Biomedical Physics & Engineering Express最新文献

Background.Varicocele is a common cause of male infertility, with ultrasound (US) serving as the primary diagnostic tool. Current practice relies on manual, subjective measurements of the spermatic vein, which are time-consuming and lack reproducibility. Developing automated tools is hindered by scarce annotated data and intrinsic US challenges like low contrast and high noise.Obejectives.This study aimed to: (1) develop and validate an efficient semi-automated annotation workflow; (2) establish the first performance benchmark for automated spermatic vein segmentation using deep learning; (3) critically evaluate the efficacy of state-of-the-art and customised segmentation models for this specific task.Methods.We proposed a semi-automated pipeline using the Segment Anything Model (SAM) with clinician refinement. Using the resulting dataset, we conducted a comprehensive benchmark, evaluating a baseline U-Net, advanced models (U-Net++, Attention U-Net, and RPA-UNet), and a proposed U-Net with deep supervision (UNet-DS). All models were assessed via leave-one-patient-out cross-validation and statistical tests.Results.The 'SAM+clinician' workflow showed excellent agreement with expert annotation (Dice Similarity Coefficient(DSC) = 92.66%; Kappa = 91.92%). In segmentation, the baseline U-Net achieved a mean DSC of 61.33%. Only Attention U-Net showed a statistically significant improvement (p= 0.0391). UNet-DS attained the mean DSC (64.65%) but this was not statistically significant (p= 0.0781). All models plateaued in a narrow range (DSC: 61%-65%), far below performance in mature US segmentation domains.Conclusion.This work validates an efficient semi-automated annotation solution and establishes the first performance benchmark for this task. Results reveal a distinct performance ceiling, indicating the primary barrier is the inherent data limitations, not model architecture. Future breakthroughs require a shift towards bespoke, physics-informed algorithms rather than applying generic deep learning models.

Accurate classification of electrocardiogram (ECG) signals is essential for automated arrhythmia detection and clinical decision support. Existing deep learning methods still struggle to jointly characterize morphological patterns, multi-lead interactions, and temporal dependencies, leading to limited representation of waveform details, rhythm dynamics, and class boundary separability. To address these challenges, we propose a NeuroCardioSense (NCS) framework, built upon a convolutional neural network (CNN) backbone, comprising a base model, NeuroCardioSenseNet (NCSN), and an enhanced variant,NeuroCardioSenseNet-Fusion (NCSNF). NCSN constructs a novel Time-Aware Gated Convolution (TAG-Conv) layer together with a Time-Aware Gating Mechanism (TAGM), which adaptively modulate convolutional filters and cross-lead feature contributions based on local temporal context and channel energy distributions. This design enables joint modeling of morphological features and short-range temporal dynamics while reinforcing inter-lead coherence. Building upon NCSN, NCSNF incorporates a Time-Fuzzy Integration Module (TFIM) that constructs a learnable fuzzy subspace by jointly encoding features and membership degrees, effectively mitigating class boundary ambiguity and improving discriminability in limited-sample conditions. Extensive experiments on the MIT-BIH Arrhythmia Database demonstrate the superiority of the NCS framework. NCSN achieves 98.77% intra-patient and 87.82% inter-patient accuracy, while NCSNF further improves performance to 99.16% and 90.85%, respectively, outperforming existing baseline methods.

Maintaining optimal health and preventing diabetes-related complications requires accurate and timely monitoring of blood glucose levels. In line with this, the present study focuses on developing an affordable, reliable, and precise Point-of-Care (POC) diagnostic platform for glucose detection by integrating microfluidic and colorimetric principles. The system employs a custom-fabricated microfluidic chip designed to facilitate efficient enzymatic color reactions using only ~20 μl of sample per microwell, achieving complete color development within 3-4 min. This chip is housed inside a compact, USB-powered 3D-printed imaging module equipped with a high-resolution fixed-focus camera, enabling consistent control over imaging parameters such as focal distance, camera alignment, and illumination conditions. The overall workflow is optimized for seamless compatibility with embedded systems or laptops, eliminating the dependency on smartphones or external calibration tools and making the setup well-suited for real-time diagnostic use in POC environments. A total of 1280 images, representing 16 glucose concentration levels ranging from 50 to 200 mg dl^-1, were captured under standardized conditions, labelled according to known concentrations, and processed through uniform preprocessing steps. Engineered image features extracted from the preprocesses images were then analysed using supervised machine learning models, including Random Forest, Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and a Feedforward Neural Network, to establish a robust predictive framework capable of delivering fast, consistent, and accurate glucose estimation for practical healthcare applications. Among the evaluated models, the Random Forest (RF) classifier achieved the highest cross-validation precision of 98% and an exceptional specificity approaching 100%. This clearly describes its ability to distinguish between different glucose concentration levels. Further, the confusion matrix and the ROC curve analysis have validated the model's reliability, with very minimal chances of misclassifications and a high mean AUC value of around 1. These results ensure the potential of the image-based glucose concentration estimation as a cost effective and a reliable, scalable solution for real time monitoring in various medical related industries.

Low-dose computed tomography (LDCT) is crucial for reducing radiation exposure in medical imaging, but it often yields noisy images with artifacts that compromise diagnostic accuracy. Recently, Transformer-based models have shown great potential for LDCT denoising by modeling long-range dependencies and global context. However, standard Transformers incur prohibitive computational costs when applied to high-resolution medical images. To address this challenge, we propose a novel pure Transformer architecture for LDCT image restoration, designed within a hierarchical U-Net framework. The core of our innovation is the integration of an agent attention mechanism into a variable shifted-window design. This agent attention module efficiently approximates global self-attention by using a small set of agent tokens to aggregate and broadcast global contextual information, thereby achieving a global receptive field with only linear computational complexity. By embedding this mechanism within a multi-scale U-Net structure, our model effectively captures both fine-grained local details and long-range structural dependencies without sacrificing computational efficiency. Comprehensive experiments on a public LDCT dataset demonstrate that our method achieves state-of-the-art performance, outperforming existing approaches in both quantitative metrics and qualitative visual comparisons.

Skin cancer and dermatological diseases are among the most prevalent global health conditions, where early and accurate diagnosis is critical for improving patient outcomes. Although deep learning models have achieved strong performance in dermoscopic image classification, many existing approaches primarily rely on visual features and make limited use of complementary clinical metadata and language-based context routinely considered by dermatologists. Recent vision-language models (VLMs), including medical-domain adaptations such as MedCLIP, have begun to show promise in dermatology; however, their integration with structured clinical metadata and the impact of different multimodal fusion strategies have not been systematically analyzed. In this work, we address the binary skin lesion classification problem by conducting a structured evaluation of MedCLIP-based multimodal embeddings combined with classical machine learning and neural classifiers. Image-text representations are extracted using MedCLIP and fused with patient metadata through early and attention-based fusion mechanisms, followed by multilayer perceptron (MLP) and ensemble classifiers. Experiments are performed on a curated subset of the ISIC 2024 dataset comprising 1,600 training and 400 test dermoscopic images with associated metadata. The proposed multimodal approach achieves an accuracy of 96% (95.7% exact) with AUROC = 0.987, outperforming unimodal baselines and demonstrating the complementary value of language and metadata for skin lesion diagnosis. This study provides a comprehensive analysis of MedCLIP-based multimodal learning in dermatology and highlights the importance of fusion design in vision-language-metadata systems for computer-aided diagnosis.

Gold nanoparticle (GNP) enhanced radiosensitization was studied across three tumour cells lines that vary in radiosensitivity for a 2.5 MV photon beam. The intrinsic sensitivity of the cell lines is studied for the effects of GNPs. Three cell lines which exhibit differences in radiation response: prostate adenocarcinoma (PC-3), breast adenocarcinoma (MDA-MB-231), and cervical adenocarcinoma (HeLa) were used to determine thein vitrodose enhancement effects of GNPs combined with a 2.5 MV photon beam. Cells were incubated with 20 μg ml^-1GNPs for 24 h, and any extracellular GNPs were washed out prior to each assay. The cellular uptake of GNPs was assessed with inductively coupled plasma optical emission spectroscopy (ICP-OES). Clonogenic assays were conducted to assess cell viability after irradiation. The biological damage was assessed through DNA damage using terminal deoxynucleotidyl transferase dUTP nick end labeling assay (TUNEL). The production of reactive oxygen species (ROS) was assessed using CellRox assays. The enhancement factor when cells were irradiated in the presence of GNPs with 2.5 MV was 1.35 ± 0.46 for HeLa, 1.35 ± 0.11 for MDA-MB-231, and 0.99 ± 0.08 for PC-3 cells. On average, the level of DNA damage increased in MDA-MB-231 and HeLa cells when irradiated with 2.5 MV in the presence of GNPs. Increase in the ROS levels were detected in all cell lines when irradiated in the presence of GNPs. The enhancement effects with GNPs combined with a 2.5 MV photon beam were dependent on the cell line. The enhancement factor for HeLa and MDA-MB-231 supports further investigation of intermediate photon-energy beams in combination with GNPs. The combination of using a conventionally lower energy megavoltage beam with gold nanoparticles may become applicable in the clinical setting due to reduced skin dose and enhanced secondary electron production.

Introduction. The HARMONIC study (health effects of cardiac fluoroscopy and modern radiotherapy in paediatrics) investigates, among other objectives, the relationship between ionising radiation and cancer incidence in children treated by cardiac fluoroscopy (HARMONIC-Cardio). This requires estimation of organ doses for a large patient cohort. This article describes the development and validation of a framework for calculations of cardiac fluoroscopy doses, and its application to build a software tool for rapid cohort dosimetry.Materials and methods. Organ doses were calculated with MCNP6.2, a Monte Carlo particle transport code. Realistic, anthropomorphic phantoms representing patients of both sexes at specific ages (newborn, 1, 5, 10, 15 year old and adult) were used. A large range of technical parameters was covered in the simulations, including 11 primary beam angles and seven secondary angles, four levels of beam filtration, three tube voltages and three field sizes. The absorbed dose was computed for 32 organs and tissues. The calculated organ doses were normalised to the air kerma-area product (P_KA), a common modality-specific dose index, resulting in P_KA-to-organ-dose conversion coefficients.Results. Organ dose conversion coefficients were calculated for 22,176 exposure configurations and extended to a total of 2,667,600. A coefficient database for 12 organs of interest for radiation protection and effective dose, was embedded in HARMONIC-CardioDose, a software tool that enables the estimation of organ doses for any exposure scenario within the simulation range. The program is in the form of a Python script or an executable file (.exe), and uses an Excel document for inputting the calculation parameters.Conclusion. A framework for the calculation of cardiac fluoroscopy doses was developed and validated. It was used as the basis for HARMONIC-CardioDose, a rapid software tool for organ dose estimates for epidemiology studies. The tool is also freely available to the medical and research community for supporting patient dosimetry.

Segmentation of high-resolution CT enables the assessment of bone microstructure and provides relevant information in determining bone quality and fracture risk. Segmentation is typically performed with a global threshold value, expressed in bone mineral density (BMD). While consistent, a global threshold can lead to a poor representation of bone microstructure; amongst others, caused by altered bone physiology, image noise, image inhomogeneities, and partial volume effects. Adaptive threshold segmentations can preserve fine features, yet, it does not retain any quantitative information on BMD. We propose a new method for bone microstructure segmentation based on a Gaussian mixture model (GMM). This technique models the normalized image histogram as a bi-modal Gaussian distribution, reflecting bone and non-bone voxels, and gives an analytical description of the intensity threshold where probability of belonging to either class is equal. Next, the technique calculates a range of intensities around this threshold, where the tissue class is uncertain; only intensities within this range are adaptively segmented; others with a single threshold. Verification n was performed using simulated images witha prioriknown distributions. The GMM subsequentially reconstructed the input models. Next, high-resolution peripheral quantitative CT (HR-pQCT) and photon-counting CT (PCCT) images of cadaveric wrists were segmented, and segmentations were scored against reference segmentations from micro-CT. GMM segmentation accuracy was compared to adaptive and global thresholding. The optimal threshold from simulated images could accurately be determined, provided the bi-modal components did not accumulate into a single Gaussian. For HR-pQCT, full adaptive segmentation achieved the highest segmentation accuracy in this specific dataset (86.2 ± 3.7%), although the GMM led to comparable results (83.0 ± 2.3%). For PCCT, the GMM led to a slightly higher segmentation accuracy (77.9 ± 2.3%) than adaptive segmentation did (76.4 ± 2.4%). We conclude that this new method can segment bone microstructure with comparable accuracy as conventional techniques. Through the definition of uncertain intensities, the GMM method has the benefit that it provides the opportunity to tune the segmentation towards higher sensitivity or specificity, depending on the objective.

Breast cancer poses a significant threat to women's health, and early diagnosis is crucial for reducing mortality rates. Automatic breast tumor segmentation is important in medical image processing, but existing methods face challenges with breast pathology images due to sample scarcity, image degradation after data augmentation, and limitations in feature extraction. Traditional networks like U-Net often lose small lesions and edge details during downsampling and struggle with complex images and class imbalance. To address these issues, this study proposes DetailEdgeSkipBalance-Net (DESB-Net), an improved segmentation model based on U-Net. DESB-Net includes several innovations: Enhanced Detail-aware Multi-scale Re-parameterized Convolution (EDConv) for enhanced feature extraction, HistoEdge Focus Module (HEFM) for edge enhancement, Multi-Path Fusion Module (MPFM) for multi-scale feature fusion, and Binary Cross-Entropy Dice Loss (BD Loss) to balance class imbalance and boundary accuracy. These improvements significantly enhance the model's ability to capture small lesions and edge details, improve segmentation accuracy and robustness, and maintain high computational efficiency. On the UCSB dataset, DESB-Net achieved an mIoU of 79.53% and accuracy of 97.02%, outperforming U-Net by 6.5% and 1.89%, respectively, without increasing parameters or computational load. On the BCSS dataset, it achieved an mIoU of 63.4% and accuracy of 85.8%, surpassing U-Net by 4.2% and 2.6%. DESB-Net also outperformed mainstream models like DeepLabv3+, SegFormer, ResUNet, and Connected-UNets, demonstrating its effectiveness in breast pathology image segmentation. These results highlight the potential of DESB-Net to improve diagnostic accuracy and efficiency in clinical settings, making it a promising tool for early detection and treatment of breast cancer.

The purpose of this study is to address two questions regarding Gamma Knife Stereotactic Radiosurgery (GKSRS) planning: 1) Are there shape differences among disease types? and 2) Does plan quality vary by disease type? We considered meningioma, acoustic neuroma, and pituitary adenoma. For analysis, we exported a retrospective dataset from the treatment planning system (TPS) in DICOM format, including single fraction treatments with prescriptions at the 50% isodose line from 232 patients, completed between February 2018 and June 2023. The analysis included TPS-reported planning parameters, 3D radiomic shape features, and Gaussian Weighted Conformity Index (GWCI). The results were analyzed using two-sample t-tests and Pearson correlation coefficients. Although there were no statistically significant differences in volume, the cross-section of meningioma was the most circular, as defined by its major and minor axes. Pituitary adenoma exhibited the most flattened shape along its least axis. These results indicate that pituitary adenomas have distinct 3D shape characteristics. Pituitary adenoma required more shots, indicating they are more complex to plan. Acoustic neuromas had a similar number of shots to meningioma but showed better selectivity, implying it was easier to achieve planning guidelines, particularly for coverage. The normalized beam-on time for meningiomas was the shortest and the GWCI of acoustic neuromas was higher than that of the other two diseases, both statistically significant. A weak correlation between normalized BOT and the number of shots was found, suggesting that other factors beyond target shape influence plan complexity. Based on this study, shape differences exist among the considered diseases. Plan quality also varies by disease type. Pituitary adenomas are complex to plan, acoustic neuromas have better selectivity, and meningiomas have the shortest beam-on time. Factors beyond shape can also influence plan complexity.