Biomedical Physics & Engineering Express最新文献

Objective. To validate a respiratory motion model that uses real-time electromagnetic (EM) surface tracking acquired concurrently with time-resolved multi-cycle 4D MRI (TRMC-MRI) to estimate respiration-induced changes within the entire irradiated volume.Approach. Four volunteer participants with no self-reported history of lung cancer or other pulmonary disease underwent TRMC-MRI using a golden-angle stack-of-stars 3D GRE sequence while breathing freely for 2 min. Concurrently, real-time thoracoabdominal surface motion was recorded using four MR-compatible electromagnetic (EM) sensors (EndoScout II). Each MR volume was temporally aligned with corresponding EM data, resulting in 2,000 paired samples per participant. Deformation vector fields (DVFs) were generated through deformable image registration between a selected reference volume and all subsequent volumes. To capture temporal anatomical changes, additional DVFs were computed via consecutive volume-to-volume registration (e.g., volume 2 to 1, 3 to 2, and so on). Two machine learning models were developed to map surface motion to internal DVFs using dimensionality reduction: one employing Principal Component Analysis (PCA), and the other using Independent Component Analysis (ICA). Estimated DVFs were applied to the reference volume to reconstruct dynamic MR images, which were evaluated against ground truth using mutual information (MI) and an image-derived diaphragm profile-based root mean squared error (RMSE).Main results. Our preliminary results demonstrated that both PCA- and ICA-based models achieved comparable MI scores (mean 62%; one-way ANOVA, p > 0.05). Adaptive median filtering significantly improved MI to approximately 66% on average (one-way ANOVA, p < 0.001), outperforming no filtering across all participants (Tukey's HSD, p < 0.05). Diaphragm profile analysis showed close agreement with ground truth, with mean RMSE of 3.77-4.71 mm (SD: 1.07-1.63 mm).Significance. This proof-of-concept study demonstrates the feasibility of a non-invasive respiratory motion model derived from EM-based surface tracking combined with TRMC-MRI, with potential applications in MR-guided and conventional radiotherapy.

A flexible scintillator sheet made of Ce-doped Gd₃Al₂Ga₃O₁₂(GAGG) powder mixed with a two-part silicone adhesive enabled high-intensity imaging of proton beams on complex surfaces; however, its effectiveness for x-ray imaging remained unclear. To investigate this, we tested surface imaging with the GAGG sheet for both x-rays from computed tomography (CT) system and high-energy x-rays from linear accelerator (LINAC). For imaging of x-ray beam from CT system, we irradiated a flexible GAGG sheet with 140 kV x-rays, positioning it on a patient bed or curved surface of a cylindrical phantom. The resulting light emission on the GAGG sheet was captured by a CMOS camera. For LINAC x-ray imaging, we irradiated the GAGG sheet with 6 MV x-rays, setting it on a flat phantom or a human head phantom, and again captured the emitted light using a CMOS camera. In both CT and LINAC x-ray imaging, the light produced on the GAGG sheet was successfully captured at intervals as short as 100 ms, enabling real-time tracking of beams. The light output from CT x-rays was more than 50 times higher than that of plastic scintillator, while the light output from LINAC x-rays was 1.2 times higher. Given its adaptability to complex surfaces and high light emission for x-rays as well as real-time imaging capability, the flexible GAGG sheet shows potential for efficient surface beam imaging in both CT and LINAC x-ray applications.

Emotion recognition using electroencephalogram (EEG) signals is a growing focus in affective computing due to its wide-ranging applications in human-computer interaction. However, many existing studies process EEG signals as independent one-dimensional time series, overlooking their multidimensional structure and dynamic segment relationships. To address this, we propose a novel Hybrid Graph Attention Network (H-GAT) model that integrates Graph Attention Networks (GAT), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks. Our model effectively captures multi-dimensional dependencies in EEG data by modeling functional relationships between EEG segments, extracting local temporal patterns, and learning temporal dynamics. The EEG signals are transformed into a graph representation, where the adjacency matrix is generated from dynamic similarity measures computed by evaluating Pearson correlations between segment-wise feature vectors. This enables the GAT module to effectively model the functional relationships across EEG segments. Additionally, a CNN layer extracts local temporal patterns, while an LSTM captures the overall temporal dynamics. These features are then fused and passed through a fully connected layer for classification. Extensive experiments conducted on the SEED and DEAP datasets demonstrate the superiority of our model, achieving an average accuracy of 97.89 ± 0.50% for binary, 96.67 ± 0.72% for ternary, on SEED data, and 93.09 ± 1.02% for valence and 93.93 ± 1.14% for arousal on the DEAP dataset. Further, cross-dataset validation experiments confirm the model's strong generalization ability across heterogeneous EEG datasets. These results not only highlight the power of hybrid neural architectures but also demonstrate the model's transformative potential to progress emotion recognition, making it a robust and highly interpretable solution in the rapidly advancing field of affective computing.

Background:Inner-Speech (IS) based Brain-Computer Interfaces (BCIs) offer potential communication solutions for individuals with disabilities by decoding brain signals generated during speech imagination. While most IS-BCI systems rely on time-frequency EEG features, this study investigates functional connectivity-specifically, motif synchronization (MS)-to determine whether interactions between brain regions improve the discrimination of imagined words.Methods: We analyzed Electroencephalography (EEG) data from the 'Thinking Out Loud' dataset by Nietoet al2022, involving 10 participants mentally simulating four Spanish words ('Up,' 'Down,' 'Left,' 'Right'). Connectivity matrices were derived using the MS method, and node metrics (strength, PageRank) were calculated across 11 frequency bands. A two-stage classification framework was implemented: a support vector machine (SVM) ranked features by discriminatory power, followed by 5-fold cross-validation to identify optimal feature combinations.Results:The proposed approach achieved an average classification accuracy of 43.7%, comparable to previously reported results on the same dataset.Conclusions:Motif synchronization-based functional connectivity and graph metrics provide informative features for imagined speech BCIs by capturing cross-regional brain interactions. Further work is needed to improve robustness and generalization across sessions and subjects.

Non-invasive biosensing faces major challenges due to impedance mismatch at the skin interface, which causes significant signal reflection and limits power transmission through biological tissues. In this paper, we propose and evaluate a novel, subwavelength-thick impedance-matching metasurface (MTS) designed for direct skin contact to enhance electromagnetic (EM) wave transmission in the Ka-band. Our evaluation includes controlled laboratory experiments and a first-in-human study. Using an in-house developed benchtop system, we performed transmission measurements through aqueous glucose solutions and, most importantly, through the hands of six human volunteers. Our results show that the MTS significantly enhances signal transmission into human skin tissue, yielding an average improvement of up to 5 dB in the 36-37 GHz frequency range compared to the bare-skin condition, thereby improving sensitivity for analyte detection without increasing system size or power consumption. These findings demonstrate the potential of MTS-based impedance-matching layers as practical, integrable solutions to overcome key hardware limitations in wearable biomedical sensing devices. The study represents the first human investigation of an impedance-matching MTS designed to improve microwave signal penetration for non-invasive sensing applications.

Accurate electrode placement is critical for improving image fidelity in lung Electrical Impedance Tomography (EIT), yet current systems rely on simplified circular templates that neglect patient-specific anatomical variation. This paper presents a novel, low-cost pipeline that uses smartphone-based photogrammetry to generate individualized 3D torso reconstructions for boundary-aligned electrode placement. The method includes automated video frame extraction, mesh post-processing, interactive 2D boundary extraction, real-world anatomical scaling, and both manual and automatic electrode detection. We evaluate two photogrammetry pipelines-commercial (RealityCapture) and open-source (Meshroom + MeshLab)-across five subjects including a mannequin and four human participants. Results demonstrate sub-centimeter Mean Absolute Error (MAE 0.42-0.60 cm) and Mean Percentage Error (MPE 8.56-11.51%) in electrode placement accuracy. Repeatability analysis shows good consistency with Coefficient of Variation below 15% for MPE and 19% for MAE. The generated subject-specific finite element meshes achieve 98.79% accuracy in cross-sectional area compared to direct measurements. While the current implementation requires 15-30 minutes processing time and multiple software tools, it establishes a foundation for more precise and personalized bioimpedance imaging that could benefit both clinical EIT and broader applications in neurological and industrial domains.

Boron neutron capture therapy (BNCT) dose calculation often relies on fixed relative biological effectiveness (RBE) and compound biological effectiveness (CBE) values, despite their dependence on beam quality and tumor biology. We developed a microdosimetry-driven framework that predicts cell survival and RBE for BNCT by coupling PHITS lineal energy (T-SED) calculations with the microdosimetric kinetic (MK) model. MK parameters (α₀,β,r_d,y₀) were derived for BNCT relevant cell lines (U87 glioblastoma, NB1RG skin fibroblasts, SAS human squamous carcinoma, and SCC7 murine squamous carcinoma) using low-LET reference datasets curated in the PIDE database and irradiation conditions reproduced in PHITS. The derived parameters successfully reproducedin vitrosurvival curves for various charged particles across different energies, and when applied to neutron fields representative of BNCT systems (Kyoto University Reactor thermal neutron beam, cyclotron-based epithermal neutron source using a beryllium target, and linear accelerator system using a lithium target), the framework also reproduced measuredin vitrodata. Predicted RBE at 10% survival (RBE₁₀) agreed with measurements across cell lines and beam qualities, with only a slight deviation for SCC7 under the CICS spectrum and moderate deviations for SAS due to limited and heterogeneous low-LET datasets in PIDE. This method enables spectrum and cell-line specific estimation of biological effect, supporting replacement of fixed RBE/CBE with spectrum aware quantities to improve BNCT dose prescription and safety. The framework can also guide neutron-beam design by providing preliminary RBE estimates prior to construction of the moderator and beam shaping assembly. Incorporating intracellular boron microdistribution in future work is expected to refine CBE estimates and enhance biological accuracy in BNCT treatment planning. This framework provides a physics-based alternative to fixed RBE/CBE values.

The approach of drug repositioning, recognized as an efficient strategy to reduce costs and shorten development timelines in the pharmaceutical industry, has garnered significant interest among researchers. With vast datasets rich in valuable insights, machine learning models are not only widely applied but also continuously enhanced to improve accuracy and reliability. However, the process still faces notable challenges due to the complexity of relationships between drugs, diseases, and target proteins within biological networks. This paper focuses on optimizing the predictive performance of drug-disease relationship models by identifying potential connections through a belief network framework. The approach enables selection of the most effective graph path for each drug-disease pair, thus improving prediction accuracy by prioritizing critical biological interactions. Leveraging an ensemble of four classification methods along with a voting mechanism, the belief network ensures high adaptability and reliability across various datasets and disease types. Experimental results demonstrate that proposed analytical framework achieves not only superior performance but also high reliability in identifying potential drug-disease associations. Case studies within the paper further validate the model's capability to discover promising drug candidates for challenging diseases, underscoring the practical significance.

Magnetic nanoparticles (MNPs) have emerged as a powerful tool in cancer theranostics due to their unique size-dependent magnetic properties, surface functionalization capabilities, and responsiveness to external magnetic fields. This review outlines different types of MNPs, including those composed of pure metals, metal oxides, and metallic alloys, and highlights their size-dependent magnetic behavior, such as superparamagnetism and dynamic magnetizations. We also explore the critical role of surface modification strategies in enhancing MNPs' biocompatibility, colloidal stability, and functional versatility for targeted biomedical applications. The applications of MNPs in cancer therapy are discussed, with a focus on magnetic hyperthermia, drug and gene delivery, and a combination of various therapies. Additionally, we examine their cancer diagnostic roles in imaging techniques such as magnetic resonance imaging (MRI) and magnetic particle imaging (MPI), and emerging magnetic biosensing technologies such as giant magnetoresistance (GMR), magnetic tunnel junction (MTJ), magnetic particle spectroscopy (MPS), and nuclear magnetic resonance (NMR)-based platforms. These advances collectively establish MNPs as key components in the future of personalized cancer diagnosis and treatment.

Functional near-infrared spectroscopy (fNIRS) is a portable, non-invasive brain imaging method with growing applications in neurorehabilitation. However, signal variability, driven in part by differences in data processing pipelines, remains a major barrier to its clinical adoption. This study compares the robustness of two common processing approaches, General Linear Model (GLM) and Block Averaging (BA), in detecting cortical activation across task complexities. Eighteen neurotypical, healthy adults completed a simple hand grasp task and a more complex gross manual dexterity task while fNIRS data were recorded and analyzed using the BA and GLM pipelines. Results revealed significant effects of both pipeline and task complexity on oxygenated and deoxygenated hemoglobin amplitudes. BA produced significantly larger responses than GLM, and complex tasks elicited significantly greater activation than simple tasks. Notably, only the BA-Complex subgroup showed significant differences from all other conditions, suggesting BA more effectively detects task-related hemodynamic changes. These findings emphasize the need for careful analysis pipeline selection to reduce variability and enhance fNIRS reliability in neurorehabilitation research.