coil sensitivity estimation

Long MRI acquisition times hinder clinical efficiency and exacerbate motion artifacts. Conventional parallel imaging techniques (e.g., SENSE) rely on pre-acquired coil sensitivity maps, entailing complex calibration procedures and susceptibility to spatial misalignment. This paper proposes an end-to-end deep learning framework that jointly estimates coil sensitivity maps and reconstructs images directly from 4× undersampled multi-channel k-space data—eliminating the need for separate calibration scans. To our knowledge, this is the first method enabling joint sensitivity and image learning from a single undersampled acquisition. The architecture features two co-optimized branches: a sensitivity estimation module and a U-Net-based reconstruction module. Evaluated on brain MRI data from 10 subjects, the method yields reconstructions with visual quality comparable to SENSE; although PSNR and SSIM are marginally lower, it achieves substantially improved robustness and clinical practicality.

This work proposes UEPS, a novel architecture for accelerated MRI reconstruction that addresses the domain shift and degraded generalization performance commonly caused by inaccurate coil sensitivity map estimation in deep unrolling models. UEPS eliminates reliance on coil sensitivity maps by employing per-coil independent reconstruction, progressive resolution refinement from k-space to image space, and an MRI-specific sparse attention mechanism tailored for 1D undersampling patterns. The method demonstrates superior robustness across ten out-of-distribution test sets encompassing diverse anatomies, views, contrasts, scanner vendors, field strengths, and coil configurations. Furthermore, UEPS enables low-latency inference, offering both computational efficiency and strong potential for clinical deployment.

This work addresses the limitations of existing diffusion-based MRI reconstruction methods, which rely on large networks, opaque time-conditioning mechanisms, and offline-estimated coil sensitivity maps, leading to poor interpretability and limited adaptability across acquisition protocols. To overcome these issues, the authors propose a novel paradigm that jointly reconstructs both the image and coil sensitivities within an end-to-end framework. Central to this approach is a parameter-efficient Gaussian mixture product diffusion model serving as an interpretable image prior, complemented by a smoothness prior on coil sensitivities. The resulting method enables k-space trajectory-adaptive reconstruction, achieving rapid convergence and strong robustness under shifts in contrast and anatomical distribution as well as varying sampling trajectories, thereby significantly enhancing denoising and reconstruction performance.

This work addresses the limitations of traditional MRI reconstruction methods, which discretize both image content and coil sensitivities, resulting in high memory consumption and poor structural awareness—particularly challenging in highly undersampled dynamic 3D cardiac imaging. The authors propose the first model-driven framework based on continuous neural fields, representing magnetization and coil sensitivities through tensor products of univariate neural fields to yield continuously differentiable representations. Integrated with a differentiable MRI physical forward model, this approach overcomes the constraints of discrete formulations. Evaluated under extreme undersampling conditions with acceleration factors up to 16, the method substantially outperforms existing model-driven techniques, effectively preserving fine anatomical details and temporal motion consistency.

Conventional quantitative MRI (qMRI) struggles to jointly suppress motion and magnetic field inhomogeneity artifacts under highly accelerated acquisitions, degrading R2* quantification accuracy. To address this, we propose the first end-to-end interpretable deep-unfolding framework that unifies motion modeling, field-map correction, and biophysical signal modeling directly within a k-space-domain reconstruction architecture—enabling artifact-free R2* map estimation from undersampled multi-echo gradient-recalled echo (mGRE) raw data. Crucially, we introduce a novel self-supervised training paradigm that requires no prior parameter estimation, eliminating dependence on ground-truth R2* maps or field-map labels. Evaluated on real accelerated mGRE data, our method significantly reduces R2* quantification error compared to conventional two-step approaches, while demonstrating superior robustness and cross-scanner generalizability.

This work addresses the significant degradation in k-space reconstruction quality of magnetic resonance imaging (MRI) under undersampled and high-noise conditions by proposing a unified high-dimensional k-space reconstruction framework. The method enhances the representational capacity of the data space through the incorporation of high-dimensional embedding priors, enabling existing diffusion-based inverse problem solvers to operate more robustly in an enriched k-space without altering the underlying diffusion model architecture or optimization procedure. As a model-agnostic enhancement mechanism, the proposed framework consistently improves reconstruction performance across multiple public and internal datasets, with the most pronounced gains observed in high-noise scenarios.

Highly accelerated magnetic resonance fingerprinting (MRF) imaging is prone to undersampling aliasing artifacts and lacks large-scale training data with quantitative ground truth. To address these challenges, this work proposes MRI2Qmap, a novel framework that, for the first time, integrates a physics-driven compressed sensing model with a deep denoising autoencoder pretrained on large-scale conventional weighted MRI data. By employing a plug-and-play optimization strategy, the method enables quantitative multi-parameter map reconstruction without requiring quantitative ground truth labels. It effectively leverages anatomical priors embedded in routine clinical MRI scans, significantly improving MRF reconstruction quality. Evaluated on highly accelerated 3D whole-brain MRF data, the proposed approach achieves performance comparable to or better than existing methods, thereby overcoming the dependency on quantitative ground truth for training.

This study addresses the significant performance degradation of radiomics-based AI models caused by multicenter heterogeneity in CT acquisition protocols, which hinders their clinical translation. To tackle this issue, the authors introduce mixed-effects modeling into radiomics sensitivity analysis for the first time, establishing a performance-oriented analytical framework that simultaneously accounts for subject-specific random effects and fixed effects of CT acquisition parameters. This approach enables quantitative assessment of how key scanning parameters influence model robustness. Leveraging multicenter CT data, multiple deep learning architectures, and cross-dataset validation, the study identifies an optimal CT protocol—tube current ≥200 mA, pitch ≤1.5, and slice thickness ≤1.25 mm—that substantially improves diagnostic sensitivity from 0.79 to 0.90 and specificity from 0.47 to 0.79, thereby markedly enhancing cross-center generalizability.

This work addresses the problem of magnetic resonance image reconstruction and uncertainty quantification from undersampled k-space data by formulating reconstruction as a Bayesian linear inverse problem. A total variation prior is introduced to capture the sparsity of image gradients, and an efficient split-augmented Gibbs sampler is designed for posterior inference. The proposed method achieves, for the first time in MRI reconstruction, simultaneous high-fidelity imaging and pixel-wise uncertainty estimation, with the quantified uncertainties showing strong correlation with actual reconstruction errors. Experiments on both single-coil and multi-coil datasets demonstrate that the approach outperforms conventional compressed sensing algorithms in reconstruction quality while providing reliable and interpretable uncertainty measures.

This work addresses the challenge of motion artifacts in 3D brain MRI, particularly under highly accelerated acquisitions where joint image reconstruction and motion parameter estimation remain difficult. The authors propose a unified Bayesian framework that unsupervisedly and jointly estimates the anatomical image, rigid-body motion parameters, and coil sensitivity maps directly from motion-corrupted k-space data. The key innovation lies in the first-time integration of a pre-trained 3D complex-valued diffusion model as an image prior within a physics-driven motion-compensated reconstruction pipeline. Efficient inference is achieved through alternating diffusion posterior sampling and proximal optimization. Experiments demonstrate that the proposed method significantly outperforms existing classical and learning-based approaches on both simulated and real motion-contaminated data, delivering superior image quality and robustness—especially under severe motion and high acceleration scenarios.