Hybrid learning: a combination of self-supervised and supervised learning for joint MRI reconstruction and denoising in low-field MRI
Clinical Snapshot
PICO Framework
| P — Population | MRI datasets from breast, lung, and brain imaging acquired across field strengths of 0.3 T to 3 T, including both simulated and real noisy MRI data with varying undersampling ratios and sampling trajectories (Cartesian, spiral, radial) |
| I — Intervention | Hybrid learning framework — a two-stage training pipeline combining self-supervised learning (SSL) to generate pseudo-reference images from fully sampled low-SNR data, followed by supervised learning using those pseudo-references as training targets for joint reconstruction and denoising of undersampled, noisy k-space data |
| C — Comparator | Standard supervised learning using noisy references; standard self-supervised learning (SSL) alone |
| O — Outcomes | Image reconstruction quality measured by Structural Similarity Index Measure (SSIM), Normalised Mean Squared Error (NMSE), and High-Frequency Error Norm (HFEN) across varying acceleration rates, noise levels, and sampling patterns |
Bottom Line
This paper presents a technically innovative two-stage hybrid learning framework for MRI reconstruction and denoising that combines self-supervised and supervised deep learning, specifically targeting the low-SNR conditions inherent to low-field MRI. The reported quantitative improvements over both supervised and self-supervised baselines are numerically substantial. However, the study is a proof-of-concept technical validation without clinical endpoints, statistical inference, or patient-level outcomes. The absence of confidence intervals, sample size reporting, and diagnostic accuracy data means that clinical significance cannot be determined from the available evidence. The hallucination risk inherent to deep learning reconstruction — where the network may generate anatomically plausible but diagnostically incorrect features — is not addressed. For Australian clinicians and medical physicists, the framework is conceptually relevant to expanding low-field MRI utility in rural and remote settings, but TGA SaMD regulatory review and local clinical validation studies would be mandatory prerequisites to any deployment. Senior clinicians should regard this as promising early-stage technical research warranting prospective clinical validation rather than practice-changing evidence.
Key Findings
P Value: Not reported
Effect Size: Versus noisy-reference supervised learning: up to 167.70% higher SSIM, 95.41% lower NMSE, 90.70% lower HFEN. Versus standard SSL: up to 23.88% higher SSIM, 60.85% lower NMSE, 49.13% lower HFEN
Primary Outcome: Image reconstruction quality assessed by SSIM, NMSE, and HFEN across four experiments using simulated and real noisy MRI data of breast, lung, and brain at 0.3 T–3 T
Nnt Or Sensitivity: Not applicable — no clinical diagnostic or therapeutic endpoints assessed; no NNT, sensitivity, specificity, or hazard ratio calculable from reported data
Confidence Interval: Not reported
Clinical Application
The framework requires deep learning infrastructure for two-stage training, which may limit immediate adoption in smaller radiology departments. Inference deployment after training could be feasible on standard clinical workstations, but computational requirements are not specified. Integration into existing PACS or scanner reconstruction pipelines would require vendor collaboration and regulatory clearance. Low-field MRI is an emerging area of interest in Australia, particularly for point-of-care applications in rural and remote settings where high-field MRI access is limited. The TGA would need to evaluate any clinical implementation of this reconstruction software as a Software as a Medical Device (SaMD) under the ARTG framework. The RACGP and RANZCR have not yet issued specific guidance on deep learning MRI reconstruction in low-field settings. PBS implications are indirect — improved low-field MRI quality could support expanded Medicare Benefits Schedule (MBS) item utilisation for MRI in underserved regions. Australian researchers and clinical physicists should note that local scanner validation studies would be required before clinical deployment, as performance on Australian low-field platforms (e.g., Hyperfine Swoop or similar) has not been demonstrated. Patients undergoing MRI at low-field strength scanners (particularly 0.3 T), including settings where high-SNR reference data for deep learning training are unavailable — potentially relevant to point-of-care MRI, paediatric imaging, implant-compatible scanning, and resource-limited healthcare environments
Abstract
Objective.Deep learning has demonstrated strong potential for magnetic resonance imaging (MRI) reconstruction. However, conventional supervised learning requires high-quality, high-signal-to-noise-ratio (SNR) reference data for network training, which are often difficult or impossible to obtain, particularly in low-field MRI. Self-supervised learning (SSL) eliminates the need for reference training data but may suffer from degraded performance under low-SNR conditions. To address these limitations, we propose hybrid learning, a new training framework that integrates self-supervised and supervised learning for joint MRI reconstruction and denoising when only low-SNR training data are available.Approach.Hybrid learning is implemented in two sequential stages. In the first stage, SSL is applied to fully sampled low-SNR data to generate higher-quality pseudo-references. In the second stage, these pseudo-references are then used as targets for supervised learning to reconstruct and denoise undersampled, noisy data. The proposed method was evaluated in four experiments using simulated and real noisy MRI data of the breast, lung, and brain across different field strengths (0.3 T to 3 T), sampling trajectories (Cartesian, spiral, and radial), noise levels, and undersampling ratios.Main Results.Hybrid learning consistently improved reconstruction quality relative to both supervised and self-supervised baselines under different acceleration rates, noise levels, and sampling patterns in all experiments. Compared with standard supervised learning using noisy references, it achieved up to 167.70% higher structural similarity index measure (SSIM), 95.41% lower normalized mean squared error (NMSE), and 90.70% lower high-frequency error norm (HFEN). Compared with standard SSL, it achieved up to 23.88% higher SSIM, 60.85% lower NMSE, and 49.13% lower HFEN.Significance.Hybrid learning enables improved MRI reconstruction under low-SNR imaging conditions by jointly addressing noise and undersampling. It provides a practical solution for robust deep learning-based reconstruction and is particularly well suited for applications such as low-field MRI, where image quality is limited by reduced SNR.
References
- 1.Pei, H., Janjušević, N., Luo, R., Xia, D., Xu, X., Moore, W., Wang, Y., Chandarana, H., & Feng, L. (2026). Hybrid learning: a combination of self-supervised and supervised learning for joint MRI reconstruction and denoising in low-field MRI. Physics in Medicine and Biology. https://doi.org/10.1088/1361-6560/ae792b
Related Research
Physics in medicine and biology
Anatomically adaptive feature-wise linear modulation for deep learning-based low-dose CT denoising
28 July 2026
Science advances
Interpretable dynamic quantitative vascular morphometry features using SHAP for anti-angiogenic therapy response prediction
27 July 2026
Pediatric surgery international
Pediatric colonic diverticulitis: clinical presentation, management, and review of the literature
21 July 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service