Physical and Engineering Sciences in Medicine最新文献

Cardiovascular diseases (CVDs) are still the leading cause of death worldwide, emphasizing the critical need for reliable diagnostic systems. This study aims to create a standardized electrocardiogram (ECG) dataset that can be used to detect and classify six major CVDs using machine learning techniques and investigate feature selection and extraction methods for improved performance. A large dataset of 34,580 12-lead ECG recordings was collected from Sher-i-Kashmir Institute of Medical Sciences (SKIMS), Srinagar, Jammu and Kashmir spanning six clinically confirmed classes: Normal, Cardiac Arrhythmia, Coronary Heart Disease, Cardiomyopathy, Stroke, and Heart Failure. Data pre-processing involved baseline correction, removal of artifacts and the extraction of 14 clinically informative features. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied, resulting in an equal distribution of 16.7% of the data across each class. Ten Machine learning and deep learning models-Logistic Regression, Decision Tree, Random Forest, SVM, KNN, Naive Bayes, Gradient Boosting, MLP, DNN, and RNN-were trained and tested. SHAP and LIME methods were used for interpretability. On the raw dataset, Random Forest and Gradient Boosting produced highest performance with test accuracy of 99.88%, precision of 99.88%, recall of 99.88%, and F1-score of 99.88%. After SMOTE, DNN significantly improved (Accuracy: 97.62%, Precision: 97.66%, Recall: 97.62%, F1-score: 97.64%), while MLP obtained an F1-score of 98.49% and RNN obtained 94.76%. All models exhibited better generalization and stability after SMOTE. The balanced, heterogeneous, and clinically verified ECG dataset supported the highly accurate, interpretable, and real-time classification of CVD. SMOTE significantly improved the performance of the model, particularly for deep networks, substantiating its effectiveness in the class imbalance problem. These results place the proposed model and dataset as effective tools for clinical decision support in the diagnosis of cardiovascular disease.

Wrist pulse measurement offers significant insights into cardiovascular health. However, the application of various sensors, such as optical, pressure, image, and ultrasonic, is limited due to issues like bright environments, incompatibility with pressure adjustments, and system complexity. Recent studies suggest condenser microphones as promising alternatives, though the optimal type among various condenser microphones remains unclear. This study explores the application of three different condenser microphones using four regression-based machine learning models (Partial Least Square Regression, Ridge Regression, Principal Component Regression, and Nu-Support Vector Regression) for wrist pulse measurement based on pulse rate accuracy. One omnidirectional condenser microphone, previously used for wrist pulse measurement, and two commonly available unidirectional condenser microphones were evaluated. A mechanical system for pulse acquisition was developed, and data were collected from 27 healthy subjects using each microphone alternatingly. Extracted time-domain and statistical features were used as inputs to compare the predicted pulse rates with the ground truth pulse rate values. Results indicated that unidirectional condenser microphones were more accurate than the omnidirectional type. Among the unidirectional microphones, the one with a sensitivity range of - 50 to - 44 dB outperformed the microphone with a sensitivity range of - 40 to - 34 dB. The Nu-Support Vector Regression model exhibited the least errors, indicating superior predictive capabilities compared to the other models. In conclusion, this study provides valuable insights into selecting appropriate condenser microphones for wrist pulse measurement, offering a guiding framework for future research in this domain.

This study evaluated whether half-acquisition (180° scan) pediatric cone-beam computed tomography (CBCT; 3D Accuitomo F17, J. Morita, Kyoto, Japan) reduces radiation exposure while maintaining sufficient diagnostic image quality for identifying ectopic eruptions and impacted teeth. Additionally, it was investigated whether a low-noise reconstruction filter (G_101) mitigates image quality degradation in 180° scans. Three board-certified oral and maxillofacial radiologists certified by the Japanese Society for Oral and Maxillofacial Radiology visually evaluated clinical images from 12 pediatric patients (aged 6-10 years). The image quality was objectively assessed using phantom-based analyses of the modulation transfer function (MTF), noise power spectrum (NPS), and comprehensive objective image quality calculated from MTF and NPS values. Although 180° images showed increased noise and slightly lower visual assessment scores compared with 360° images, they remained diagnostically acceptable. In 180° reconstructions, the median visual scores with the G_101 filter were slightly higher than those with the standard G_001 filter, with small differences (within approximately 0-3 points on a 100-point scale), although the differences were not statistically significant. Interestingly, in approximately 28% of 180 evaluations, 180° images scored higher than 360° images, likely due to reduced motion artefacts from shorter acquisition. In a previous phantom experiment, the dose area product (DAP) for 360° and 180° scans was 490 mGy cm² and 249 mGy cm², respectively, indicating that 180° scan reduces radiation exposure while maintaining clinically acceptable image quality. These findings suggest that half-acquisition, when combined with an appropriate reconstruction filter, may offer a practical, low-dose alternative for pediatric dental imaging.

Diabetic foot ulcers (DFUs) pose a significant complication of diabetes with the potential to lead to amputation if not effectively managed. Current DFU treatments require rigorous monitoring by both healthcare professionals and patients, which is challenging due to the high costs associated with diagnosis, treatment and long-term care. A major limitation of these approaches is their limited capacity to identify highly relevant pattern connections and broad contextual correlations resulting inaccuracies in classifying regions of interest. This research introduces an attention enhanced deep learning-based automated approach for assessing DFUs using images to expedite the investigation process and offer optimal recommendations. Adaptive thresholding is employed to enhance the contrast and uniformity of DFU images and thereby improves the feature extraction. A hybrid model incorporating coordinate attention enhanced ConvNeXt is used for effective DFU image classification to enhance the representation of complex patterns through efficient parameter utilization. The ConvNeXt architecture is designed to scale efficiently across various sizes by utilizing depthwise separable convolutions and improved image normalization. This model is augmented with coordinate attention, which captures spatial information in both horizontal and vertical directions, aiding in the extraction of long-range dependency features for more accurate classification of DFU images. Experimental results demonstrate that the model achieves an accuracy of 97.16% and F1-score of 0.97.

Schizophrenia (SCZ) is a complex neurological disorder characterized with deformed understanding. Traditionally, neurologists rely on interviews and visual analysis for SCZ detection and treatment. This work aims to present an automatic classification of the SCZ by electroencephalogram (EEG) signals, which can obtain variations in neural activity associated with cognitive changes in SCZ. This work presents an Optimized Gated Temporal Convolutional Network (OGTCN) for SCZ detection. The suggested OGTCN is the integration of the networks like Gated Recurrent Unit (GRU), Improved Temporal Convolutional Network (ITCN) and the Enhanced Mountain Gazelle Optimizer (E-MGO). The dataset utilized in this work comprises 19 channel EEG signals from 28 individuals, and the second dataset includes 64 channel EEG signals from 81 individuals. Here, the accuracy values achieved are 99.89% (Dataset1) and 99.99% (Dataset2). This research highlights the effectiveness of the OGTCN in enhancing EEG data for supporting proper detection of the SCZ. By integrating the DL model with E-MGO, this approach provided a promising solution to enhance diagnosis of the mental disorder via analysis of the EEG signal.

High-frequency electrocardiography (HF-ECG) enhances ischemia detection by capturing microvolt-level changes in the QRS complex; however, clinical adoption requires validating digital systems against Analog standards. We recorded HF-ECG signals simultaneously from 12 healthy subjects (84 beats total) using a five-stage Analog reference (100-500 Hz band-pass, gold connectors) and Sydäntek's 10× capacitive sensors; both outputs were digitized using a Texas Instruments ADS1298. Signals underwent 10× amplification with a low-noise op-amp, the Analog output was scaled to match, and data were processed in PulseTek™ and stored in PulseVault™. Root mean square (RMS), Amplitude, Kurtosis, and Frequency content were compared using Bland-Altman analysis (Analog as the reference); values reported reflect pre-amplified measurements in microvolts (µV). Mean differences between the Analog setup and Sydäntek fell within the 95% limits of agreement (LOA): RMS, 6.39 µV (- 49.74 to 62.52 µV); amplitude, 1.82 µV (- 57.09 to 60.73 µV); kurtosis, 1.93 (- 5.13 to 1.54); and frequency, 2.1 Hz (- 5.8 to 6.2 Hz), all within a 5% clinical tolerance when scaled 10× (~ 10-20 mV). Sydäntek matched analog fidelity, with frequency peaks near ~ 150 Hz, indicating digital HF-ECG performance equivalent to that of the Analog system on key metrics. Its wearable design and cloud integration provide a portable, reliable alternative for ischemia detection with broader clinical applicability.

Phantom experiments are widely used for standardisation in positron emission tomography (PET), but current practices do not necessarily reflect clinical reality and require meticulous phantom preparation for repeatability. 3D printing can reduce these limitations by optimizing preparatory methods and improving phantom features. This work proposes employing 3D-printed porous grids as an alternative mechanism to emulate targets with contrast. Acrylonitrile butadiene styrene (ABS) cubic grids (4 cm/side) with varying design characteristics and targets were printed. Grids were immersed in a [¹⁸F]FDG solution with soap within a conventional phantom. Five consecutive acquisitions were repeated on five different days (Day 0, 1, 4-6) using a Discovery MI PET/CT. Target representation index (TRI) (analogous to recovery coefficient) and dilution coefficient (DC) were the metrics used for the analysis. Friedman test was utilized for statistical inference across days. PET images resulted in clear demarcation of various contrast regions produced by the dilution grid. Quantitative metrics showed consistent results across trials, confirming robustness. Dilution coefficient achieved (mean ± std. dev.) were 0.55 ± 0.05, 0.41 ± 0.06, and 0.33 ± 0.03 versus 0.5, 0.4 and 0.3 (theoretical), falling within 10% threshold. Observed TRI_{max, mean} were in range of 0.4-1.2. Correlation across days was strong for TRI_{max, mean} (p-values ≥ 0.67) but the DC_max (p-values ~ 0.03) values denoted minor bias in generated contrast due to noise. 3D-printed grids offer a reliable, reproducible alternative for PET/CT assessment. 27 hot targets with varying contrasts and size were produced with a single tracer administration and the metrics stayed stable across different acquisitions.

Evaluation of red blood cell (RBC) aggregation is crucial for early detection of diseases such as ischemic cardiovascular disease, type II diabetes mellitus, deep vein thrombosis, and sickle cell disease. Ultrasound, a non-invasive and real-time technique, is widely used for monitoring RBC behavior. However, measurement inaccuracies caused by instrumentation and human error can introduce data anomalies, degrading the generalization capability of deep learning models. To address this issue, we propose a Multidimensional Transformer-CNN (MTCN) model trained on incomplete data. Specifically, 20% of the original ultrasonic data is randomly masked via a Mask-Head module, followed by a multidimensional Transformer encoder. A multi-dimensional adaptive fusion module aggregates features across various dimensions, which are then passed through a classification layer. Considering that ultrasonic RF signals contain negative-valued components, we employ a Gaussian Error Linear Unit (GELU) activation function to preserve this information while ensuring model efficiency. Experimental results on RBC aggregation dataset demonstrate that MTCN outperforms existing models by achieving an accuracy of 96.89% and an F1-score of 96.92%. These findings confirm the model's robustness and strong generalization capability, providing a promising approach for the accurate and non-invasive monitoring of RBC aggregation.

Prostate cancer is a common disease among men worldwide, and patients are frequently treated with external beam radiation therapy (EBRT). Precise radiation dosage planning is required for treatment application to minimise side effects. Identification of tissue boundaries is crucial for radiation therapy treatment planning by providing essential shape information to clinicians, allowing optimisation of treatment delivery while limiting damage to healthy tissues. In this study, two approaches incorporating Principal Component Analysis (PCA) were proposed to investigate patterns of shape variation during radiotherapy. Trajectory analysis utilized PCA to track the evolution of prostate shape throughout the course of treatment, while variation pattern analysis examined the overall range of changes in target tumor volume over the treatment period. The data used consisted of 261 mesh prostate surfaces generated from radiation oncologist manual contours from 33 patients across eight weeks of treatment. Trajectory analysis revealed significant shape variations in the left superior region and from the inferior to the posterior-right region of the prostate throughout the treatment period. Variation pattern analysis indicated an overall increase in target tumor volume during treatment, with the highest average variation observed in the anterior superior region. Notably, most shape variations occurred during the first week of treatment, suggesting that implementing a second set of updated contours after the initial week's scan could improve accuracy in defining target volumes for subsequent treatments.

Photon starvation in computed tomography, which occurs when insufficient photon counts allow electronic noise to dominate the signal, leads to severe degradation in reconstructed images. This paper proposes a pre-correction method that combines a negative feedback mechanism with an adaptive diffusion filter to mitigate photon-starved effects by suppressing electronic noise in the sinogram prior to logarithmic transformation. The method was evaluated using ultra-low-dose scans of an anthropomorphic torso phantom and clinical patient data. For comparison, several sinogram-based denoising methods were also applied. The proposed method produced reconstructed images with the lowest noise, highest structural similarity, and superior spatial resolution, along with significantly reduced streaking and bias artifacts. Experimental results demonstrate that the proposed method effectively suppresses noise, streaking artifacts and large-scale bias artifacts in low-signal anatomical regions under severe photon starvation in low-dose conditions, while maintaining acceptable resolution.