-71. [Crossref] [PubMed]
12. Aggarwal HK, Mani MP, Jacob M. MoDL: Model-Based Deep Learning Architecture for Inverse Problems. IEEE Trans Med Imaging 2019;38:394-405. [Crossref] [PubMed]
13. Guo P, Mei Y, Zhou J, et al. ReconFormer: Accelerated MRI Reconstruction Using Recurrent Transformer. IEEE Trans Med Imaging 2024;43:582-93. [Crossref] [PubMed]
14. Kim S, Park H, Park SH. A review of deep learning-based reconstruction methods for accelerated MRI using spatiotemporal and multi-contrast redundancies. Biomed Eng Lett 2024;14:1221-42. [Crossref] [PubMed]
15. Knoll F, Zbontar J, Sriram A, et al. fastMRI: A Publicly Available Raw k-Space and DICOM Dataset of Knee Images for Accelerated MR Image Reconstruction Using Machine Learning. Radiol Artif Intell 2020;2:e190007. [Crossref] [PubMed]
16. Nayak KS, Lim Y, Campbell-Washburn AE, et al. Real-Time Magnetic Resonance Imaging. J Magn Reson Imaging 2022;55:81-99. [Crossref] [PubMed]
17. Wang Y, Zeng T, Liu F, et al. Illuminating the unseen: Advancing MRI domain generalization through causality. Med Image Anal 2025;101:103459. [Crossref] [PubMed]
18. Cordero-Grande L, Ferrazzi G, Teixeira RPAG, et al. Motion-corrected MRI with DISORDER: Distributed and incoherent sample orders for reconstruction deblurring using encoding redundancy. Magn Reson Med 2020;84:713-26. [Crossref] [PubMed]
19. Huang J, Fang Y, Nan Y, et al. Data and physics-driven learning models for fast MRI: fundamentals and methodologies from CNN, GAN to attention and transformers. IEEE Signal Process Mag 2022;
20. Schlemper J, Caballero J, Hajnal JV, et al. A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction. IEEE Trans Med Imaging 2018;37:491-503. [Crossref] [PubMed]
21. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical image computing and computer-assisted intervention-MICCAI 2015. Lecture Notes in Computer Science. Cham: Springer; 2015:234-41.
22. Sriram A, Zbontar J, Murrell T, et al. End-to-end variational networks for accelerated MRI reconstruction. In: Martel AL, Abolmaesumi P, Stoyanov D, et al., editors. Medical image computing and computer assisted intervention-MICCAI 2020. Lecture Notes in Computer Science. Cham: Springer; 2020:64-73.
23. Knoll F, Hammernik K, Kobler E, et al. Assessment of the generalization of learned image reconstruction and the potential for transfer learning. Magn Reson Med 2019;81:116-28. [Crossref] [PubMed]
24. Muckley MJ, Riemenschneider B, Radmanesh A, et al. Results of the 2020 fastMRI Challenge for Machine Learning MR Image Reconstruction. IEEE Trans Med Imaging 2021;40:2306-17. [Crossref] [PubMed]
25. Hammernik K, Küstner T, Yaman B, et al. Physics-Driven Deep Learning for Computational Magnetic Resonance Imaging: Combining physics and machine learning for improved medical imaging. IEEE Signal Process Mag 2023;40:98-114. [Crossref] [PubMed]
26. Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J Comput Phys 2019;378:686-707.
27. Zhu Y, Cheng J, Cui ZX, et al. Physics-driven deep learning methods for fast quantitative magnetic resonance imaging: performance improvements through integration with deep neural networks. IEEE Signal Processing Magazine 2023;40:116-28.
28. Ahmadi M, Biswas D, Lin M, et al. Physics-informed machine learning for advancing computational medical imaging: integrating data-driven approaches with fundamental physical principles. Artif Intell Rev 2025;58:297.
29. Wang J, Liu C, Zhong Y, et al. Deep plug-and-play MRI reconstruction based on multiple complementary priors. Magn Reson Imaging 2025;115:110244. [Crossref] [PubMed]
30. Eo T, Jun Y, Kim T, et al. KIKI-net: cross-domain convolutional neural networks for reconstructing undersampled magnetic resonance images. Magn Reson Med 2018;80:2188-201. [Crossref] [PubMed]
31. Yaman B, Gu H, Hosseini SAH, et al. Multi-mask self-supervised learning for physics-guided neural networks in highly accelerated magnetic resonance imaging. NMR Biomed 2022;35:e4798. [Crossref] [PubMed]
32. Huang B, Yang F, Yin M, et al. A Review of Multimodal Medical Image Fusion Techniques. Comput Math Methods Med 2020;2020:8279342. [Crossref] [PubMed]
33. Yang Y, Sun J, Li H, et al. Deep ADMM-Net for compressive sensing MRI. NIPS'16: Proceedings of the 30th International Conference on Neural Information Processing Systems; 2016; Barcelona, Spain. IEEE; 2016:10-18.
34. Endres J, Weinmüller S, Dang HN, et al. Phase distribution graphs for fast, differentiable, and spatially encoded Bloch simulations of arbitrary MRI sequences. Magn Reson Med 2024;92:1189-204. [Crossref] [PubMed]
35. Yaman B, Hosseini SAH, Moeller S, et al. Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data. Magn Reson Med 2020;84:3172-91. [Crossref] [PubMed]
36. Yaman B, Hosseini SAH, Moeller S, et al. Self-supervised physics-based deep learning MRI reconstruction without fully sampled data. In: Proceedings of the IEEE 17th International Symposium on Biomedical Imaging (ISBI); 2020; Virtual (Iowa City, IA, USA). p. 921-5.
37. Kaissis GA, Makowski MR, Rückert D, et al. Secure, privacy-preserving and federated machine learning in medical imaging. Nat Mach Intell 2020;2:305-311.
38. McMahan HB, Moore E, Ramage D, et al. Communication-efficient learning of deep networks from decentralized data. In: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS 2017). Fort Lauderdale, Florida; PMLR; 2017:1273-82.
39. Sheller MJ, Edwards B, Reina GA, et al. Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Sci Rep 2020;10:12598. [Crossref] [PubMed]
40. Kairouz P, McMahan HB, Avent B, et al. Advances and open problems in federated learning. Found Trends Mach Learn 2021;14:1-210.
41. Alçalar YU, Cao Y, Akçakaya M. Edge computing for physics-driven AI in computational MRI: a feasibility study. arXiv:2506.03183 [Preprint]. 2025 [cited 2025 Oct 3]. Available online: https://arxiv.org/abs/2506.03183
42. Xu Y, Khan TM, Song Y, et al. Edge deep learning in computer vision and medical diagnostics: a comprehensive survey. Artif Intell Rev 2025;58:93.
43. Chatterjee N, Duda J, Gee J, et al. A Cloud-Based System for Automated AI Image Analysis and Reporting. J Imaging Inform Med 2025;38:368-79. [Crossref] [PubMed]
44. Gayathri S, Gowri S. Securing medical image privacy in cloud using deep learning network. J Cloud Comp 2023;12:40.
45. Shokri R, Shmatikov V. Privacy-preserving deep learning. Proceedings of the 22nd ACM SIGSAC Conference on Computer and Communications Security (CCS 2015); Denver, CO, USA. 2015:1310-21.
46. Moulahi W, Jdey I, Moulahi T, et al. A blockchain-based federated learning mechanism for privacy preservation of healthcare IoT data. Comput Biol Med 2023;167:107630. [Crossref] [PubMed]
47. Zbontar J, Knoll F, Sriram A, et al. fastMRI: an open dataset and benchmarks for accelerated MRI. arXiv:1811.08839 [Preprint]. 2018 [cited 2025 Oct 3]. Available online: https://arxiv.org/abs/1811.08839
48. Zhang J, Zhao C, Gao W. Optimization-Inspired Compact Deep Compressive Sensing. IEEE Journal of Selected Topics in Signal Processing. 2020;14:765-774.
49. Liu J, Qin C, Yaghoobi M. High-fidelity MRI reconstruction using adaptive spatial attention selection and deep data consistency prior. IEEE Transactions on Computational Imaging 2023;9:298-313.
50. Souza R, Lucena O, Garrafa J, et al. An open, multi-vendor, multi-field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement. Neuroimage 2018;170:482-94. [Crossref] [PubMed]
51. Beauferris Y, Teuwen J, Karkalousos D, et al. Multi-Coil MRI Reconstruction Challenge-Assessing Brain MRI Reconstruction Models and Their Generalizability to Varying Coil Configurations. Front Neurosci 2022;16:919186. [Crossref] [PubMed]
52. Vornehm M, Wetzl J, Giese D, et al. CineVN: Variational network reconstruction for rapid functional cardiac cine MRI. Magn Reson Med 2025;93:138-50. [Crossref] [PubMed]
53. Chen C, Zhang Y, Xu J, et al. OCMR: open-access multi-coil k-space dataset for cardiac MRI. arXiv:2008.03410 [Preprint]. 2020 [cited 2025 Oct 3]. Available online: https://arxiv.org/abs/2008.03410
54. Wang C, Lyu J, Wang S, et al. CMRxRecon: A publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI. Sci Data 2024;11:687. [Crossref] [PubMed]
55. Boudlal H, Serrhini M, Tahiri A. Cloud computing application in medical imaging: challenges and opportunities. 2022 5th International Conference on Networking, Information Systems and Security: Envisage Intelligent Systems in 5g//6G-based Interconnected Digital Worlds (NISS), Bandung, Indonesia, 2022:1-5.
56. Güngör A, Dar SU, Özsavaş EE, et al. Adaptive diffusion priors for robust MRI reconstruction. Med Image Anal 2023;88:102872. [Crossref] [PubMed]
57. Ouyang C, Schlemper J, Biffi C, et al. Generalising deep learning MRI reconstruction across different domains. arXiv:1902.10815 [Preprint]. 2019 [cited 2025 Oct 3]. Available online: https://arxiv.org/abs/1902.10815
58. Ramanarayanan S, Palla A, Ram K, et al. Generalizing supervised deep learning MRI reconstruction to multiple and unseen contrasts using meta-learning hypernetworks. Appl Soft Comput 2023;146:110633.
59. Inati SJ, Naegele JD, Zwart NR, et al. ISMRM Raw data format: A proposed standard for MRI raw datasets. Magn Reson Med 2017;77:411-21. [Crossref] [PubMed]
60. Abbasi S, Lan H, Choupan J, et al. Deep learning for the harmonization of structural MRI scans: a survey. Biomed Eng Online 2024;23:90. [Crossref] [PubMed]
61. van der Velden BHM, Kuijf HJ, Gilhuijs KGA, et al. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal 2022;79:102470. [Crossref] [PubMed]
62. Edupuganti V, Mardani M, Vasanawala S, et al. Uncertainty Quantification in Deep MRI Reconstruction. IEEE Trans Med Imaging 2021;40:239-50. [Crossref] [PubMed]
63. Selvaraju RR, Cogswell M, Das A, et al. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int J Comput Vis 2020;128:336-59.
64. Wilkinson MD, Dumontier M, Aalbersberg IJ, et al. The FAIR Guiding Principles for scientific data management and stewardship. Sci Data 2016;3:160018. [Crossref] [PubMed]
65. Zeng G, Guo Y, Zhan J, et al. A review on deep learning MRI reconstruction without fully sampled k-space. BMC Med Imaging 2021;21:195. [Crossref] [PubMed]
66. ISMRM MR-Hub & MR-Pub. Open-access software tools for MR research. 2019-2025 [cited 2025 Sep 29]. Available online: https://ismrm.github.io/mrhub
67. Koçak B, Ponsiglione A, Stanzione A, et al. Bias in artificial intelligence for medical imaging: fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects. Diagn Interv Radiol 2025;31:75-88. [Crossref] [PubMed]
68. Tripathi S, Gabriel K, Dheer S, et al. Understanding Biases and Disparities in Radiology AI Datasets: A Review. J Am Coll Radiol 2023;20:836-41. [Crossref] [PubMed]
69. U.S. Food and Drug Administration. Artificial intelligence/machine learning (AI/ML)-based software as a medical device (SaMD) action plan. Silver Spring (MD): FDA; 2021 Jan [cited 2025 Sep 29]. Available online: https://www.fda.gov/media/145022/download
70. Guan H, Liu M. Domain Adaptation for Medical Image Analysis: A Survey. IEEE Trans Biomed Eng 2022;69:1173-85. [Crossref] [PubMed]
71. Ranschaert E. Artificial intelligence in radiology: hype or hope? J Belg Soc Radiol 2018;102:20.