Physical and Engineering Sciences in Medicine最新文献

Bone scintigraphy is an important tool for detecting bone lesions. This study aimed to improve and evaluate the performance of our previously-developed deep learning-based model called MaligNet in helping nuclear medicine (NM) physicians interpret bone scan. Bone scintigraphy of 553 patients with imaging data from six-month follow-up records were split into training, validation, and test sets in a ratio of 353:100:100 to re-train MaligNet. Seven nuclear medicine physicians, including two junior and five senior physicians, were asked to segment and classify lesions in the test set images without and with AI assistance, which was the prediction of MaligNet. The improved performance of MaligNet was evaluated using the precision-recall (PR) and receiver operating characteristic (ROC) curves for lesion-based and patient-based classifications, respectively. The impact of AI assistance on physician reading was evaluated using reading time per case and malignancy diagnostic performance metrics. The re-trained MaligNet yielded considerably higher area under the PR curve (0.334 vs. 0.225) and higher area under the ROC curve (0.881 vs. 0.789) than the original model. For patient-based classification, AI assistance improved the average accuracy, sensitivity, specificity, and precision of the physician by 2.14%, 0.89%, 2.38%, and 1.97%, respectively, while reducing the average reading time by 31.14%. For lesion-based classification, it improved physicians' average precision by 2.95%, but did not improve sensitivity. With AI assistance, junior physicians achieved diagnostic performances comparable to those of senior physicians. AI assistance with MaligNet improved bone scintigraphy diagnostic performance and showed promise in clinical practice.

Atrial fibrillation (AFB) and atrial flutter (AFL) are cardiac arrhythmias very often associated with the aggravation of other cardiac pathologies and increase the risk of stroke and heart failure. Their detection is therefore crucial. Automated analysis of the ECG signal has been suggested to assist cardiologists in the diagnosis of AFB and AFL. In this paper, a novel automated electrocardiogram (ECG) signal analysis method to aid in the detection of AFB and AFL is presented. The first step of the method consists of processing the original ECG signal. The second step carries out the classification using a modified MobileNetV2 convolutional neural network (CNN) powered by transfer learning. This CNN classifies the fed-in ECG signals into atrial fibrillation (AFB), atrial flutter (AFL), other (OTH), normal sinus rhythms (NOR), and noisy (NOI) recordings. The performance of the proposed method was assessed and scored using the Physio Net/Computing in Cardiology (CinC) 2017 dataset and the MIT-BIH Atrial Fibrillation Database (MIT-BIH). The experimental results showed that the proposed method gave an F1 score of 96.08%, sensitivity of 97.1%, specificity of 99.53%, and accuracy of 95.1% for atrial fibrillation, for the CinC 2017 dataset. For the MIT-BIH dataset, an F1 score of 99.54%, sensitivity of 99.51%, specificity of 99.64%, and accuracy of 99.5% were obtained. The results disclosed above on 2 databases prove that the proposed algorithm is efficient, robust, and can be used to assist cardiologists.

Purpose: Electron beam radiotherapy is a crucial modality for treating superficial tumors. Accurate dose calculation is essential for treatment efficacy and minimizing side effects. While Monte Carlo (MC) simulations are considered the gold standard for dose calculation, their computational cost can be prohibitive. The electron Monte Carlo (eMC) algorithm offers a faster alternative, but its accuracy, especially in heterogeneous environments, remains a concern.

Methods and materials: This study compares electron beam dose distributions calculated using the eMC algorithm in a treatment planning system (TPS) with those obtained from full MC simulations using the GATE platform. We evaluated the eMC algorithm's performance across various electron energies (6, 9, and 12 MeV) and field sizes (6 × 6 cm² to 20 × 20 cm²), in both homogeneous water phantoms and heterogeneous phantoms incorporating lung-equivalent and bone-equivalent materials.

Results: Results in homogeneous phantoms demonstrated generally good agreement between eMC and GATE, with some discrepancies observed in penumbra regions and at higher energies, particularly for larger field sizes. In heterogeneous phantoms, significant deviations were observed, particularly in lateral dose profiles near density interfaces and at higher beam energies, with percentage of points with less than 3% difference dropping considerably.

Conclusion: These findings highlight the limitations of the eMC algorithm in accurately modeling complex tissue heterogeneities. While eMC provides acceptable accuracy in relatively simple scenarios, its performance degrades significantly in clinically realistic heterogeneous environments, necessitating caution in treatment planning and highlighting the ongoing need for improved dose calculation algorithms.

Obsessive-compulsive disorder (OCD) causes unwanted thoughts and repetitive actions and leads to many problems in a person's life. In this study, Electroencephalography (EEG) signals and deep learning methods were used to diagnose OCD patients early. Three popular pre-trained convolutional neural network (CNN) models are developed for scalp-EEG data analysis: EEGNet, Shallow ConvNet, and Deep ConvNet. Three pre-trained CNNs were utilized as transfer learning models. Following the fine-tuning of models with our raw EEG data, an ensemble of three scalp EEG-based CNN models was used, employing weighted majority voting, in which weights of these base classifiers were optimized by the Differential Evolution (DE) algorithm. Shallow ConvNet has the highest performance with an accuracy of 85.91±0.72, sensitivity of 82.19±0.72, and specificity of 93.34±2.91 among all models. Ensemble these three scalp EEG-based CNN models achieved superior performance with an accuracy of 87.03±0.46, sensitivity of 82.21±0.56, and specificity of 96.69±1.28. Consequently, a hybrid proposed model based on pre-treatment raw EEG signals can independently extract distinctive characteristics and accurately identify OCD patients.

Electrocardiogram (ECG) signals are usually contaminated by numerous artefacts during the recording process, and the quality of physiological information related to the heart is compromised. Due to this, artefact cancellation has become necessary for ECG signals. In this paper, swarm intelligence-based optimally tuned adaptive noise cancellers (ANCs) have been proposed and applied to denoise the ECG signal. The results have been analysed both qualitatively and quantitatively for noise cancellation from ECG signals through the ANCs optimized by using the seagull optimization algorithm (SOA), the Neighbourhood-based lineal population size success history-based adaptive differential evolution (NLSHADE) algorithm and the hyperbolic gravitational search algorithm (HGSA). The performance of the proposed methodology has been validated by using the additive white Gaussian noise at a diverse signal-to-noise ratio (SNR) on two publicly available datasets of ECG signal from the arrhythmia database (ADB) and QT ECG database (QTDB). The reference noise for ANC was considered using the noise stress test database (NSTDB). The performance of SOA-assisted ANC has been tested with the help of the Wilcoxon signed-rank test. The proposed technique-based ANCs supplied an enhanced percentage root mean squared deviation (PRD) value of 3.40E-03, mean squared error (MSE) value of 1.35E-11 and mean SNR improvement of 10.986 dB as compared to the reported state-of-the-art methods along with the benchmark competent algorithms, namely NLSHADE and HGSA.

This paper reviews the latest methods for creating control interfaces for intention detection in active transfemoral prosthetic devices. A literature review over the past two decades identified several control algorithms for intention detection. Sources included scientific publications, books, and online resources focusing on knee prostheses. Three main areas of research were identified. The studies were assessed using the Downs and Black checklist, detailing their control techniques and performance assessments. Initially, 213 studies were retrieved; 33 were selected for this review. Fifteen (15) papers examined control strategy frameworks and goal outputs of active prosthetic legs. Two (2) papers discussed conventional control methods for transfemoral prosthetic legs. Four (4) studies explored potential implementations of intention detection, and twelve (12) papers investigated machine learning algorithms for active prosthetic legs. The review suggests using a simpler sensory system paired with innovative control algorithms to translate limited sensor data into a broader set of relevant information. Effective sensory systems and intention detection algorithms are crucial for active transfemoral prosthetic limbs. This review presents the feasibility of control interfaces that enable intention detection for active prosthetic legs, offering multiple references and classifying different works in the field.

The Omni Legend (GE Healthcare), equipped with a digital bismuth germanium oxide PET/CT system, has been recently developed. However, the performance of the Omni Legend without a time-of-flight (TOF) system for ¹⁸F-fluciclovine imaging is still unclear. Therefore, this study evaluated the image quality of the Omni Legend according to the Japanese brain tumor phantom test (JBT) criteria, and assessed its potential use for ¹⁸F-fluciclovine imaging. This study followed the JBT procedures. A brain tumor phantom, which includes six hot spheres of different diameters, was filled with an ¹⁸F-fluorodeoxyglucose solution with a radioactivity concentration ratio of 3 (spheres):1 (background). PET scanning was performed using the Omni Legend with a 30-min list mode acquisition. The PET data were reconstructed using an ordered subset expectation maximization (OSEM), an OSEM with point spread function (OSEM + PSF), and a Bayesian penalized likelihood (BPL) under standard clinical parameters. The image quality was evaluated using the JBT criteria, including contrast for a 7.5-mm sphere, recovery coefficient (RC) for a 10.0-mm sphere, standardized uptake value of total background (SUV_TOT), and detectability for a 7.5-mm sphere. The contrast, RC, and SUV_TOT were 25.1%, 0.70, and 1.00, respectively in OSEM; 25.8%, 0.80, and 0.99 in OSEM + PSF; and 33.8%, 0.93, and 0.99 in BPL. The 7.5-mm sphere was detected by all three methods. All of the JBT criteria were satisfied, regardless of the PET image reconstruction methods. This study demonstrated that the Omni Legend without TOF satisfies all JBT criteria and has the potential to provide high-quality images in ¹⁸F-fluciclovine imaging.

Colorectal cancer starts in the large intestine and rectum. It develops when small, usually harmless growths called polyps become cancerous over time. Early diagnosis increases the chances of successfully treating colorectal cancer. A new hybrid model was developed to detect colorectal tissue types. In the first step of the model, the quality of the images was increased using Denoising Convolutional Neural Network (DNCNN) networks. The feature maps of the images were then obtained using DarkNet53 and shrunk using the Gorilla Troops Optimization Algorithm (GTO) to speed up the proposed model's performance and boost the performance. Finally, a support vector machine (SVM) classifier was used to classify the feature maps. The proposed model obtained an accuracy of 95.5% in classifying eight tissue types in colorectal cancer histopathology specimens (Adipose, Complex, Debris, Empty, Lympho, Mucosa, Stroma, and Tumor). To make the developed model more generalizable, robust, and accurate, it needs to be tested with a huge dataset collected from various centers and races.

The increasing demand for secure, high-quality medical image transmission across healthcare institutions has posed a significant challenge to modern telemedicine systems. Traditional network infrastructures often fail to provide sufficient bandwidth and low latency required for transferring large volumes of high-resolution medical images, such as MRI and CT scans, over long distances. To address this limitation, a fiber-optic transmission framework was designed and evaluated with the objective of enhancing the speed, reliability, and accuracy of inter-hospital medical image sharing. In this study, a simulation-based approach was employed using OPTISYSTEM and MATLAB to model the optical transmission chain, including stages of image digitization, modulation, fiber propagation, and optical-to-electrical conversion at the receiving end. Various performance parameters such as Bit Error Rate (BER), Quality Factor (Q), transmission power, and noise levels were analyzed for different image resolutions and transmission distances. The results showed that Q-Factor values between 8.5 and 9.5 were obtained, with BER reaching values as low as 10⁻²⁰, even for high-resolution images transmitted over distances up to 90 km. These results were compared to existing benchmarks in the literature and demonstrated superior performance. The proposed system exhibited strong robustness in handling large image datasets, with minimal signal distortion and negligible transmission errors. It was concluded that the adoption of this fiber-optic architecture could significantly improve the efficiency of telemedicine applications, offering a reliable and high-capacity solution for real-time diagnostic collaboration and patient monitoring between geographically distributed medical facilities.