Biomedical Physics & Engineering Express最新文献

Electronic health records (EHRs) capture longitudinal multi-visit patient journeys but are difficult to analyze due to temporal irregularity, multimorbidity, and heterogeneous coding. This study introduces a temporal and comorbidity-aware trajectory representation that restructures admissions into ordered symbolic visit states while preserving diagnostic progression, secondary comorbidities, procedure categories, demographics, outcomes, and inter-visit intervals. These symbolic states are subsequently encoded as fixed-length numerical vectors suitable for computational analysis. Validation was conducted in two stages: Stage I assessed construction fidelity using coverage metrics, comorbidity preservation, diagnostic transition structures, and exact inter-visit gap encoding and Stage II assessed analytical utility through clustering experiments using different clustering approacheslike sequence similarity, Gaussian Mixture Models (GMM), and a temporal LSTM autoencoder (TS-LSTM). Proof of concept was done by encoding subset of patient cohorts from the MIMIC-IV database consisting of 2,280 patients with 8,849 admissions having complete primary diagnosis coverage and near-complete secondary coverage. Stage 1 assessment consisting of cohort-level coverage metrics confirmed that the transformation preserved essential clinical information and key properties of longitudinal EHRs. In Stage 2, clustering experiments validated the analytical utility of the representation across sequence-based, Gaussian mixture, and temporal LSTM autoencoder approaches. Ablation studies further demonstrated that both multimorbidity depth and inter-visit gap encoding are critical to maintaining cluster separability and temporal fidelity. The findings show that explicit encoding of comorbidity and timing improves interpretability and subgroup coherence. Although evaluated on a single dataset, the use of standardised ICD-10 EHR structure supports the assumption that the framework can generalise across healthcare settings; future work will incorporate multimodal data and external validation.

Research on deep learning for medical image segmentation has shifted from single-modality networks to multimodal data fusion. Updating the parameters of such deep learning models is crucial for accurate segmentation predictions. Although existing optimizers can perform global parameter updates, the fine-grained initialization of learning rates across different network hierarchies and its influence on segmentation performance has not been sufficiently explored. To address this, we conducted a series of experiments showing that the initialization of a differentiated learning rate across network layers directly affected the performance of medical image segmentation models. To determine the optimal initial learning rate for each network level, we summarized a general statistical relationship between early-stage training results and the model's final optimal performance. In this paper, we proposed a fine-grained learning rate configuration algorithm. To verify the effectiveness of the proposed algorithm, we evaluated 10 segmentation models on three benchmark datasets: the colon polyp segmentation dataset CVC-ClinicDB, the gastrointestinal polyp dataset Kvasir-SEG, and the breast tumor segmentation dataset BUSI. The models that achieved the most significant improvement in mIoU on these three datasets were H-vmunet, MSRUNet, and H-vmunet, with increases of 3.87%, 4.67%, and 6.22%, respectively. Additionally, we validated the generalization and transferability of the proposed algorithm using a thyroid nodule segmentation dataset and a skin lesion segmentation dataset. Finally, a series of analyses, including segmentation result analysis, feature map visualization, training process analysis, computational overhead analysis, and clinical relevance analysis, confirmed the effectiveness of the proposed method. The core code is publicly available athttps://github.com/Lambda-Wave/PaperCoreCode.

Purpose. Mammographic density is associated with the risk of developing breast cancer and can be predicted using deep learning methods. Model uncertainty estimates are not produced by standard regression approaches but would be valuable for clinical and research purposes. Our objective is to produce deep learning models with in-built uncertainty estimates without degrading predictive performance.Approach. We analysed data from over 150,000 mammogram images with associated continuous density scores from expert readers in the Predicting Risk Of Cancer At Screening (PROCAS) study. We re-designated the continuous density scores to 100 density classes then trained classification and ordinal deep learning models. Distributions and distribution-free methods were applied to extract predictions and uncertainties. A deep learning regression model was trained on the continuous density scores to act as a direct comparison.Results. The root mean squared error (RMSE) between expert assigned density labels and predictions of the standard regression model were 8.42 (8.34-8.51) while the RMSE for the classification and ordinal classification were 8.37 (8.28-8.46) and 8.44 (8.35-8.53) respectively. The average uncertainties produced by the models were higher when the density scores from pairs of expert readers density scores differ more, when different mammogram views of the same views are more variable, and when two separately trained models show higher variation.Conclusions. Using either a classification or ordinal approach we can produce model uncertainty estimates without loss of predictive performance.

Objective. Proton boron capture therapy (PBCT) is a novel approach that utilizes alpha particles generated through the proton induced capture reaction with¹¹B. Early studies reported substantial dose enhancements of 50%-96% near the Bragg peak, suggesting a promising therapeutic advantage. However, subsequent investigations have raised critical concerns regarding the practical feasibility of PBCT, primarily due to the relatively low reaction cross section in the Bragg peak region and the need for clinically unrealistic boron concentrations. The aim of this study is to evaluate Relative Biological Effectiveness (RBE) enhancement in PBCT using microdosimetric analysis across a wide range of boron concentrations.Approach. In the present work, we have employed Monte Carlo model using Particle and Heavy Ion Transport code System (PHITS) package combined with the Microdosimetric Kinetic Model to quantify both physical and biological dose enhancements at varying concentrations of¹¹B.Main Results. Microdosimetric analysis revealed that the total dose is dominated by protons, although alpha particles dominate in regions of higher linear energy deposition. The resulting RBE enhancement factors were 1.0011, 1.0080, and 1.1275 (for Human Salivary Gland (HSG) cell type) for 100, 1000, and 10,000 ppm boron concentrations, respectively. While the enhancements at lower concentrations are negligible, a modest increase is observed at very high boron levels.Significance. Based on the resulting RBE enhancement factors, it can be concluded that although alpha particles generated via thep+¹¹B → 3αreaction contribute high-Linear Energy Transfer (LET) energy at the cellular level, the overall biological dose enhancement remains rather minimal. These results indicate that under clinically achievable boron concentrations, the therapeutic benefit of PBCT may be limited.

In recent years, unsupervised domain adaptation (UDA) has emerged as a promising approach for constructing cross-subject emotion recognition models. However, most existing UDA methods do not fully exploit class information in the target domain, resulting in a relatively coarse-grained process of domain alignment and a higher risk of incorrect class matching. To address this issue, this paper proposes a novel adversarial training-based domain adaptation framework. The proposed method leverages emotion class prototypes to enhance intra-class correlation between the source and target feature distributions. Meanwhile, soft pseudo-labels generated by prototype clustering are utilized to further improve the inter-class discriminability within each domain. In order to enhance the robustness and quality of hard pseudo-labels in the target domain, a dual pseudo-labeling strategy is introduced. Finally, adversarial training is conducted to achieve a more fine-grained alignment of data distributions across domains. We conduct cross-subject and cross-session evaluations on the SEED and SEED-IV datasets, respectively. Experimental results demonstrate the effectiveness of our method and its advantages over several state-of-the-art UDA approaches. By introducing dual pseudo-labels, our study incorporates additional supervision, enabling a more refined domain adaptation process and significantly improving the generalization capability of EEG-based emotion recognition models.

Magnetic induction technology (MIT), as a non-contact and non-invasive sensing approach, has shown great potential for detecting brain lesions since it is unaffected by skull shielding. However, most MIT-based studies on intracerebral hemorrhage (ICH) have mainly focused on identifying the presence or estimating the volume of bleeding, while research on spatial localization has remained limited. In this study, a magnetic induction differential localization (MIDL) method was proposed to detect and localize ICH. A pair of symmetrically arranged detection coils was designed to sense the differential magnetic field perturbations caused by variations in the electrical conductivity and permittivity of brain tissues. The feasibility and response characteristics of the system were verified through numerical simulations and physical phantom experiments, followed byin vivovalidation on eight New Zealand white rabbits with unilateral induced hemorrhages. The real and imaginary components of the differential signals were analyzed to investigate their correlation with the side and volume of hemorrhage. Both simulations and phantom experiments demonstrated opposite variation trends of the real and imaginary components for left- and right-side hemorrhages. Animal experiments further confirmed that, after the injection of 1 ml of blood, the signal variation amplitudes significantly exceeded the baseline deviation (P < 0.05), exhibiting opposite directions of change between the two hemispheres. These results indicate that the proposed MIDL method can effectively distinguish the hemorrhage side and provide a theoretical and experimental foundation for non-invasive localization of intracerebral hemorrhage using MIT.

Conventional color flow processing is primarily optimized for qualitative visualization of flow dynamics, limiting its diagnostic use in regions where vascular structures are small relative to the ultrasound beamwidth. Leveraging the statistical properties of color flow data may provide a pathway toward quantitative discrimination between blood and tissue signals. This could enhance detection of vascular abnormalities, improve diagnostic accuracy, and support monitoring in diseases with small hemodynamic changes. Experimental data was obtained using a clinical GE LOGIQ 9 ultrasound system with a 10L linear array probe (3.75 MHz) positioned on an in-house made half-space flow phantom with the focus located at 3 cm depth. The simulation data obtained from Field II used a setup analogous to the experimental settings. Theoretical probability density function of ultrasound color flow power was derived using a gamma distribution. Shape parameters for blood and tissue were estimated using maximum likelihood estimation (MLE) in both simulation and experimental data. Color flow power was found to follow the gamma distribution in both simulation and experimental data. The estimated shape parameters aligned with theoretical predictions and distinguished between blood and tissue. Estimated shape parameters are less than or equal to 1 for tissue samples and greater than 1 for blood samples. This study presents a statistical modeling approach to enhance blood-tissue differentiation in color flow ultrasound, enabling blood characterization and perfusion quantification for improved detection and monitoring of vascular abnormalities.

Electroencephalogram (EEG)-based emotion recognition holds great potential in affective computing, mental health assessment, and human-computer interaction. However, EEG signals are non-stationary, noisy, and composed of multiple frequency bands, making direct feature learning from raw data particularly challenging. While end-to-end models alleviate the need for manual feature engineering, advancing the performance frontier of lightweight architectures remains a crucial and complex challenge for practical deployment. To address these issues, we propose LMSA-Net (Lightweight Multi-Scale Attention Network), a lightweight, interpretable, and end-to-end model that directly learns spatio-temporal features from raw EEG signals. The architecture integrates learnable channel weighting for adaptive spatial encoding, multi-scale temporal separable convolution for rhythm-specific feature extraction, and Sim Attention Module for parameter-free saliency enhancement. Our proposed LMSA-Net is evaluated on three benchmark datasets, SEED, SEED-IV, and DEAP, under subject-dependent protocols. It achieves top performance on SEED (65.53% accuracy), competitive results on SEED-IV (48.52% accuracy), and strong performance in arousal classification on DEAP, demonstrating good generalization. Ablation studies confirm the critical role of each proposed module. Frequency analysis reveals that our multi-scale temporal kernels inherently specialize in distinct EEG rhythms, validating their neurophysiological alignment. Lightweight design is evidenced by minimal parameters (7.64K) and low latency, ideal for edge deployment. Interpretability analysis further shows the model's focus on emotion-related brain regions. LMSA-Net thus delivers an efficient, interpretable, and high-performing solution. The code is available athttps://github.com/rhr0411/LMSA-Net.git.

This study aimed to assess a visual-tactile breath-hold (BH) workflow integrated with Elekta Unity's comprehensive motion management (CMM) system for gated MR-guided radiotherapy in situations where verbal coaching is impractical. A visual guidance program and a 3D-printed couch-mounted tactile pointer were implemented to instruct patients and stabilize voluntary BH. Two patients, one with pancreatic cancer and one with lung cancer, were treated using this workflow. Treatment beam gating was driven by CMM BH criteria, and audit log files from CMM-guided treatments were analyzed. Expected gating efficiencies were 40% for the pancreas case and 51.4% for the lung case, while measured efficiencies were 42.59 ± 2.56% and 54.95 ± 0.54%, respectively. The corresponding beam-on times were 14.75 ± 0.96 and 16.25 ± 0.50 min. The workflow reduced reliance on motion prediction for gating and mitigated frequent beam holds typically observed with free-breathing strategies, thereby decreasing dosimetric uncertainty. These findings indicate that a visual-tactile BH workflow on a 1.5 T MR-Linac is feasible and practical, supporting efficient gated delivery and reproducible breath-holds when verbal coaching is limited.

Improving medical image diagnosis performance relies on effectively representing features across various scales and accurately capturing local lesion characteristics and spatial context. While traditional convolutional neural networks are limited by fixed local receptive fields, hindering their ability to model global semantic relationships, transformers with self-attention mechanisms excel at capturing long-range contextual information but struggle with identifying small lesions. To overcome these challenges, this study introduces Hires-Diagnoser, a dual-stream framework for medical image diagnosis that supports multiple resolution levels. This framework combines ConvNeXt and Swin-Transformer branches in a parallel architecture. The ConvNeXt branch focuses on extracting local texture features through convolutions, while the Swin-Transformer branch captures global contextual dependencies using window-based self-attention. Additionally, a cross-modal correlation module (LCA) facilitates dynamic interaction and adaptive fusion of features across different resolutions. Experimental assessments on four datasets (RaabinWBC, Brain Tumor MRI, LC25000, and OCT-C8) demonstrated accuracy rates of 98.59%, 95.45%, 99.43%, and 95.23%, respectively, surpassing existing methods. By incorporating a cross-modal feature interaction mechanism, this framework achieves high performance and precise pathological interpretations, offering an effective solution for medical image diagnosis with certain practical implications.The source code of this proposal can be found at https://github.com/si-yuan20/hire-diagnoser.