CDP-KDNet: Curriculum-Guided Dynamic Pruning and Knowledge Distillation for Resource-Efficient Ultrasound Elastography.

To develop a deep learning-based auto-navigation technique for free-breathing golden-angle radial MRI named RANGR (Respiratory Auto-Navigator for Golden-angle Radial free-breathing abdominal MRI). RANGR computes a one-dimensional (1D) respiratory motion signal from 1D projections along the superior-inferior dimension (z) extracted directly from the acquired golden-angle stack-of-stars k-space data. The motion signal is used to retrospectively sort k-space data into different undersampled motion states, which are reconstructed using Movienet, a neural network developed for dynamic reconstruction. RANGR was trained using PCA (Principal Component Analysis) as a reference in cases where PCA was successful. The performance of RANGR is evaluated against PCA using a dynamic phantom with programmable motion waveforms on a 1.5 T MR-Linac system and free-breathing imaging of patients with abdominal tumors. In vivo image quality was qualitatively scored by body radiologists. For phantom studies, RANGR can track motion with less than 1 mm error with respect to the ground-truth. For in vivo studies, RANGR generalizes to cases where PCA failed due to limited liver coverage and/or the presence of high intensity regions outside the liver dome. RANGR outscores PCA on all qualitative image criteria as evaluated on a 5-point Likert scale by two expert body radiologists (p < 0.05). RANGR estimates motion faster than PCA on GPU (1.7 ± 0.3 vs. 168.7 ± 172.4 ms, p < 0.005). The total time from motion estimation to Movienet reconstruction is 2.91 ± 1.07 s. RANGR presents a robust auto-navigation solution based on deep learning for free-breathing MRI using golden-angle radial MRI acquisition.

Diffusion-weighted steady-state free precession (DW-SSFP) is a diffusion imaging sequence achieving high SNR efficiency. A key challenge for in vivo DW-SSFP is the sequence's severe motion sensitivity, currently limiting investigations to low or no motion regimes. Here we establish a framework to both (1) model and (2) correct for the impact of subject motion associated with the underlying magnetisation distribution of DW-SSFP. An extended phase graphs (EPG) representation of the 1D DW-SSFP signal was established incorporating a motion operator describing rigid body and pulsatile motion. The representation was validated using Monte Carlo simulations, and subsequently integrated into a data fitting routine for motion estimation and correction. The fitting routine was evaluated using both simulations and a voxelwise correction applied to in vivo experimental 2D low-resolution single-shot timeseries DW-SSFP data acquired in the human brain in three healthy volunteers, with a tensor reconstructed from the motion-corrected experimental DW-SSFP data. The proposed EPG-motion framework gives excellent agreement to complementary Monte Carlo simulations, demonstrating that diffusion coefficient estimation is robust over a range of motion and SNR regimes. Tensor estimates from the motion-corrected experimental DW-SSFP data give good visual agreement to complementary diffusion-weighted spin-echo (DW-SE) data acquired in the same subject, considerably reducing orientation-dependent motion-induced biases. Temporal information capturing the evolution of the DW-SSFP signal can be used to retrospectively (1) estimate subject motion and (2) reconstruct motion-corrected DW-SSFP data. Open-source software is provided, facilitating future investigations into the impact of subject-motion on DW-SSFP acquisitions.

Objective.Accurate lung motion estimation from 4D computed tomography (CT) is essential for image-guided radiotherapy and thoracic motion analysis, but remains challenging in vessel-rich regions where weak contrast, fine structures, and heterogeneous mechanics limit conventional registration. This work aims to develop an anatomy-informed finite-element digital volume correlation (FE-DVC) framework for accurate and mechanically plausible lung motion estimation.Approach.We propose a structure-tensor-guided heterogeneous anisotropic FE-DVC method. Structure tensors extracted from the reference CT image are used as heuristic anatomical priors to modulate element-wise regularization strength and align an effective anisotropic regularization frame with bronchovascular directions. A lung-specific multi-mesh strategy uses finer elements in vessel-containing regions and coarser elements in parenchyma. A practical workflow based on L-curve analysis and Jacobian-based deformation regularity is introduced to reduce empirical parameter tuning. The method was evaluated on three lung 4D-CT datasets and compared with Demons and pTV registration.Main results.The proposed method consistently reduced landmark target registration error across all datasets, with most errors concentrated below 2 mm and fewer large outliers. In a representative DIR-Lab case, normalized correlation residuals were reduced to 0.059 and 0.100 in axial and coronal views, compared with 0.172/0.256 for Demons and 0.088/0.138 for pTV. The recovered strain fields showed coherent vessel-oriented deformation, and boundary-driven FE simulations confirmed vessel-aligned strain concentrations only under heterogeneous anisotropic regularization. Singular value decomposition further showed that the first motion mode explained more than 90% of deformation energy and the first three modes exceeded 99%.Significance.This framework integrates CT-derived anatomical organization with mechanically interpretable regularization, improving vessel-scale lung motion recovery and supporting future strain-based biomarkers and pulmonary structure-function modeling.

MR imaging (MRI) plays an increasing role at different stages in the radiation oncology workflow, including tumor detection and delineation, the prediction of margins, fiducial marker detection, the estimation of organ and tumor motion, biofunctional avoidance and treatment response assessment using quantitative imaging biomarkers. Where conventional anatomical contrast-weighted MRI scans are a first-line strategy used for tumor detection and delineation, quantitative biofunctional MRI can provide predictors of tumor response or regional information on functional organs at risk. Unfortunately, the implementation of advanced quantitative MRI in radiation oncology may be prone to bias and uncertainties as a result of an over-reliance on vendor provided pulse sequences, image processing and analysis methods that have not been designed in the light of robust imaging biomarkers for radiotherapy guidance. Bias in quantitative MRI (qMRI) measurements can originate from the implementation of oversimplified physical models, unrecognized image artifacts and poor parametric fitting methods. Moreover, inconsistency in the use of imaging parameters and image post-processing methods across different centers results in a large variability in reported quantitative parameters. Much bias slips under the radar of both the medical physicist and MRI technologists because of the intrinsic variability between different human subjects which may remain undetected even when the imaging protocol is optimized with commercial quality assurance (QA) phantoms. Adequate pulse sequence optimization, harmonization of protocols and QA is essential to safeguard the reliability and robustness of the MRI protocols and can provide objective measures of uncertainty in quantitative parametric MRI maps. The increasing introduction of compressed sensing and data driven approaches in MRI image reconstruction demands for a different approach in imaging QA. In this first part, in-house fabricated phantoms are discussed that have proven useful in the optimization and QA of MRI sequences, protocols and post-processing methods. Sources of measurement bias and uncertainties in quantitative MRI, including diffusion mapping,T1andT2mapping will be discussed in parts 2 and 3. Novel imaging QA methods using anthropomorphic phantoms are introduced and the discrepancy between conventional inline approaches and in-house developed offline processing is discussed.

We aim to address the technical limitations in 3D respiratory estimation and image reconstruction from ultra-sparse views, overcoming data-acquisition constraints that often hinder conventional deformable image registration and volumetric imaging in real-time clinical applications. &#xD;We propose a Latent Accelerated Diffusion framework for Deformation Estimation enabling Real-time volumetric imaging (LADDER) framework, which integrates: (1) a deformation network (VoxelMorph) that generates a patient specific baseline deformation vector field (DVF) from pre treatment imaging, and (2) a latent diffusion model (LDM)-based DIR model that estimates DVF scaling factors and residual corrects to generate intra-treatment real-time DVF and dynamic volumetric images. The LDM compresses the baseline DVF into a compact latent manifold, enabling fast, projection conditioned refinement guided by anatomical cues. A physics informed loss enforces anatomical regularity and consistency with measured projections. LADDER was trained on the Learn2Reg dataset and evaluated on 10 DIR Lab lung datasets . &#xD;Main Results: With a dual projection input, an end-inhale to end-exhale baseline DVF and a compression and down-sampling factor of 8, on 10 test cases, LADDER achieves a mean target registration error (TRE) of 0.87±0.33 mm, and high volumetric structure similarity (3D SSIM > 0.95) and low volumetric reconstruction error (3D NMSE < 0.006), while maintaining real-time inference (~0.11-0.12s). Further analysis shows the range of DVF span impacts deformation accuracy, and the dual-projection input improves deformation fidelity and reduced variability across breathing phases compared to the single-projection. &#xD;Significance: LADDER enables real time 3D motion estimation and volumetric reconstruction from ultra sparse X ray views by conditioning diffusion in a patient specific DVF latent space. Its submillimeter TRE and real time speed indicate strong potential for next generation motion management and image guidance, supporting on the fly anatomical modeling to enhance the safety and efficacy of lung SBRT.

Magnetic resonance imaging (MRI) is highly susceptible to patient motion due to its relatively long acquisition times and the fact that data are acquired sequentially in k-space. Even small patient movements introduce phase inconsistencies across measurements, leading to severe artifacts such as blurring, ghosting, and geometric distortions that can compromise diagnostic quality. Retrospective motion compensation remains challenging, particularly in accelerated acquisitions, due to the ill-posed nature of the joint reconstruction and motion estimation problem. In this work, we propose a unified Bayesian framework for motion-compensated 3D MRI that jointly estimates the anatomical image, rigid-body motion parameters, and coil sensitivity maps directly from motioncorrupted k-space data. Our approach integrates pretrained 3D complex-valued score-based diffusion models as expressive anatomical image priors within a physics-based forward model. Inference is performed by alternating diffusion posterior image updates with efficient proximal optimization steps for motion and coil sensitivity estimation, enabling fully unsupervised reconstruction without the need for paired motion-free training data. Experiments on simulated and real-motion brain MRI datasets demonstrate that the proposed method achieves improved image quality and motion robustness compared to state-of-the-art classical and learning-based motion correction techniques, particularly in the presence of severe motion and high acceleration.

Reliable cloud motion estimation and interpolation of cyclones through fractional order variational technique

To develop a data-driven technique, Scout-based Multi-Echo NAvigator (SMENA), for joint estimation of motion and B 0 inhomogeneity ( δ B 0 ) at a temporal resolution of ∼200 ms with minimal additional scan time for gradient-echo acquisition. SMENA consists of two key acquisition components: SMENA-scout and SMENA-nav. SMENA-scout is a rapid 3D 4-mm multi-echo acquisition completed in less than 8 seconds, providing images with matched contrast and phase at multiple echo times. SMENA-nav captures signal variations induced by motion and δ B 0 during the scan using compact multi-echo navigator trajectories (3.5 ms) embedded within each TR. Motion and δ B 0 maps were jointly estimated every ∼200 ms through a model-based optimization framework relating SMENA-scout to SMENA-nav. The estimation accuracy and correction performance of SMENA were evaluated in simulations and in vivo using multi-echo GRE and GRE-EPTI acquisitions. Multiple prospective motion experiments, including large continuous movement and deep breathing, were investigated. In both simulations and in vivo experiments, accurate motion and δ B 0 estimation were achieved. Compared with motion-only estimation, joint estimation reduced rotation and translation errors. Joint motion and δ B 0 correction resulted in substantial improvements in image quality, particularly at longer echo times, producing an NRMSE of 10.4% compared to 31.6% with motion-only correction. High-temporal-resolution tracking of motion and δ B 0 enabled improved reconstruction quality in scenarios involving continuous motion and deep breathing. SMENA enables high-temporal-resolution joint estimation of motion and δ B 0 with minimal additional acquisition cost, providing a practical solution for motion- and δ B 0 -robust MRI.

Low-frame-rate flat-panel detector and gantry design limit Cone-beam computed tomography (CBCT) scanning speed, increasing susceptibility to patient motion during prolonged scanning. Rigid motion invalidates the predetermined imaging geometry, leading to degraded image quality. To ensure diagnostic reliability and support downstream tasks, accurate motion estimation and subsequent motion-compensated reconstruction are essential. This work presents a general rigid motion correction framework for CBCT using differentiable 3D-2D registration. We introduce a Lie group manifold-constrained motion modeling approach that conforms to the definition of rigid transformation and analytically derive the gradient of the CBCT forward projection operator with respect to rigid motion parameters. Furthermore, we identify the spatial deviation relationship between motion-corrupted and ground truth images, proposing a motion estimation constraint to enhance convergence. Experimental comparisons with state-of-the-art methods demonstrate that our method achieves superior motion estimation and image restoration. It effectively recovers anatomical details and substantially reduces rigid motion artifacts, even under severe motion conditions.

CardioMorphNet: Cardiac motion prediction using a shape-guided Bayesian recurrent deep network.

Cardiac quantitative MRI (qMRI) is a powerful imaging technique for diagnosing pathologies such as diffuse myocardial fibrosis. One main challenge is cardiac motion, which requires synchronization of data acquisition with the heartbeat, leading to long scan times. We present a novel deep learning-based image registration method for cardiac qMRI that enables non-rigid motion correction of data acquired continuously over multiple cardiac cycles, thereby reducing scan times. Our method is a zero-shot approach that utilizes the physical qMRI signal model for accurate motion estimation. Non-rigid motion of dynamic images is estimated with a U-Net-based architecture. This exploits the intrinsic smoothness of cardiac motion, allowing sharing information between neighboring images. The approach is robust to undersampling artifacts, enabling motion estimation from dynamic images reconstructed from very few k-space data even without advanced image reconstruction methods. We evaluated the method for fast cardiac T1 mapping using a Golden radial sampling scheme on numerical simulations and in-vivo acquisitions. On numerical simulations, our method achieved a 61.64% improvement in T1 accuracy. On in-vivo data, our approach yielded a 45.13% improvement in sharpness of T1 maps, and temporal image alignment of motion-corrected dynamics improved on average by 11.78%. Our method enables accurate non-rigid motion correction of highly undersampled cardiac qMRI data obtained from continuously acquired data. As our method is individually optimized for each scan without the need for training on large datasets, it can easily be adapted to other cardiac qMRI approaches.

Abstract Recent advancements in brain-inspired complementary vision chips (CVS) with intensity, multi-bit temporal difference (TD) and spatial difference (SD) sensing capabilities offer a promising solution to the limitations of traditional image sensors by enabling high-speed, high-precision sensing with reduced bandwidth consumption. However, noise characterizations and denoising strategies for these sensors remain underexplored. In this study, leveraging a recently developed state-of-the-art CVS, Tianmouc, we present a theoretical analysis of its noise characteristics, revealing the main challenge for denoising: a distinctive distribution that varies with local illumination. Building on this analysis, we develop a suite of novel denoising algorithms named locally adaptive direction-aware filter (LADF). LADF implements multi-stage denoising algorithms consisting of preprocessing followed by an adaptive threshold filter that adjusts parameters locally to mitigate noise variability. Additionally, considering the distinct characteristics of SD, we develop a multi-directional and polarity-aware separation strategy, while for TD, we exploit the inherent time-space correlation between TD and SD to suppress noise further. To enable rigorous evaluation, we construct a large-scale paired dataset through a novel synthetic-real approach that combines accurately labeled synthetic images with real-world captured data. Experimental results demonstrate that LADF achieves an average signal-to-noise ratio (SNR) of 10.11 in SD, outperforming two baseline methods by factors of 1.54× and 2.73×, respectively, and an average SNR of 4.52 in TD, surpassing the baselines by 1.47× and 3.57×, respectively. Furthermore, our method reduces errors in motion estimation by 28.9%, and enhances the peak SNR in reconstruction by 3.35 dB, demonstrating its effectiveness in downstream tasks. Our algorithm establishes a new benchmark for CVS denoising and demonstrates significant potential to enhance the performance of application systems utilizing CVS.

Purpose. Automated myocardial strain analysis requires accurate segmentation and motion tracking, but deep learning methods trained on one dataset often fail when applied to images from different scanners or institutions. We investigate how general-purpose foundational models can be efficiently adapted for cardiac strain estimation, establishing the minimal data requirements for robust cross-domain deployment.Methods. We propose CardiacSAM, a lightweight adaptation of the Segment Anything Model (SAM) fine-tuned with Low-Rank Adaptation, for cardiac structure segmentation, paired with uniGradICON for myocardial motion estimation. We compute global strain in cardiac coordinates for both short-axis (SAX) and long-axis (LAX) views.Results. Validated on four public and one private dataset comprising over 900 subjects, CardiacSAM achieved mean left ventricle (LV) Dice scores of0.90±0.02using only 10 training studies per domain for SAX (50 for LAX). Motion estimation achieved median average endpoint error (AEPE) of 3.08 mm on 15 volunteers without known cardiac disease with expert landmark annotations. Strain measurements discriminated significantly between reference and disease groups across all datasets. Comparing SAX and LAX strain in 360 subjects showed moderate correlation for radial strain (r=0.64) but weak correlation for longitudinal strain (r=0.17), suggesting SAX and LAX capture complementary information about cardiac function.Conclusion. Parameter-efficient adaptation of foundational models enables automated cardiac strain assessment with high data efficiency and robust cross-dataset generalization.

This article introduces a distributed deep neural network (NN)-based adaptive control framework for cooperative object manipulation in human-robot teams with unknown agent dynamics by using three distinct multilayer NN observers (MNNOs). The first observer, termed the reference point estimator, enables each robotic agent to estimate the object's reference center using consensus-based learning, even without direct access to global reference trajectories. The second observer, referred to as the human force-to-trajectory estimator, uses human-applied forces to infer the intended position, velocity, and acceleration of the object, enabling real-time estimation of human intent. Together, these two observers allow distributed estimation of human-intended motion. In addition, a third observer, the distributed NN dynamics observer, is integrated into the control layer to simultaneously estimate the agent's own state and unknown system dynamics while incorporating the state vector of all other agents. Weight update laws for the multilayer NN observers are developed using singular value decomposition (SVD), enabling stable and efficient parameter tuning in multiagent settings. The framework combines the observer estimates with a distributed online multilayer actor-critic NN controller to compute Pareto game theoretic optimal effort that coordinates robot actions while considering neighborhood interactions. Safety is enforced via barrier Lyapunov functions (BLFs) formulated using Karush-Kuhn-Tucker (KKT) conditions, which dynamically adjust safety constraints based on both the agent's own state and its neighbor state vector. Simulation results demonstrate that the proposed approach achieves accurate intent estimation, robust control, and a 60% reduction in total cost compared to baseline methods.

Multi-attention-aware motion estimation for cardiac MR imaging based on a feature pyramid network

SLAM (Simultaneous Localisation and Mapping) is an important component in robotics, providing a map of an environment and enabling localisation and navigation. While 3D LiDAR odometry and mapping systems have advanced in recent years, producing accurate motion estimates and detailed 3D maps, high-quality 2D occupancy grid maps (OGMs) remain challenging to obtain in large, complex indoor environments. OGMs are often degraded by drifts in odometry, sensor artefacts, and partial observability, resulting in maps with fractured walls, double boundaries, and artefacts that limit readability for mapping-centric tasks such as floor plan creation. To address this, we propose Transformation & Translation Occupancy Grid Mapping (TT-OGM), a system-level pipeline that targets map fidelity. TT-OGM leverages 3D scan registration to stabilise 2D map construction via projection and standard occupancy updates, then applies a learned GAN-based refinement module as post-processing to remove artefacts, regularise structure, and complete small missing regions. To enable training at scale, we introduce an offline DRL-based data generation process that produces paired but weakly aligned erroneous/clean OGMs spanning diverse error modes and severities. We demonstrate TT-OGM in real-time on a building-scale dataset collected at Loughborough University and evaluate map fidelity against a registered floor-plan reference using mIoU, masked SSIM, and occupied-boundary F1. We additionally report localisation accuracy on S3Ev2 using translation ATE (RMSE) against Cartographer and SLAM Toolbox (Karto). Our results show that 3D registration improves baseline 2D map quality over standard 2D SLAM outputs, and that GAN refinement further increases structural consistency and boundary accuracy in our pipeline. Additional ablations on synthetic stress tests and qualitative transfer to unseen Radish sequences show that the refinement module consistently improves OGM readability under common noise, moderate drift, and clutter conditions. • System-level Transformation-Translation (TT) pipeline for generating 2D OGMs from 3D LiDAR registration/odometry. • Learned post-processing module that refines OGMs by reducing common mapping artefacts and improving boundary fidelity. • DRL-driven synthetic data generation producing paired degraded/clean OGMs with controllable error modes for training and stress testing.

Self-determining multiderivative-enhanced phase motion estimation for accurate visual vibration measurement

OMFlow: Optimizing optical flow via occlusion motion estimation

Segmented 3D Gradient and Spin Echo (GRASE) is commonly used in Arterial Spin Labeling (ASL) perfusion imaging. However, it is vulnerable to inter-shot motion, leading to subtraction errors that cannot be corrected. We developed a retrospective self-navigated inter-shot motion correction method for segmented 3D-GRASE ASL imaging with Controlled Aliasing in Parallel Imaging (CAIPI). Multiple shots, each uniformly covering k-space at distinct sample locations, allow a self-navigator image to be reconstructed using SENSE for each shot. Rigid-body motion estimation across the self-navigators is incorporated into a motion-compensated forward model for image reconstruction. To support self-navigation, two CAIPI-sampled segmented 3D-GRASE trajectories ensuring full k-space coverage were explored for point spread function profiles and g-factor effects. Our approach was evaluated against conventional inter-volume registration and a previously proposed method, alignedSENSE. Additionally, we compared tag-control interleaving strategies to assess the impact on motion robustness in five healthy volunteers with instructed head motion. With instructed moderate head motion, our method effectively reduced motion artifacts and outperformed conventional inter-volume correction by 12.3% in Pearson correlation coefficient, 4.5% in Structural Similarity Index Measure, and 40.1% in temporal SNR. It matched alignedSENSE performance while requiring only 20% of the computational time. All evaluated CAIPI sampling variants enabled robust motion correction, although tradeoffs were observed between through-plane blurring and SNR. The tag-control (T/C) inner loop acquisition yielded better motion robustness across quantitative metrics. Self-navigated inter-shot motion correction using CAIPI sampling and a T/C inner loop for segmented 3D-GRASE ASL can improve image quality and motion robustness.