Joint Supervised and Self-supervised Learning (JSSL) trains an MRI reconstruction model on a fully-sampled proxy domain and a subsampled target domain simultaneously. However, the supervised loss on the proxy domain converges quickly and produces small gradients, while the self-supervised loss on the target domain retains large, noisy gradients. This persistent magnitude imbalance allows the proxy domain, despite its smaller gradients, to dominate optimization, degrading target-domain reconstruction. The Pareto Optimization Strategy (POS) addresses gradient conflict via minimum-norm aggregation but fails in this setting because it is magnitude-sensitive. It allocates weight inversely to gradient norm, silencing the target domain when its gradients are large. A simple lower-bound constraint on the target weight recovers substantial performance, revealing that the failure stems from asymmetric gradient scaling rather than directional conflict. We propose Target-Anchored Gradient Integration (TAGI), which resolves the imbalance by operating on the unit sphere of gradient directions, decoupling direction from magnitude. TAGI anchors the target domain at unit weight, treats the proxy as a tunable regularizer, and applies conflict-aware projection to neutralize opposing gradient components. We establish a formal equivalence between TAGI and constrained POS, showing that constrained POS implicitly performs norm-ratio-adaptive weighting. This reduces TAGI to a single mixing ratio c, for which we derive a data-driven adaptive scheduler. Across six cross-domain configurations at 4× and 8× acceleration, TAGI attains the highest PSNR on ten of twelve settings, with a peak gain of ＋4.12 dB over JSSL, and the best SSIM on seven of twelve settings. Ablation studies confirm that the norm-ratio-adaptive trust coefficient provides the largest individual contribution, while conflict projection and the data-driven scheduler each yield measurable improvements, and the method is robust to the choice of the mixing ratio.

Major Depressive Disorder (MDD) is a psychiatric disorder associated with altered neuro-metabolism. This study investigated the effect of antidepressant therapy on the metabolism of the right hippocampus and anterior cingulate cortex (ACC) in patients with MDD, using in vivo proton magnetic resonance spectroscopy (MRS) at 3.0 T. MRS was performed in 42 patients with MDD (at baseline and after 8 weeks of antidepressant therapy) and 35 healthy controls (HCs). Metabolite concentrations were quantified using Osprey software with partial-volume corrections. The Hamilton Depression Rating Scale (HDRS) was used to assess response to antidepressant therapy. Patients with MDD who had a > 50% reduction in HDRS scores were classified as remitted, and those with a < 50% reduction were grouped as non-remitted after therapy. Compared with HCs, baseline levels of hippocampal total N-acetylaspartate (tNAA), total creatine (tCr), total choline (tCho), and myo-inositol (mI) were reduced. Baseline levels of ACC tNAA and tCr were higher in patients with MDD than in HCs. After antidepressant therapy, a significant increase in tNAA and tCr levels was observed in the hippocampus of patients with remitted MDD, indicating a positive impact of therapy on energy metabolism and neurodegenerative processes, thereby supporting remission. Hippocampal metabolite levels remained unaltered after therapy in the non-remitted subgroup. ACC metabolites remained unaltered after therapy in both remitted and non-remitted subgroups. Thus, findings highlighted the importance of hippocampal metabolism in MDD and the possibility that metabolite levels may serve as biomarkers of antidepressant therapy response.

Magnetic resonance imaging (MRI) is the preferred modality for non-invasive examination of the human brain, but the acquired images are inevitably corrupted by Rician noise and by intensity non-uniformity (INU), also called the bias field, which arises from imperfect radio-frequency coils and from subject-induced field perturbations, and these artefacts degrade subsequent registration, segmentation and tumour-detection pipelines. This article reviews 151 publications spanning four decades (1985 to early 2026) on the removal of noise and intensity inhomogeneity from brain MRI, located through a structured search of IEEE Xplore, PubMed, Scopus, Web of Science and Google Scholar and complemented by manual snowballing. The methods surveyed are organised into prospective approaches (phantom-based, multi-coil and special-sequence techniques), retrospective approaches (filtering, surface fitting, segmentation-based and histogram-based correction), classical machine learning, and deep learning (convolutional networks, generative adversarial networks, denoising autoencoders, transformer-based architectures, denoising diffusion probabilistic models and self-supervised frameworks). Deep-learning approaches to bias-field correction are reviewed separately and classified into direct field regression, adversarial translation, joint estimation of the field and the uniform image, and frequency-domain probabilistic formulations, a distinction that clarifies why supervision, rather than architecture, is the limiting factor in that literature. On the BrainWeb T1 benchmark at Rician noise σ = 9%, 3D-Parallel-RicianNet reaches a published peak signal-to-noise ratio (PSNR) of 37.11 dB and structural similarity index (SSIM) of 0.9859, compared with 28.91 dB and 0.9584 for BM3D. The 2023-2026 generation of models-Swin-UNet Transformers, the Imaging Transformer for SNR ≪ 1, hybrid HTC-Net, the lightweight CHARMS CNN-Transformer, denoising diffusion probabilistic models and self-supervised frameworks such as Coil2Coil, SURE-Net, Rep2Rep and the Efficient Collaborative Diffusion Model-report further methodological gains and remove the requirement for paired noise-free training data, which had been the principal obstacle to clinical deployment. Open challenges remain in cross-scanner generalisation, in clinically meaningful evaluation beyond PSNR and SSIM, and in the construction of a publicly available multi-vendor benchmark for brain-MRI denoising.

The problem of spatially selective radio-frequency (RF) pulse design in magnetic resonance imaging (MRI) is typically stated in the form of determining, analytically or numerically, RF waveforms to be applied in synchrony with one or more predetermined gradient waveforms. In most cases, the dynamics of the nuclear spin magnetization under the RF and gradient fields is described in a global rotating frame that cancels the effect of the static (main) magnetic field B0. In this work, we consider an alternative frame of reference, which can be called a local rotating frame, where total longitudinal magnetic field (B0 plus gradient) in every voxel is zero. In this frame, the effect of time-dependent gradient field is integrated out, and the remaining magnetization dynamics, governed by much weaker RF fields, becomes both simpler and slower. We show that recasting existing RF design methods in such a frame provides useful insights and techniques that are not obvious in the conventional description. The methods we consider include (i) two-dimensional spatial RF pulse design in the excitation k-space, (ii) Shinnar-Le Roux RF pulse design, (iii) residual phase calculation in slice-selective excitation, and (iv) iterative and numerical solutions for multi-coil RF pulse design. In particular, we show that the new formalism can substantially reduce the Bloch simulation time which can greatly benefit iterative pulse design in parallel transmit. In all, the proposed framework provides considerable theoretical insights and practical utility for RF pulse design in MRI.

Magnetic resonance spectroscopy (MRS) provides non-invasive in vivo assessment of brain tumour metabolism and complements conventional magnetic resonance imaging (MRI) by adding biochemical information beyond structural imaging. This narrative state-of-the-art review summarises current clinical and emerging applications of proton MRS and magnetic resonance spectroscopic imaging (MRSI) in brain tumours, with emphasis on metabolic biomarkers, technical developments, clinical translation, and current limitations. Established spectroscopic patterns, particularly elevated choline, reduced N-acetylaspartate, and variable lactate and lipid signals, support lesion characterisation, glioma grading, and differentiation of tumour recurrence from treatment-related effects. MRSI further enables spatial mapping of metabolic heterogeneity, helping to identify infiltrative and non-enhancing tumour regions relevant to biopsy targeting, surgical planning, and radiotherapy guidance. Molecularly informative metabolites, especially 2-hydroxyglutarate in isocitrate dehydrogenase-mutant gliomas, extend the role of MRS toward non-invasive molecular imaging, while emerging markers such as alanine and glycine remain promising but require further validation. Advances in localisation, acceleration, spectral editing, automated quantification, and artificial intelligence-assisted analysis have improved feasibility and reproducibility, although variability in acquisition protocols, quantification methods, quality control, and reporting continues to limit widespread clinical adoption. Hyperpolarized [13]C and deuterium-based methods offer dynamic assessment of tumour metabolism, but largely remain translational or research-oriented techniques. Overall, MRS and MRSI are increasingly positioned as useful components of multiparametric neuro-oncology imaging, provided that future work continues to prioritise protocol harmonisation, multicentre validation, clinically actionable thresholds, and integration into routine diagnostic and treatment-planning workflows.

Q-space trajectory imaging (QTI) provides promising markers of tissue microstructure, but clinical translation requires shorter acquisitions, faster analysis, and more robust parameter estimation at high spatial resolution. To address these barriers, we trained a voxel-wise multilayer perceptron (MLP) to infer QTI-derived scalar parameters directly from the diffusion signal. We established reference QTI parameters of the brain in 18 healthy subjects using constrained fitting on 50-min QTI scans. The MLP was trained to estimate those reference parameters from a five-minute subset of the diffusion data. We compared the MLP with the constrained fit applied to the same short-protocol input, computing normalized root mean squared error, peak signal-to-noise ratio, and structural similarity with respect to the reference. Here, the MLP consistently achieved better performance metrics, with normalized root mean squared errors up to two-fold lower. For one whole-brain dataset, MLP inference reduced computation time from more than an hour with constrained fitting to a few seconds. Robustness to lower SNR was tested in a separate 1.7 mm isotropic voxel size acquisition of the full protocol, in which the MLP retained lower errors and less visually apparent noise. Finally, we demonstrate qualitative feasibility in two glioma patients scanned with the short protocol. We conclude that a simple MLP can provide high-quality QTI parameter estimates from short tensor-valued diffusion acquisitions. This enables five-minute, high-resolution QTI and may encourage further clinical studies of markers such as microscopic fractional anisotropy.

Low-field magnetic resonance imaging (MRI) is an affordable medical imaging technique used to assess the structural and functional features of internal and external tissues and organs. However, its clinical effectiveness is often constrained by prolonged scan times and reduced image quality due to compromised signal-to-noise ratios. Deep learning (DL) has emerged as an effective solution for reconstruction of undersampled MRI data. This study presents a comprehensive comparative evaluation of six U-Net variants, including U-Net, Attention U-Net, Residual U-Net, Recurrent-Residual (R2) U-Net, Residual-Attention U-Net (RAU-Net), and U-Net++, for image domain reconstruction of low-field (0.3 Tesla) MRI data from the M4RAW dataset. Root Sum of Squares (RSS) magnitude images were used for training and testing. The models were trained on 1024 MRI volumes comprising 18,432 image slices. Reconstruction performance was evaluated under Cartesian undersampling at acceleration factors of 2, 4, 8, and 16, and estimated different statistical indices including Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Normalized Mean Squared Error (NMSE) and Coefficient of Determination (R[2]). The results demonstrate that the vanilla U-Net performs well at lower acceleration factors, achieving average SSIM and PSNR values of 0.9303 and 39.18 dB, respectively, at an acceleration factor of 2. In contrast, more complex architectures, particularly U-Net++ and RAU-Net, exhibit superior performance at higher acceleration factors. In particular, U-Net++ achieved average SSIM values of 0.8754 and 0.8731, and PSNR values of 35.21 dB and 35.06 dB at acceleration factors of 8 and 16, respectively. These findings inform the development of efficient and accessible imaging solutions, promote the broader adoption of low-field MRI, and encourage further innovation in cost-effective medical imaging technologies.