IEEE Transactions on Medical Imaging

Voxel-level annotation for volumetric medical imaging is expensive and difficult to scale, which makes training highcapacity 3-D segmentation models challenging in practice. Transfer learning (TL) from large public datasets is a common remedy, but it can under-perform when the source domain differs from the target anatomy and acquisition characteristics, as is often the case for pulmonary nodules. In this work, we propose a masked autoencoder (MAE) pretraining-based approach to break the data efficiency wall of domain difference and present a focused empirical study of domain-specific self-supervised learning (SSL) for 3-D lung nodule segmentation. We evaluate two experimental settings: first, Masked Autoencoder (MAE) pretraining versus random initialization across representative baselines; second, MAE versus Decathlon TL for UNETR++ while testing whether MAE-based pretraining also benefits a CNN baseline (V-Net). MAE pretraining on target-domain CT volumes achieves a Dice Similarity Coefficient (DSC) of 0.307, outperforming random initialization (0.136) and Decathlon weights (0.257). In addition, MAE improves the stability of V-Net in a "low-data" regime (i.e., with "insufficiently labeled" data), increasing DSC from 0.010 to0.071. Overall, these results suggest that MAE-based pretraining can provide a practical and robust initialization strategy for volumetric segmentation when labeled data are limited.

Microbubble contrast-enhanced ultrasound (CEUS) relies on discriminating nonlinear bubble signals from linear tissue backscattering. While Singular Value Decomposition (SVD) filtering improves this discrimination, existing techniques often fail to retain the slowly-moving microbubble signals from static clutter. Here, we present a novel multi-stage singular value decomposition (MS-SVD) framework for ultrafast CEUS imaging. Our method employs plane-wave transmissions at multiple angles and acoustic pressure levels (implemented via duty-cycle modulation) and alternating transmit polarity. The beamformed data are then processed by three sequential SVD filters: (1) spatial-angular SVD to extract coherent signals across all transmit angles, (2) spatial-pressure SVD to separate linear fundamental and nonlinear harmonic components, and (3) spatiotemporal SVD to isolate moving microbubble echoes from tissue clutter. In in vitro flow phantoms and in vivo rat brain through a cranial window, MS-SVD dramatically improves microbubble detection compared to conventional SVD filtering, MS-SVD yields much stronger vascular contrast and suppresses tissue clutter to a greater extent. The resulting power-Doppler and super-resolution maps are notably cleaner and more complete: MS-SVD detects substantially more microbubble events in ULM, revealing finer vessel details and more accurate flow speeds. By capturing the full acoustic signature of microbubbles (both fundamental and harmonic), MS-SVD achieves higher contrast-to-noise and sensitivity in CEUS. These gains make it a powerful front-end for super-resolution ultrasound localization microscopy and other high-sensitivity microvascular imaging applications.

Three-dimensional photoacoustic tomography (3D-PAT) enables noninvasive structural and functional imaging with optical absorption contrast and ultrasonic detection depth. However, its spatial resolution is limited by acoustic diffraction, and incomplete detection geometry can substantially degrade image fidelity and quantitative accuracy. Here, we present a ULM-guided model-based reconstruction framework, termed 3D-PAULMprior that incorporates sub-diffraction vascular priors from concurrent ultrasound localization microscopy (ULM) into 3D photoacoustic reconstruction. The method uses weighted regional Laplacian regularization to integrate high-resolution vascular information into the inverse problem, thereby enhancing vascular sharpness, suppressing limited-view artifacts, and improving blood oxygen saturation estimation. We validated 3D-PAULMprior using numerical simulations, tissue-mimicking phantoms, and in vivo mouse brain imaging. Compared with conventional reconstruction, 3D- PAULMprior improved spatial resolution by over 50%, increased contrast-to-noise ratio by 261.2%, and enhanced structural similarity index by 24.6%. In vivo, 3D-PAULMprior recovered vascular structures that were poorly resolved or missing in conventional reconstructions and produced more spatially confined sO2 maps. These results establish 3D-PAULMprior as a robust multimodal reconstruction strategy for high-resolution structural and functional photoacoustic imaging.

Biomedical image segmentation is a fundamental problem in computational biomedicine that aims to precisely delineate anatomical and biological structures, tissue types, or pathological regions in biomedical images. Accurate segmentation is essential for interpretation, decision-making, and quantitative analysis across a wide range of biological and medical applications. Over the past decade, the field has undergone a profound paradigm shift, evolving from task-specific specialist models to universal foundation models. This review provides an in-depth analysis of the evolution, tracing how the limitations of local discriminative learning drove the transition toward transformer-based global modeling, and large-scale generative pre-training. To help navigate the diverse landscape of interaction paradigms, we introduce the first systematic taxonomy of promptable biomedical image segmentation, categorizing existing methods into six distinct types, enabling users to intuitively select appropriate prompting strategies based on visual demonstrations and quickly pinpoint relevant literature (Prompt Type Visualization). Beyond model architectures, we discuss parallel advancements in dataset development, evaluation protocols, and application-specific adaptations across radiology, pathology, and biology. Integrating these powerful foundation models with rigorous domain-specific adaptation has great potential to improve patient outcomes and healthcare efficiency. Finally, we highlight key challenges in trustworthiness and clinical integration that must be overcome to realize the potential of the next generation of biological and medical generalists.

Image-based liver Couinaud segmentation is designed to automatically provide the locations of suspicious objects in liver CT/MR images. Once achieved, the physicians will be guided to the target slice and area where the suspicious node is located. However, conventional algorithms trained primarily on healthy liver images often fail to generalize to Hepatocellular Carcinoma (HCC) cases due to pathological structural distortions. In this work, we propose a robust two-stage framework that integrates a 3D Unet with a 3D Anatomical Structure-Guided Graph Convolutional Network (3D GCN). This two-stage strategy effectively isolates the liver volume to eliminate structural noise from neighboring organs, such as the spleen, allowing the framework to focus exclusively on the complex 3D anatomical relationships among the eight segments. To ensure the topological consistency required for global spatial reasoning, we implement a standardized preprocessing pipeline that normalizes liver-only volumes to exactly 50 frames along the z-axis. By combining a lightweight 3D UNet backbone with the 3D GCN for refined boundary reasoning, our model demonstrates superior generalization performance on unseen clinical datasets, achieving a mean Dice score of 0.828 in blind testing. By releasing our code and pretrained weights, we aim to provide the first publicly available deep learning resource for robust Couinaud segmentation.

Viscoelastic characterization of tissue has significant diagnostic value in oncology, as tumor progression alters both elasticity and viscosity in ways that neither property alone can fully capture. Existing acoustic radiation force (ARF)-based methods such as Viscoelastic Response (VisR) ultrasound estimate relative elasticity and viscosity through per-A-line nonlinear model fitting, which is computationally intensive and requires auxiliary simulations to correct elasticity-dependent bias. This work presents VESTA (Machine Learning-Enabled Estimation of ViscoElastic Ratios from On-Axis Spatio-Temporal ARFI Features), a two-stage data-driven pipeline that predicts elasticity ratio (ER) and viscosity ratio (VR) directly from seven normalized ARFI displacement features at the A-line level, without model fitting or compensation. Stage~1 is an MLP classifier that detects inclusion boundaries from normalized peak displacement and negative peak velocity ratios; Stage~2 is a dilated Conv1D regression model that estimates ER and VR along the full axial sequence using the predicted mask alongside displacement features. The pipeline was trained on 500 simulated inclusion scenarios spanning three geometries, five focal depths, two F-numbers, and a broad range of material contrasts. In silico, mean predicted ER and VR were within 12\% of ground truth across all geometries, with performance best when ER and VR were moderate or decoupled. Experimental validation on a chicken breast phantom demonstrated plausible generalization to real tissue heterogeneity. Applied to an in vivo murine 4T1 breast cancer model, the pipeline tracked treatment-related attenuation of mechanical contrast in paclitaxel-treated tumors relative to controls over a 36-day imaging period, supporting its relevance for tumor monitoring.

The development of digital twins in medicine, i.e., virtual replicas of human organs, offers a promising path toward precision medicine by enabling interpretable, mechanistic, and actionable insights. In the brain, cerebrovascular twins support individualized modeling of hemodynamics and bio-transport, with broad applications. A major bottleneck, however, is the lack of robust methods to transform in vivo cerebrovascular images into simulation-ready cerebrovascular meshes or graphs. Here, we present CerebroVascular Imaging to Graph reconstruction (CVIG), a robust and multiscale framework for reconstructing whole brain cerebrovascular graphs from in vivo cerebrovascular images. CVIG integrates vessel vectorization, with tolerance to discontinuity in vessel structures, using a topology-guided assembly of vessel trees to generate cerebrovascular graphs from medical images. We demonstrate the ability of CVIG to generate vascular graphs with improved vascular coverage and topological correctness, the capability essential for high fidelity brain biophysical simulations. This work establishes a vascular graph framework for individualized modeling and analysis, providing a key foundation for digital twins of the human brain.

High bit-depth medical images preserve subtle intensity variations that are important for quantitative analysis and clinical interpretation, but their large dynamic range poses challenges for efficient compression. We propose a bit-plane-aware dual-stream compression framework for 16-bit medical images by separately modeling the most significant bit (MSB) and least significant bit (LSB) components. The MSB structural stream is encoded using JPEG coding with a Duplicate Segment Skipping (DSS) strategy to exploit spatial and segment-level redundancy, while the LSB detail stream is compressed using learned image compression to represent residual variations and fine-grained details. Experiments on four MRI and CT datasets show that the proposed method consistently outperforms representative traditional and learning-based codecs, achieving the lowest bit rate across all datasets. Meanwhile, it preserves high reconstruction fidelity. As a downstream application, we further demonstrate that the compressed bitstreams can be effectively integrated with DNA encoding and converted into sequences with favorable biochemical properties.

Constitutive models of myocardial mechanics form a cornerstone of personalized cardiac simulations and cardiac digital twins. Researchers traditionally prescribe these models a priori and calibrate them from ex vivo tissue experiments, even though tissue excision alters loading conditions, removes residual stresses, and eliminates important physiological interactions. Multimodal cardiac MRI now provides subject-specific ventricular geometry, deformation, and myocardial microstructure, yet current inverse approaches still rely on predefined constitutive laws. Here we present the first framework to discover constitutive models of passive myocardial mechanics directly from in vivo cardiac imaging data by embedding a constitutive artificial neural network within a nonlinear finite element model of ventricular filling. Using multimodal cardiac MRI that combines ventricular geometry, deformation, and microstructure from a representative healthy individual, the framework identifies sparse, mechanically admissible strain-energy functions without prescribing their form a priori. The best-performing model contains only two fiber- and two sheet-invariant terms, achieves a mean displacement error of 1.62 mm, and reduces the error of the widely used Guccione and Holzapfel models by 34.14% and 26.01%. The discovered models indicate that fiber- and sheet-related anisotropic mechanisms dominate the passive mechanical response during physiological ventricular filling. More broadly, this work establishes a non-invasive strategy for subject-specific constitutive discovery from cardiac imaging data and lays the foundation for personalized cardiac simulations and cardiac digital twins.

Accurate and robust classification of medical pathology images is pivotal for computer-aided diagnosis. However, the deployment of deep learning models in high-throughput clinical screening faces a fundamental challenge: the trade-off between diagnostic accuracy and computational efficiency. Current lightweight architectures, while reducing parameter complexity through grouped convolutions, often lead to cross-channel information isolation and diminished representational capacity. In this paper, we propose TetraFuse, a novel framework that systematically integrates features from four complementary domains: space, channel, statistics, and frequency. TetraFuse introduces a novel Cross-Channel Dynamic Aggregation (CCDA) paradigm that reconstructs global channel topology with negligible computational overhead, resolving the inter-group isolation issue. To balance perceptual fidelity and efficiency, we design a stage-aware local enhancement mechanism: Local Variance-Guided Enhancer (LVGE) is employed to filter out shallow-stage background noise, while High-Frequency Boundary Injection (HFBI) reinforces deep-stage pathological contours, preventing spatial over-smoothing. Experimental results on the COVID-19, ISIC 2018, and Kvasir datasets confirm that TetraFuse outperforms state-of-the-art (SOTA) methods. Notably, TetraFuse-Tiny achieves a transformative 91.53% reduction in FLOPs compared to ResNet50; on the Kvasir dataset, it achieved an accuracy of 0.926 and an AUC of 0.994 with only 0.345G FLOPs. By combining high representational power with minimal computational demand, TetraFuse offers a scalable solution for large-scale medical image analysis, especially in resource-constrained clinical environments.

In modeling complex probability distributions, normalizing flows provide exact-likelihood, bijective mappings between empirical data and tractable latent spaces. Building on this foundation, latent-aligned multiview normalizing (LAMNr) flows leverage these salient properties to learn shared latent subspaces across heterogeneous, multimodal datasets while simultaneously topologically unfolding the sampled data manifold into a continuous vector space. Formal latent-alignment constraints are used to model shared structural features separate from view-specific variations, coordinating latent projections into a shared geometric subspace. By applying this transformation in the context of biological imaging, the framework establishes a potential basis for a deep learning interpretation of foundational computational anatomy concepts, such as the population template, latent distances, and geodesic pairwise image interpolation. Additionally, the proposed framework enables closed-form conditional modeling for exact cross-view imputation and other latent space manipulations. Evaluations and illustrations on both imaging-derived phenotypes (IDPs) and multimodal MRI demonstrate the proposed framework and potential applications. To further motivate our work, we provide a robust and comprehensive, 2D and 3D open-source implementation in PyTorch, natively integrated with the ANTsX ecosystem (i.e., ANTsTorch) for efficient training and subsequent data transformation, manipulation, and analysis.

Hyperspectral imaging (HSI) has gained increasing attention for bone assessment because it captures rich wavelength dependent information associated with mineralised tissue. HSI provides detailed spectral information related to material composition, while 3D geometric information supports the analysis of surface morphology and structural detail. However, integrating spectral and geometric information remains challenging, particularly when conventional reconstruction pipelines depend on external pose estimation. To address this challenge, we propose BoNeRF-HS, a self-calibrated hyperspectral neural radiance field for 3D reconstruction. BoNeRF-HS jointly optimises camera intrinsics, volume density, and hyperspectral radiance, removing the need for COLMAP based poses. To improve spectral modelling, we incorporate a gated spectral adapter head that learns wavelength dependent radiance features for hyperspectral view synthesis. We evaluate BoNeRF-HS on a multi-view hyperspectral dataset containing mouse bone, trabecular bone analogue, and cortical bone analogue samples. Experimental results demonstrate that our framework achieves improved reconstruction quality, and better preservation of bone surface details compared with existing approaches.

Diffusion MRI (dMRI) enables noninvasive characterization of white-matter fiber orientations and tissue microstructure, but widely used approaches, such as constrained spherical deconvolution (CSD) and parametric multicompartment models, typically address these features separately. The diffusion tensor distribution (DTD) framework jointly represents fiber orientation and microstructure, but estimating DTD from finite, noisy measurements is severely ill-posed. Existing inversion methods either rely on nonnegativity constrained basis representations, which are challenging to sale to high-dimensional and high-resolution distributions, or use sampling-based approaches with limited reliability. We propose MaxEnt-DTD, a maximum-entropy algorithm for DTD estimation from finite and noisy dMRI data. By deriving the Lagrange dual formulation, we reformulate a constrained infinite-dimensional optimization problem into a finite-dimensional unconstrained convex optimization problem, substantially reducing the parameter space and enabling tractable whole-brain DTD estimation. We evaluate MaxEnt-DTD using both synthetic and in vivo data from the Human Connectome Project protocol and a second dataset using advanced B-tensor diffusion encoding. We compare MaxEnt-DTD-derived fiber orientation distributions with results from CSD and Monte-Carlo inversion methods, and assess fiber-specific microstructure measures and rotation-invariant metrics based on the cumulants of DTD. The results demonstrate that MaxEnt-DTD provides a reliable and efficient framework for joint fiber-orientation and microstructure analysis in dMRI.

Objective: This study aims to develop and validate a quantitative, depth-resolved anisotropy imaging framework that extends ARFI-based focal degree-of-anisotropy (DoA) estimation into two-dimensional mapping by modeling the depth-dependent relationship between shear modulus ratio (SMR) and peak displacement ratio (PDR). Methods: We propose APRIL (Adaptive Polynomial Regression for anisotropy Imaging via ARFI-induced DispLacements), a framework for quantitative, depth-resolved DoA imaging that adaptively selects polynomial regression or shape-preserving spline interpolation based on excitation PSF asymmetry. Training data were generated using an LS-DYNA3D + Field II simulation pipeline in homogeneous transversely isotropic media (SMR 0.9-4.9). Testing included shifted SMRs under varied acoustic conditions and three heterogeneous inclusion configurations (anisotropic inclusion in isotropic background and vice versa). Experimental validation was performed in an in-vivo murine tumor model over the time, ex-vivo chicken breast, and tissue-mimicking gelatin phantoms, using a Verasonics system with an L11-5v transducer. Results: APRIL achieved depth-resolved SMR prediction errors below 9% over 10-30 mm, with highest accuracy in the focal region (MAE 2.3%, RMSE < 0.1) and stable performance across PSF transition zones. In heterogeneous phantoms, it reconstructed anisotropy maps with SSIM up to 86% and MPE below 7%, accurately delineating inclusion boundaries. Under acoustic parameter variations, mean absolute errors remained below 10%, demonstrating robustness to system and tissue heterogeneity. Conclusion: APRIL enables robust, two-dimensional anisotropy imaging beyond focal estimates. Significance: The method provides a physically grounded and generalizable framework for clinically viable anisotropy biomarkers in muscle, tendon, kidney, tumor and breast tissues.

Quantifying cells within intact three-dimensional biological specimens remains a major challenge, as standard optical and histological techniques are inherently two-dimensional, destructive, or constrained by light scattering. Optical clearing can extend imaging depth but is time-consuming, disruptive to tissue integrity, and often incompatible with downstream analyses, limiting its practical use for routine three-dimensional quantification. X-ray computed tomography can overcome these limitations, yet conventional micro-CT lacks the soft-tissue contrast required for cellular-scale analysis. Here, we introduce an integrated imaging framework in which propagation-based phase-contrast X-ray CT is combined with volumetric nuclear segmentation to enable three-dimensional cell quantification in unstained volumetric tissue. We imaged ex vivo human liver tissue and segmented nuclei throughout the reconstructed volume, extracting quantitative nuclear metrics and spatial organisation metrics, including equivalent diameter, minor-to-major axis ratio and nearest-neighbour distance. We assessed measurement consistency across two non-overlapping volumes of interest and benchmark slice-resolved nuclear metrics against haematoxylin and eosin histology. The resulting high-contrast volumetric datasets preserve tissue context, allowing quantitative measurements to be interpreted alongside surrounding architecture and microstructure. Together, these results show that laboratory phase-contrast X-ray CT supports nucleibased volumetric cell quantification in intact unstained tissue and provides a framework for context-preserving quantitative analysis in three dimensions.

Blood clot formation and thrombolysis are dynamic biological processes that play central roles in hemostasis, thrombosis, and thrombolytic therapy. Monitoring clot evolution is challenging, as existing approaches often rely on specialized hardware or complex acquisition protocols. This study presents dense speed-of-sound shift imaging (DSI), a noninvasive ultrasound framework for spatiotemporal monitoring of coagulation and lysis from ultrasound image sequences acquired with a single imaging transducer. DSI estimates interframe displacements using dense optical flow and reconstructs slowness-shift maps by solving a regularized inverse problem, from which relative speed-of-sound (SoS) shifts are derived. Using this approach, we quantified spatially localized acoustic signatures of material solidification, clot formation, and enzymatic clot dissolution across systems of increasing biological complexity, including thermally gelling gelatin, fibrin clots, and porcine and human whole blood. DSI detected composition-dependent clot properties, with fibrinogen primarily affecting SoS shift magnitude and thrombin primarily affecting clotting kinetics. In both porcine and human whole blood, DSI tracked the full transition from rapid clot formation to tPA-mediated thrombolysis, revealing markedly slower lysis kinetics and species-dependent differences in clotting amplitude and stabilization time. Together, these results establish DSI as a simple, ultrasound-based platform for quantitative monitoring of coagulation, thrombolysis, and related biological material transitions.

This study presents the use of sub-micron nanobubbles (NBs) as contrast agents for ultrasound localization microscopy (ULM), a super-resolution imaging technique that visualizes microvascular structure and flow beyond the acoustic diffraction limit. While ULM has traditionally relied on micron-sized microbubbles (MBs), the reduced dimensions and prolonged circulation times of NBs make them attractive candidates for localization-based imaging. However, their weaker acoustic responses present significant challenges for reliable detection and tracking. To address this challenge, we developed the ULM Master GUI, an interactive framework for optimization of the complete ULM processing pipeline. Using custom ultrasound-compatible wall-less gelatin flow phantoms containing vessel-mimicking channels and bifurcations ranging from 100 to 500 m, we demonstrate that NB-based ULM achieves velocity reconstruction and flow partitioning measurements comparable to conventional MB-based ULM. Across all investigated geometries, NBs faithfully reproduced the underlying flow patterns and hemodynamic behavior despite their substantially reduced acoustic scattering. These findings establish the feasibility of NB-based ULM, expand the range of contrast agents available for localization microscopy, and provide a foundation for future super-resolution ultrasound imaging using nanoscale acoustic contrast agents. The ULM processing GUI is publicly available at https://github.com/grisha1998/ulm-super-resolution-toolbox.

Characterising human brain connectivity remains a major challenge in neuroscience. Multimodal datasets combining diffusion MRI with high-resolution microscopy in the same brain offer a unique link between macroscopic imaging and microstructural detail, but we lack tools to leverage these data to improve connectivity estimates for in vivo human imaging. We present a deep learning model that predicts high-resolution microscopy-informed fibre orientations from diffusion MRI. The model uses microscopy-derived three-dimensional fibre orientation maps as biologically grounded training targets. It is trained on a bespoke macaque dataset integrating in vivo MRI, postmortem MRI, and whole-brain microscopy, and then translated to in vivo human imaging. We use domain adaptation to predict fibre orientations from diverse MRI datasets: first to bridge differences in tissue state in the macaque (postmortem to in vivo), and then to generalise across species (macaque to human). Our method derives microscale-informed fibre architecture from diffusion MRI without requiring microscopy at inference. It leverages data that can easily be acquired only in animal models whilst generalising to in vivo human diffusion MRI with minimal acquisition requirements. The microscopy-informed fibre orientation distributions support biologically meaningful tractography, enhancing superficial white matter and cortical-subcortical pathway delineation for in vivo human data. More broadly, this work establishes a general framework for transferring microstructural information from microscopy to non-invasive imaging, enabling biologically informed mapping of brain connectivity.

Diffusion MRI (dMRI) tractography provides a non-invasive method for mapping whole-brain structural connectivity. However, its application is limited by substantial false-positive and false-negative connections. While deep learning based methods have shown promise in improving tractography, most rely on training data derived from conventional dMRI tractography, therefore inheriting the same limitations. Here, we introduce FiberLM, an attention-based Transformer model for mouse brain tractography. The model was trained using a whole-brain streamline dataset based on viral tracer data from the Allen Mouse Brain Connectivity Atlas (AMBCA), allowing the model to learn the properties of both local and long-range axonal trajectories through self-attention. FiberLM was applied to predict anatomically plausible axonal trajectories from ex vivo high-resolution mouse brain dMRI data. Quantitative evaluations demonstrated that FiberLM significantly reduced false-positive and false-negative connections, improved spatial agreement with tracer-defined pathways, and generated whole-brain connectomes that more closely approximated AMBCA results compared to conventional tractography. These findings suggest FiberLM as a potential tool for accurate reconstruction of mouse brain structural connectomics.

PurposeTo develop a data-efficient deep learning framework for rapid reconstruction of highly accelerated 4D arterial spin labeling (ASL) magnetic resonance angiography (MRA) with robust generalization using extremely limited acquired data, addressing the challenges of prolonged acquisition and reconstruction time. MethodsA simulation-driven, few-shot transfer learning approach was adopted by leveraging publicly available 3D time-of-flight (TOF)-MRA data to generate realistic multi-coil complex-valued pseudo-ASL k-space datasets for large-scale pre-training. A 3D unrolled reconstruction network was trained on this simulated data using a histogram-weighted loss and subsequently extended to 4D using lightweight temporal fusion modules. Fine-tuning was performed using only two experimentally acquired 4D ASL-MRA datasets. The method was evaluated on retrospectively and prospectively undersampled Cartesian 4D ASL-MRA data acquired at 3T and compared with compressed sensing (CS) and locally low-rank (LLR) reconstructions. ResultsThe proposed method achieved superior reconstruction quality compared with CS and LLR, with improved vessel depiction, particularly in distal branches, and enhanced temporal fidelity. Quantitative evaluation demonstrated higher vessel-masked peak signal-to-noise ratio and structural similarity index measure, along with increased error entropy, indicating reduced noise and structured artifacts. The initial pre-trained model already outperformed conventional methods, while additional 4D fine-tuning further improved performance. Robust reconstruction was demonstrated in prospectively undersampled data and multi-slab acquisitions, enabling large-coverage, time-resolved angiography within clinically feasible scan times (4-6 min). ConclusionsSimulation-driven pre-training combined with few-shot fine-tuning enables accurate and rapid reconstruction of highly accelerated 4D ASL-MRA in data-limited settings. The proposed framework provides a practical pathway toward clinically feasible, non-contrast dynamic cerebrovascular imaging.