Biomedical Physics & Engineering Express最新文献

Purpose. This study aims to synthesize CT from MR images for radiotherapy planning of head and neck tumor using an improved three-dimensional conditional generative adversarial network (3D cGAN) based on dual-attention modules.Methods. A total of 212 paired CT and T1-weighted MRI datasets are utilized, including 180 publicly available cases and 32 clinical cases from our hospital. Building upon the 3D cGAN framework, we implement structural modifications to the generator, discriminator, and loss functions. In particular, a lightweight dual-attention mechanism module is introduced to the generator based on 3D residual network. The model is trained on 186 datasets and evaluated on 26 test cases. Quantitative metrics including normalized cross-correlation (NCC), structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE) are calculated to assess the similarity between synthetic CT (sCT) and ground-truth CT images. A comparative analysis with U-Net, CycleGAN and basic 3D cGAN is conducted to validate performance improvements.Results. The proposed dual-attention enhanced 3D cGAN generates clinically acceptable sCT images across all 26 test cases. Quantitative evaluations demonstrate high accuracy with NCC of 97.06%, SSIM of 90.24%, PSNR of 28.23 ± 0.42, and MAE of 32.53 ± 2.49 HU. In quantitative comparison, the proposed dual-attention enhanced 3D cGAN approach outperforms U-Net, CycleGAN and the basic 3D cGAN across all metrics.Conclusion. This study proposes an improved dual-attention enhanced 3D cGAN algorithm. The method can rapidly and automatically generate sCT images from MR images for patients of head and neck tumor, which holds significant importance for implementing MR-only radiotherapy planning.

Stress is a prevalent and inherent phenomenon in people. It triggers the production of hormones that assist in managing the scenarios; nevertheless, chronic stress adversely impacts physical and mental health, which may result in detrimental effects such as depression, anxiety, digestive and heart diseases. Thus, early stress detection is essential to avoiding such negative effects. Addressing this challenge, this research attempted to create a Machine Learning (ML) based stress identification model utilizing two available datasets, namely K-EmoCon and WESAD, which acquired most discriminative signals for stress identification - Inter Beat Interval (IBI), Electro Dermal Activity (EDA) using the Empatica E4 wrist band. Time-Frequency features are extracted from these signals using Ensemble Empirical Mode Decomposition (EEMD) based on Hilbert Transform (HT). Instantaneous Frequency (IF) from IBI and EDA were fed as input to traditional ML models, showing a reduction of the computational power needed, which is especially relevant for setups with limited resources. Among those models, k-NN provides the highest accuracy of about 99.85% and an F1-score of 99.87%. Furthermore, real-time data acquired using a Fitbit smartwatch is also validated using the proposed approach, thereby improving the model's efficiency.

Objective.This study aimed to develop a miniaturized low-field thoracic magnetic stimulation (LF-ThMS) device to evaluate its effects on peripheral oxygen saturation (SpO₂) in healthy rats. This investigation was motivated by prior findings that LF-ThMS at 10.5 to 13.1 mT increased SpO₂in patients with COVID-19. However, its effect on healthy subjects remains unknown. To address this gap before extending research to healthy humans, we first examined its effects in healthy animal models.Approach.A miniature low-field thoracic magnetic stimulation (LF-ThMS) device, also referred to as a pulsed electromagnetic field (PEMF) system, was developed using two 30-turn coils made of 13-gauge magnet wire, encased in nylon sheaths. The coils were powered by a 30 V, 13 A DC source to generate magnetic pulses up to 13.1 mT. A custom control circuit, featuring an ATmega328P microcontroller, relays, and MOSFETs, regulated the pulse frequency and included a safety system to maintain coil temperatures below 38 °C. The device also featured a user interface for customizable and reproducible operation. Peripheral oxygen saturation (SpO₂) was monitored using a NONIN 750 pulse oximeter.Main results.The LF-ThMS device successfully generated magnetic flux densities of 10.5, 11.6, and 13.1 mT. However, when we compared SpO₂levels between the control condition (before LF-ThMS) and the SpO₂levels after the LF-ThMS at these intensities, we did not find a statistically significant difference. Significance.These results suggest that LF-ThMS may not affect SpO₂in healthy individuals, and the improvements observed in COVID-19 patients could be due to disease-specific mechanisms or other unknown factors, rather than a general physiological effect of LF-ThMS.

Sleep arousal, characterized by emergence of light sleep or partial wakefulness, often indicates underlying physical disorders, and its detection is crucial for effective patient treatment. While the detection of arousals using multiple signals can be effective, the dependencies on multiple electrodes impose burden on patients. To resolve this issue, some effective features estimated from single-lead electroencephalography (EEG) signals were proposed to detect sleep arousal. Normalized and filtered EEG signals were segmented into 7-s frames, and scalograms were estimated using continuous wavelet transform (CWT). Scalograms and local properties such as frequency, bandwidth, band energy, band energy ratio, maxima, and regularity were derived from the coefficients of CWT. Final classification features were generated using statistical analyses. The most effective features, estimated by correlation coefficients andp-values, were subjected to an artificial neural network to evaluate the performance of the features. The maximum classification performances (86.72% accuracy, 89.26% sensitivity, 86.55% specificity, and 94.87% AUC) were achieved with 100 features. However, sixty specific features were selected from a total of 182 classification features, yielding nearly the same performance as the maximum. Finally, only 14 features were identified as making a pronounced contribution to arousal detection. These findings highlighted the potential of a feature-efficient single-channel EEG-based approach for reliable sleep arousal detection. The proposed framework can be integrated into patient monitoring systems, such as apnea detection modules, to provide a more comprehensive tool for sleep disorder management.

Continuous and interpretable monitoring of masseter muscle activity is essential for the assessment of sleep bruxism (SB) and temporomandibular dysfunction (TMD). However, existing surface electromyography (sEMG) systems remain constrained by wired power supply, data-privacy concerns, and limited real-time specificity. To address these gaps, this study introduces a self-powered, edge-intelligent monitoring framework that combines poly(vinylidene fluoride) (PVDF)-based piezoelectric patches (BP-Patch) with a dual-branch lightweight neural network, the Depthwise Separable Convolutional Network with Efficient Channel Attention (DSC-AttNet). The network leverages depthwise separable convolution (DSC) to balance computational load and feature resolution, and incorporates an Efficient Channel Attention (ECA) module to enhance the discriminability between lateralised activations. After 8-bit quantisation, DSC-AttNet is deployed on an Arm Cortex-M4 microcontroller (MCU) while occupying only 80.7 KiB Flash and 72.8 KiB RAM, enabling real-time on-device inference across five physiological states (left/right bruxism, left/right chewing, and resting) with 94.75% classification accuracy and 63.6 ms average latency on data from 12 subjects. To support trustworthy AI-driven decision-making, Gradient-weighted Class Activation Mapping (Grad-CAM) and attention-based relevance analysis are employed to identify class-specific activation patterns across both time and frequency domains. These interpretable features further enable the derivation of clinically relevant indices such as nightly bruxism count, episode duration, and the Masseter Symmetry Index (MSI). By integrating bilateral self-powered sensing, resource-efficient edge inference, and quantitative interpretability within a fully on-device framework, this work lays the groundwork for long-term, home-based assessment and privacy-preserving intervention in masseter monitoring.

Leukodystrophies are a group of inherited disorders that predominantly and selectively affect the white matter of the central nervous system. Their overlapping clinical and imaging manifestations make a timely and accurate diagnosis challenging. In this study, brain MRI images from 115 patients with confirmed Leukodystrophy representing five major subtypes were analyzed. The imaging pipeline began with comprehensive pre-processing, which included tilt correction, noise reduction, skull stripping, brain segmentation, intensity normalization, and registration. This process ensured consistency throughout the dataset. Subsequently, two main classification strategies were investigated: (1) five traditional machine learning algorithms trained on four sets of handcrafted features extracted from the white matter and whole-brain regions, and (2) deep learning models using pre-trained convolutional neural networks fine-tuned on 3D MRI volumes. The CNN-based methods consistently outperformed traditional approaches, demonstrating a greater ability to learn complex hierarchical and spatial patterns. The InceptionV3 architecture achieved the highest performance on whole-brain images, with an accuracy of 93.41%, precision of 85.49%, recall of 83.95%, specificity of 95.77%, F1-score of 84.48%, and AUC of 89.86%. These findings indicate that machine learning-based approaches provide a reliable automated tool that can support neurologists in the differential diagnosis of Leukodystrophies, facilitating targeted confirmatory genetic testing and guiding patient management strategies.

Lung cancer remains one of the deadliest forms of cancer worldwide, making early and accurate pulmonary-nodule classification essential for improving patient prognosis. This study presents a robust ensemble-stacking framework that integrates Histogram of Oriented Gradients with advanced radiomic features to distinguish benign from malignant nodules. Experiments were conducted on the publicly available LIDC-IDRI dataset, which comprises of 1,018 thoracic computed tomography scans with expert-annotated nodules. Complementary feature sets capturing both local edge patterns and high-order texture and shape descriptors were extracted. On this feature set, Random Forest, Logistic Regression, and Support Vector Machine served as base learners. Through extensive hyperparameter tuning and class-balanced training, followed by 5-fold cross-validation, the proposed ensemble achieved an accuracy of 93.26%, a sensitivity of 90.76%, and an AUC-ROC of 97.96%, outperforming individual feature-only models and several recent state-of-the-art approaches. Furthermore, feature-importance analysis highlights the importance of morphological descriptors and the complementary value of gradient-based features. These results demonstrate that integrating different imaging biomarkers within an ensemble framework can significantly enhance diagnostic performance. Future work will extend this framework to multi-modal imaging and also to incorporate semi-supervised learning to reduce manual label dependence and improve the overall generalisation.

Pre-extracted lumen information of 3D vessel in medical images can effectively assist doctors in intraoperative navigation and postoperative evaluation, which has important clinical value. The main challenge faced by fully automatic 3D vessel segmentation comes from the imbalanced proportion of the vessels in medical image, which may lead to lost target. In this paper, a fully automatic 3D vessel segmentation method based on small-object-sensitive deep reinforcement learning, is presented. The region of target is firstly detected by the bounding box of a deep reinforcement learning (DRL) network, and then is segmented with a convolutional neural network (CNN). To better detect small vessel object, we have made three improvements to the existing DRL-based detection network: 1) A novel state with random receptive field expansion is applied to provide the agent with necessary information even if part of the target is lost. 2) A Recall-priority reward is presented to provide the most complete region for the next segmentation stage. 3) The dependency of vascular spatial positions between adjacent slices is utilized to correct the errors in detection stage, and the topological integrity of the obtained vascular structure is improved. The proposed method has been extensively validated on a challenging vessel dataset with 100 computed tomography angiography (CTA) scans. The segmentation accuracy of this method is Dice=93.75%, which outperforms the baseline and other automatic 3D vessel segmentation algorithms. This method has advantages in positioning accuracy, segmentation accuracy, and operational efficiency, and can be easily applied to clinical applications.

The recognition of the subject's emotional states is of great significance for achieving humanized services in many scenarios with human-computer interaction. Recently, identification of the emotional states based on electroencephalogram (EEG) has received increasing attention. However, due to the complexity of EEG signals, EEG-based emotion recognition is very challenging. In this research, a novel BrainEmoNet with learning-based framework is proposed to improve the emotion recognition accuracy from the perspective of the asymmetry of human brain functions. The BrainEmoNet consists of frequency-domain feature network (FFN), long-term dependent feature network (LDFN) and spatial characteristic analysis network (SCAN). The parallel FFN and LDFN are suggested to extract the frequency-domain and long-term dependent features of the information in each brain channel, respectively. Meanwhile, based on the working principle of the human brain, the SCAN with channel-spatial attention mechanism is proposed to focus on the high-value information channels with assigning adaptive weights and analyze the spatial characteristics of the frequency-domain and time-domain features. The feature analysis in the time-frequency-spatial perspective can fully explore the emotional information contained in EEG information. Experimental results on multi-modal DEAP dataset presents the competitive performances of the BrainEmoNet over the existing state-of-the-art models. In the subject-dependent experiments, the proposed model achieves identification accuracies of 86.77% and 82.14% in arousal and valence dimensions, respectively, compared to 75.53% and 72.83% in the subject-independent experiments. The proposed BrainEmoNet model in this research can be used as an auxiliary tool for the assessment or monitoring of emotions.

Dosimetric biomarkers, in terms of dosiomics features, play a crucial role in modeling radiotherapy and should be analyze d for their robustness and stability. This study aims to investigate how these dosiomics features will change over variations in treatment planning parameters. Different treatment plans were created by varying such parameters as field number, dose calculation algorithm, dose grid resolution, energy, monitor units, fraction, dose, field size, multileaf collimator, collimator angle, table angle, source-to-surface distance and source-to-axis distance, and wedge for a hypothetical tumor in the CIRS phantom CT scan. Dosiomics features were extracted with different segment sizes. The coefficient of variation (COV) was used to evaluate dosiomics feature changes with consider COV ≤ 5% as robust features. Our findings showed that many of the dosiomics features had significant variations due to changes in treatment parameters. First-order and gray-level co-occurrence matrix (GLCM) features were more stable (COV ≤ 5%) compared to others. Field and wedge changes had the most significant impact on features, while the dose calculation algorithm, dose, and MU changes had the lesser effects. Dosiomics features were vulnerable over changing treatment parameters and should always be reported. The GLCM features set was the most robust. Further studies are needed to identify robust dosiomics features for future biomarker discovery.