Physical and Engineering Sciences in Medicine最新文献

Endoscopy is a procedure that allows examination of the gastrointestinal system, including the stomach, esophagus, large intestine, and duodenum, with the help of an endoscope. Processing of endoscopic images is important for early detection and treatment of gastrointestinal diseases. In this study, hybrid ConvViT was developed using CNN and ViT to increase the classification accuracy of pathologies in gastrointestinal endoscopic images. CNNs are well-suited for capturing local spatial features through hierarchical convolutions, making them highly effective in detecting fine-grained textures and edge patterns. These capabilities complement the ViT's global attention mechanism, which excels at modeling long-range dependencies in images. The motivation of this study is to increase the classification accuracy and reliability with the ConvViT model, which was developed by combining the practical features of CNN and ViT models, which are individually successful in different aspects of image processing. The ConvViT model was compared with VGG-16, ResNet-50, Inception-V3 and ViT. Comparable models were tested using a gastrointestinal endoscopic image dataset containing ulcers, polyps, inflammation, bleeding, and regular anatomical features. Experiments showed that ConvViT had better prediction performance than compared models, with 95.87% classification accuracy.

Electrocardiogram (ECG)-based arrhythmia classification is crucial for the early detection and diagnosis of cardiovascular diseases. However, the presence of noise in raw ECG signals presents significant challenges to classification performance. In this study, we propose a novel approach that combines discrete wavelet transform (DWT) for signal denoising with an Attention-Enhanced Convolutional Neural Network-Bidirectional Gated Recurrent Unit (CNN-BiGRU) model for arrhythmia classification. First, DWT is applied to eliminate noise while preserving essential morphological features of the ECG signals. To address class imbalance, the Borderline-SMOTE algorithm is applied to generate synthetic samples for minority classes. The preprocessed signals are then passed through a CNN for hierarchical feature extraction, followed by a BiGRU to capture temporal dependencies. An attention mechanism is integrated to emphasize the most informative regions of the signal, enhancing the model's discriminative capability. The proposed method was evaluated on the MIT-BIH arrhythmia database and achieved an accuracy of 99.22% across five arrhythmia categories, outperforming several existing methods. This approach provides an effective solution for automatic arrhythmia detection in clinical practice.

To determine the optimum value of head scatter correction factor ([Formula: see text]) used in Radcalc software. The head scatter factors for a selection of multi-leaf collimator fields were measured on a Varian TrueBeam Edge and TrueBeam linear accelerators using an ionisation chamber in a mini-phantom. Radcalc calculated the head scatter values for the same fields. Radcalc calculates head scatter as [Formula: see text]. The head scatter value from Radcalc was recorded when the value of [Formula: see text] was set to 0 or 1. An optimum value of [Formula: see text] was obtained by minimising the sum of the differences between measured and calculated. The optimum values of [Formula: see text], for each linear accelerator type, were applied to clinical patient volume modulated arc therapy calculations. Minimising the summed differences yielded optimum values of [Formula: see text] of 0.149 and 0.276 for the TrueBeam Edge and the TrueBeam datasets respectively. Applying these values to 100 clinical patient volume modulated arc therapy plans, for each linear accelerator type, reduced the mean difference between the primary calculation and the independent check from 0.55 ± 0.95% (µ ± σ) to -0.11 ± 0.85% for the TrueBeam Edge and from 0.79 ± 1.16% to 0.24 ± 0.90% for the TrueBeam plans compared to when a generic Sc value of 0.675 was used. Using optimal values of [Formula: see text]in Radcalc, determined by measurement, reduced the mean monitor unit difference when compared to the primary calculation of a treatment planning system compared to using the standard value of 0.675.

Undersampled magnetic resonance imaging (MRI) reconstruction reduces scanning time while preserving image quality, improving patient comfort and clinical efficiency. Current parallel reconstruction strategies leverage k-space and image domains information to improve feature extraction and accuracy. However, most existing dual-domain reconstruction methods rely on simplistic fusion strategies that ignore spatial feature variations, suffer from constrained receptive fields limiting complex anatomical structure modeling, and employ static frameworks lacking adaptability to the heterogeneous artifact profiles induced by diverse undersampling patterns. This paper introduces a Spatial Adaptive Fusion Network (SAFNet) for dual-domain undersampled MRI reconstruction. SAFNet comprises two parallel reconstruction branches. A Dynamic Perception Initialization Module (DPIM) in each encoder enriches receptive fields for multi-scale information capture. Spatial Adaptive Fusion Modules (SAFM) within each branch's decoder achieve pixel-wise adaptive fusion of dual-domain features and incorporate original magnitude information, ensuring faithful preservation of intensity details. The Weighted Shortcut Module (WSM) enables dynamic strategy adaptation by scaling shortcut connections to adaptively balance residual learning and direct reconstruction. Experiments demonstrate SAFNet's superior accuracy and adaptability over state-of-the-art methods, offering valuable insights for image reconstruction and multimodal information fusion.

Early prediction of brain metastases (BM) in epidermal growth factor receptor (EGFR) positive lung adenocarcinoma patients is critical for improving treatment strategies and prognosis. This study aimed to enhance BM risk prediction within two years for lung adenocarcinoma patients by using lung CT images and clinical data both derived from initial diagnosis. This study comprised 173 patients with EGFR positive lung adenocarcinoma who underwent diagnostic CT and was stratified into 93 patients with BM and 80 patients without BM. We extracted a total of 1334 radiomic features from each manually delineated primary pulmonary nodule. Least absolute shrinkage and selection operator (LASSO) method was applied to select the optimal image features. Subsequently, the clinical model, radiomic model and hybrid model were constructed employing logistic regression, random forest (RF), support vector machine (SVM), and light gradient boosting machine (LGBM) algorithms separately. Ultimately, the model was evaluated and interpreted utilizing the receiver operating characteristic (ROC) curve, decision curve analysis (DCA), and shapley additive explanations (SHAP). The hybrid model consistently exhibited superior predictive performance. Specifically, the logistic regression-based hybrid model exhibited the highest overall performance metrics, with an AUC of 0.94 (95% CI 0.81-0.99). This study demonstrates that the logistic regression-based hybrid model can effectively predict BM in EGFR positive lung adenocarcinoma patients at their initial diagnosis, aiding physicians in developing more accurate treatment plans.

Beam modeling for radiotherapy treatment planning systems (RTPS) can be performed using representative beam data (RBD) or direct measurements. However, RBD typically excludes output factor (OPF) measurements for fields smaller than 3 × 3 cm². The Eclipse treatment planning system addresses this limitation by incorporating measured OPF data for fields as small as 1 × 1 cm². Although existing studies have primarily examined the accuracy of small-field OPFs for plastic scintillator detectors, studies directly comparing the OPF values obtained through RBD modeling with and without OPF measurements for small field sizes are limited. Therefore, this study proposes a novel measurement approach using data averaged from an ion chamber and diode detector for small-field dosimetry to provide critical insights into the integration of OPFs for these small field sizes in RBD-based beam modeling. We systematically evaluated the impact of small-field OPF measurements on beam modeling accuracy by comparing three distinct approaches: (1) RBD-based modeling without small-field OPF data, (2) RBD-based modeling incorporating measured small-field OPF data, and (3) modeling based solely on measured data, with and without the inclusion of 1 × 1 cm² field sizes. In addition, we compared OPF values obtained from a W2 plastic scintillator detector with the averaged OPF values from a PinPoint 3D ion chamber and EDGE diode detector across multiple beam energies and flattening filter-free (FFF) configurations. Our analysis included field sizes ranging from 1 × 1 cm² to 40 × 40 cm². The results demonstrated that for square fields, OPF calculation differences between RBD modeling with and without measured data were < 1.5%, < 4.5%, and < 4.5% at 1 × 1 cm², and < 0.5%, < 1.5%, and < 1.5% at 2  ×  2  cm², respectively. The RBD group exhibited a trend in which the OPF difference increased with the expansion of the irradiation field size. Notably, the most significant variations between modeling approaches occurred along the upper jaw expansion direction in rectangular fields. This suggests that a thorough evaluation is necessary for modeling results with an OPF ≤  1 × 1 cm². This study highlights the advantages and disadvantages of beam modeling using measured OPF and RBD, providing valuable insights for future facilities that rely solely on RBD for beam modeling.

This study investigated the potential of combining baseline ¹⁸F-FDG PET tumor-to-liver ratio radiomics with demographic data, using machine learning, to classify lymphoma subtypes and differentiate between candidates for ABVD and R-CHOP therapy. Additionally, we assessed whether nodal radiomics alone is sufficient for treatment and subtype classification. We conducted a multi-center study involving 241 lymphoma patients, including 125 with Non-Hodgkin lymphoma (NHL) and 116 with Hodgkin lymphoma. Among these, 94 had high-grade NHL, whereas 110 had classical Hodgkin lymphoma. We utilized 107 radiomic features, along with demographic data, such as age, stage, gender, and weight, to develop predictive models for classifying lymphoma subtypes and selecting treatment regimens (ABVD vs. R-CHOP). Data harmonization was performed using ComBat, feature selection was done with SelectKBest, and three machine learning models (Logistic Regression, Random Forest, and XGBoost) were trained with hyperparameter tuning, followed by external validation. For the best model in each classifier on the external test, adding extra-nodal radiomic features improved performance for certain lymphoma subtypes. For NHL vs. HL, accuracy increased from 0.807-0.819, whereas NHL precision rose from 0.837-0.875. High-grade NHL precision improved notably from 0.821-0.962. In treatment classification, extra-nodal features boosted accuracy for R-CHOP from 0.783-0.839 and increased F1-scores for both R-CHOP and ABVD. This study demonstrated the promise of PET radiomics combined with demographic features for lymphoma classification and treatment decision-making. Overall, extra-nodal features enhanced high-grade NHL and treatment classification but had minimal impact on other lymphoma subtypes.

Nowadays, monitoring the health of elderly people at home or patients at the hospital on a regular basis is becoming necessary. Unfortunately, peer-to-peer treatment may require a longer time based on the availability of the doctors. In addition, it is practically impossible to go to hospitals for health checkups almost every day of the week. Hence, this research proposes an idea that can automate these processes without decreasing efficiency and reducing manual labor by integrating a healthcare system with the cyber layer to execute the automation processes. Previous text and image recognition studies used different machine learning and deep learning algorithms. However, in this study, an optical character recognition method ‛YOLO V8' is used, which provides a faster detection speed than other methods. The target was to retrofit biomedical devices such as blood pressure monitoring machines, digital thermometers, etc. using image processing techniques. To train the'YOLOv8' model, we have utilized two distinct image datasets that we have developed. The model showed an accuracy of 99.5% in detecting areas of concern on medical devices. Later, for recognition of values of different parameters from those devices a Convolutional Neural Network model is used, which confirms real-time validation employing 1000 images from different medical equipment. An accuracy of 99.7% has been achieved using this method. In the future, other medical devices such as heart rate monitors, pulse oximeters, etc. can be included in this system.

Due to high cost, training phantoms are often inaccessible and their manufacturing technologies are quite sophisticated. The purpose of this paper is to develop an inexpensive and reproducible technology for creating ultrasound training phantoms. These phantoms are a 3D printed porous medium composed of 156-µm-thick photopolymer resin fibers and include models of cysts ranging from 4 to 8 mm in diameter, effectively simulating a muscle tissue with anechoic lesions. A custom software generates a virtual phantom model, enabling precise control over its properties. We believe that the results of the acoustic characteristics' measurements for the designed phantoms provide an opportunity to mimic muscle (1547 m/s) and breast (1510 m/s) tissues. Following the creation of the phantom, a series of assessments were conducted to evaluate its efficacy for needle insertion (involving 3 observers) and to identify its mimicked tissue type (with 29 observers participating). The findings revealed that the phantom is capable of enduring up to 300 punctures in a single location without exhibiting significant decline in image quality. A subsequent survey of ultrasound specialists, who possessed a range of professional experiences, indicated that the ultrasound images produced by the phantom predominantly corresponded to those of muscle tissues upon visual examination. The 3D printing process for the phantom 60 mm × 60 mm × 30 mm in size was completed in 3 h and 23 min. The proposed technology allows creating low-cost, long-lasting phantoms for training in ultrasound diagnostics and ultrasound-guided procedures. The phantom designed using widely available photopolymer resin, while the custom software and high-resolution 3D printing ensures reproducibility of the shape and positions of the fibers and inclusions. The phantom mimics muscle tissues with multiple cysts and can be used to develop basic coordination and navigation skills required for ultrasound diagnostics.

Hepatic steatosis, affecting one-third of the global population, is a key challenge in gastroenterology with limited screening focus. It characterizes metabolic dysfunction-associated steatotic liver disease, which is increasingly prevalent and linked to metabolic issues, yet lacks accessible non-invasive early detection tools. This study evaluates an AI model's ability to predict controlled attenuation parameter (CAP) scores, providing qualitative estimates of mild and moderate or greater liver steatosis degrees. The study included 705 participants from a nutrition clinic, with data collected on 27 features such as physical exams, body measurements, and InBody270 results. CAP score was obtained from transient elastography findings. We developed a novel graph neural network (GNN) architecture that conceptualizes the human body as an interconnected graph structure to capture complex physiological relationships between different anatomical regions. The proposed GNN model significantly outperformed traditional machine learning approaches, achieving RMSE of 23.7 dB/m, MAE of 18.9 dB/m, and R² of 0.87. Attention-guided feature importance analysis identified waist circumference, trunk fat mass, and neck circumference as the most influential predictors of CAP scores. The graph-based model outperforms traditional machine learning in predicting CAP scores, leveraging body relationships for reliable, non-invasive hepatic steatosis screening across all severities.