Physics in medicine and biology最新文献

Objective.Cerebrovascular diseases are a major global health challenge due to their high morbidity and mortality rates. Accurate segmentation of cerebrovascular structures in TOF-MRA is crucial for accurate diagnosis and treatment planning. However, it remains difficult due to the variability in vessel morphology and the scarcity of annotations.Approach.In this paper, we propose BDD-CL, a boundary-aware and discrepancy-guided dynamic pseudo-labeling consistency learning framework for semi-supervised 3D TOF-MRA cerebrovascular segmentation. The framework is equipped with three carefully designed modules: (1) a boundary enhancement (BE) module that introduces shape constraints to improve vessel boundary delineation; (2) a shape-aware discrepancy (SAD) module that detects and refines prediction inconsistencies between networks, boosting robustness in regions with complex vessel morphology; and (3) a dynamic pseudo-label selection mechanism that adaptively delegates pseudo-label generation to the better-performing network, mitigating error propagation and improving label efficiency.Main results.Extensive experiments on COSTA and IXI datasets demonstrate that BDD-CL surpasses seven state-of-the-art semi-supervised methods in both quantitative and qualitative evaluations.Significance.These results highlight the framework's potential for label-efficient and reliable cerebrovascular segmentation in clinical practice. The code and model will be made publicly available athttps://github.com/nazikelsayed/Boundary-aware-and-discrepancy-guided-dynamic-pseudo-labeling-with-consistency-learning.

Objective: Surface electromyographic (sEMG) signals of the diaphragm provide a valuable physiological signal for real-time respiratory monitoring, particularly in clinical applications such as radiotherapy tracking and intensive care, where accurate estimation of respiratory motion is essential. However, these signals are often contaminated by electrocardiographic (ECG) interference. Traditional signal processing methods introduce certain delays while suppressing ECG artifacts and rely on linear assumptions for quantifying respiratory motion, limiting their real-time adaptability and accuracy in clinical applications. This study aims to develop a robust solution for real-time respiratory motion quantification form sEMG signals.

Approach: A cascaded deep learning framework was proposed which consisting of 1) a CNN-LSTM hybrid model that isolates respiratory sEMG components and 2) a multi-scale CNN with nonlinear feature abstraction for quantifying respiratory motion. sEMG and respiratory data from 45 subjects was acquired, with 20 subjects for training and 25 for validation. Cross-correlation analysis was performed to assess correlation coefficient between sEMG and respiratory signal.

Main results: The proposed method achieved superior correlation with abdominal pressure-derived respiration (Pearson's r = 0.949 ± 0.030) compared to gating (0.910 ± 0.046) and template subtraction (0.859 ± 0.081) using the same filtering post-processing technology. Notably, the proposed method demonstrated significantly higher correlation with reference signals without requiring any post-processing, highlighting its real-time processing capability in artifact suppression.

Significance: This study demonstrates that the proposed deep learning framework provides an efficient solution for high-fidelity artifact suppression and realtime respiratory monitoring in clinical settings.

Objective.Artificial intelligence (AI) can enable automation, improve treatment accuracy, allow for a more efficient workflow, and improve the cost-effectiveness of radiotherapy (RT). To implement AI in RT, clinicians have expressed a desire to understand the AI outputs. Explainable AI (XAI) methods have been put forward as a solution, but the multidisciplinary nature of RT complicates the application of trustworthy and understandable XAI methods. The objective of this review is to analyze XAI in the RT landscape and understand how XAI can best support the diverse user groups in RT by exploring challenges and opportunities with a critical lens.Approach. We performed a review of XAI in RT, evaluating how explanations are built, validated, and embedded across the RT workflow, with attention to XAI purposes, evaluation and validation, interpretability trade-offs, and RT's multidisciplinary context.Main results. XAI in RT serves five purposes: (1) knowledge discovery, (2) model verification, (3) model improvement, (4) clinical verification, and (5) clinical justification/actionability. Many studies favor interpretability but neglect fidelity and seldom include user-specific evaluation. Key challenges include stakeholder diversity, evaluation of XAI, cognitive bias, and causality; we also outline opportunities.Significance. By linking XAI purposes to RT tasks and highlighting challenges and opportunities, we provide actionable recommendations and a user-centric framework to guide the development, validation, and deployment of XAI in RT.

Objective.Ultrafast nonlinear ultrasound imaging of gas vesicles (GVs) promises high-sensitivity biomolecular visualization for applications such as targeted molecular imaging and real-time tracking of gene expression. However, separating GV-specific signals from tissue remains challenging due to tissue clutter and limitations of current methods, which require complex transmit schemes and suffer from incomplete tissue suppression. This study aims to develop and validate harmonic amplitude-modulated singular value decomposition (HAM-SVD), a novel technique that represents a shift from current GV imaging methods by exploiting the unique nonlinear pressure-dependence of the GV harmonic signature.Approach.HAM-SVD employs single-cycle plane waves transmitted at 9.6 MHz across five tilted angles at a pulse repetition frequency of 2500 Hz, under four duty cycles (DCs) with alternating polarity. Beamformed data are reshaped into a space-pressure Casorati matrix and decomposed via SVD. Tissue background is suppressed by discarding the first (weakly nonlinear tissue) and lowest (noise) singular modes, yielding images comprised solely of pressure-dependent second-harmonic GV signals. The method was validated through numerical simulations,in vitrophantom experiments, andin vivorat lower limb imaging.Main results.HAM-SVD achieved a signal-to-background ratio of 19.16 ± 1.63 dBin vivo, significantly outperforming pulse inversion (14.19 ± 1.41 dB) and amplitude modulation (AM-SVD) (15.79 ± 1.38 dB). Simulation and phantom studies demonstrated superior spatial coherence in singular vector decomposition and reduced nonlinear artifacts compared to AM-SVD. HAM-SVD enables wide-field, ultrafast imaging without complex transmit sequences while maintaining robust tissue clutter suppression across varying pressure levels.Significance.By combining harmonic imaging with AM-SVD's adaptive clutter filtering, HAM-SVD overcomes limitations of conventional nonlinear techniques, including depth restrictions in xAM and incomplete tissue cancellation in pulse inversion. This approach enhances molecular imaging specificity for GVs and holds translational potential for ultrasound localization microscopy of slow-flowing contrast agents and preclinical disease-targeted molecular imaging.

Objective.To evaluate the trade-offs among model resolution, anatomical fidelity, computational cost, and localization accuracy in electroencephalography (EEG) source imaging using personalized head models. Special emphasis was placed on segmentation-free modeling and sparse inverse analysis for rapid brain mapping.Approach. Forward problems were solved using a scalar-potential finite-difference method combined with orthogonal matching pursuit for inverse source localization. Tissue conductivity was directly estimated from magnetic resonance images using machine learning models, including CondNet and CondNet-TART, to generate personalized, continuous conductivity distributions without explicit segmentation. Localization accuracy and computational efficiency were compared across different model resolutions and against a finite-element method (FEM) using a five-tissue head model.Main results. Segmentation-free models achieved localization errors below 23 mm while maintaining stable accuracy even at coarser resolutions. Estimated current density distributions exhibited smooth transitions across tissue boundaries. Computational time was reduced by 87.9% compared with FEM at 2.0 mm resolution. Among segmentation-free models, CondNet-TART yielded the highest localization accuracy and stability.Significance.The proposed segmentation-free framework provides efficient and accurate EEG source localization with substantially reduced computational cost. These features support time-sensitive applications such as presurgical functional mapping and real-time neuroimaging.

Aortic dissection (AD) is a life-threatening cardiovascular emergency. Non-contrast-enhanced computed tomography (NCE-CT) could provide timely AD screening with fewer contraindications compared to CE-CT imaging. However, NCE-CT examinations lack distinctive imaging characteristics of AD, leading to high rates of missed diagnoses and misdiagnoses, and increased radiologist workload. In this paper, we propose a novel end-to-end multi-task framework for automated aortic segmentation and AD detection using NCE-CT images. The framework comprises three main components: a deformable feature extractor enhancing aorta tubular-feature attention, an adaptive geometric information extraction module to optimize feature sharing between segmentation and classification tasks via the transformer cross-attention mechanism, and a knowledge distillation module transferring diagnostic information from the CE-CT-based teacher model to the NCE-CT-based student model. Multi-center tests across 3 internal and 2 external centers demonstrated that our model outperformed existing methods both for segmenting the aorta and detecting AD. Specifically, for segmenting the aorta, our framework achieved dice of 0.928 and 0.909, Jaccard index (JI) of 0.867 and 0.858, mean intersection over union (MIoU) of 0.932 and 0.913, and frequency-weighted IoU (FWIoU) of 0.995 and 0.994, in internal and external testing datasets, respectively. For identifying AD patients from non-AD patients, our framework achieved accuracies of 0.911 and 0.840, sensitivities of 0.925 and 0.888, andF1-scores of 0.922 and 0.836, in internal and external testing datasets, respectively. Ablation experiment demonstrates the effectiveness of each module. The proposed model may serve as an effective diagnostic assistant for radiologists, acting as a 'second pair of eyes' to assist in AD screening using NCE-CT images.

Objective: Spectral CT data from Photon-Counting CT (PCCT) enables material decomposition. Mechanistic approaches such as Maximum Likelihood Estimation (MLE) are noise sensitive. Deep learning alternatives mitigate this issue, but their accuracy remains limited due to lack of incorporation of underlying physics principles and lack of ground truth data. This study aims to develop and validate a physics-informed deep-learning model, trained on validated simulated data, to decompose spectral CT images into density (rho) and effective atomic number (Zeff) maps. Methods: The training dataset included simulated abdominal PCCT scans from 32 human models with corresponding ground truth. The scans were obtained at two clinical dose levels, four detector energy thresholds, different iodinated contrast agent concentrations and reconstructed using three clinically-used kernels. A Generative Adversarial Network (GAN) was trained with and without a physics-informed regularization loss to estimate ρ and Zeff maps. Model performance was evaluated on 16 computational phantoms and validated on 6 clinical cases. A reader study was performed on 30 image slices to assess the comparative performance of ρ and Zeff maps to multi-rendered Virtual Monochromatic Images (VMIs) for assessing liver lesion conspicuity. Main Results: With physics-informed regularization, NRMSE of 1.29% and 0.68%, SSIM of 0.99 and 0.99, and PSNR of 29.8dB and 29.04dB were achieved. A maximum RMSE of 5.45% was achieved on clinical data. Reader study results showed ρ and Zeff images had higher conspicuity scores compared to VMIs (median: 4.52 vs. 4.13; 95% CIs: [4.19, 4.52] vs. [4.01, 4.31]). The study showed equivalent conspicuity between VMIs and material images within a ±0.5 margin, though the small sample limits generalization. Significance: This study demonstrates the feasibility of material decomposition using a physics-informed GAN model trained on realistic simulated data. The maps provided equivalent conspicuity under a clinically acceptable margin, with a significantly small number of images for interpretation. .

Multimodal images contain complementary information that is valuable for deep learning (DL)-based image segmentation. To enable effective multimodal feature learning and fusion for accurate segmentation, multimodal images usually need to be registered to achieve anatomical alignment.However, in clinical settings, multimodal image registration is often challenging. For instance, to reduce radiation exposure, CT scans usually have a smaller field of view (FoV) than MR, i.e., inconsistent anatomical content in CT and MR images, hindering accurate registration. Using such misaligned multimodal images, segmentation performance could be significantly degraded. This study aims to develop a DL-based multimodal image segmentation method that is capable of learning high-quality and strongly related image features from misaligned multimodal images without registration and produce accurate segmentation results comparable to that obtained with well-aligned multimodal images. In our method, a unified body space (UBS) module is presented, where image patches cropped from misaligned modalities are encoded to positions and projected into a unified body space, thereby largely mitigating the misalignment among multimodal images. Built upon the UBS module, a new spatial-attention is proposed and integrated into a multilevel feature fusion (MFF) module, where features learned from misaligned multimodal images are effectively fused at internal-, spatial-, and modal-levels, leading the segmentation of misaligned multimodal images to a high accuracy level. We validate our method on both public and inhouse multimodal image datasets containing 1472 patients. Experimental results demonstrate that our method outperforms state-of-the-art (SOTA) methods. The ablation study further confirms that the UBS modules can accurately project image patches from different modalities into the unified body space. Moreover, the internal-, spatial-, and modal-level feature fusion in the MFF module substantially enhances segmentation accuracy for misaligned multimodal images. Codes are available at https://github.com/BH-MICom/Patch2Space.

Saturation transfer MR fingerprinting (ST-MRF) is a quantitative molecular MRI method that simultaneously estimates parameters of free water, solute, and semisolid macromolecule protons. The accuracy of these quantifications is highly dependent on the choice of acquisition parameters, and thus, the optimization of the data acquisition schedule is crucial to improve acquisition efficiency and quantification accuracy. Herein, we developed a learning-based optimization framework for ST-MRF, incorporating a deep Bloch equation simulator as a surrogate model for the forward Bloch equation solver to enable rapid simulations. Notably, the deep Bloch equation simulator overcomes the non-differentiability of the original model by enabling gradient computation during backpropagation within the physics-informed optimization framework, thereby allowing iterative updates of the acquisition schedule to minimize quantification error. In addition, the proposed method estimated an accurate ∆B0 map with the inclusion of a minimal number of scans to address B0 inhomogeneity. B1 inhomogeneity was corrected by providing a relative B1 map as an input to the quantification network. We validated our approach using Bloch-McConnell equation-based digital phantoms and further evaluated the performance of the proposed optimized ST-MRF framework in in vivo experiments. Our results showed that the optimal ST-MRF schedule outperformed other data acquisition schedules with regard to quantification accuracy. In addition, we enhanced the in vivo quantitative maps by correcting motion artifacts and suppressing noise using self-supervised learning techniques. The optimal ST-MRF approach could generate accurate and reliable multi-tissue parameter maps within a clinically acceptable time.