Physical and Engineering Sciences in Medicine最新文献

Radiotherapy is a crucial treatment for brain tumor malignancies. To address the limitations of CT-based treatment planning, recent research has explored MR-only radiotherapy, requiring precise MR-to-CT synthesis. This study compares two deep learning approaches, supervised (Pix2Pix) and unsupervised (CycleGAN), for generating pseudo-CT (pCT) images from T1- and T2-weighted MR sequences. 3270 paired T1- and T2-weighted MRI images were collected and registered with corresponding CT images. After preprocessing, a supervised pCT generative model was trained using the Pix2Pix framework, and an unsupervised generative network (CycleGAN) was also trained to enable a comparative assessment of pCT quality relative to the Pix2Pix model. To assess differences between pCT and reference CT images, three key metrics (SSIM, PSNR, and MAE) were used. Additionally, a dosimetric evaluation was performed on selected cases to assess clinical relevance. The average SSIM, PSNR, and MAE for Pix2Pix on T1 images were 0.964 ± 0.03, 32.812 ± 5.21, and 79.681 ± 9.52 HU, respectively. Statistical analysis revealed that Pix2Pix significantly outperformed CycleGAN in generating high-fidelity pCT images (p < 0.05). There was no notable difference in the effectiveness of T1-weighted versus T2-weighted MR images for generating pCT (p > 0.05). Dosimetric evaluation confirmed comparable dose distributions between pCT and reference CT, supporting clinical feasibility. Both supervised and unsupervised methods demonstrated the capability to generate accurate pCT images from conventional T1- and T2-weighted MR sequences. While supervised methods like Pix2Pix achieve higher accuracy, unsupervised approaches such as CycleGAN offer greater flexibility by eliminating the need for paired training data, making them suitable for applications where paired data is unavailable.

Radiomic biomarkers have demonstrated significant potential in non-invasively assessing tumor biology and providing essential insights for precision medicine. However, the clinical translation is often hindered by challenges in multicenter studies, primarily due to a lack of standardization, such as variations in scanner models, acquisition protocols, reconstruction techniques, etc. This study aims to assess the impact of various harmonization methods in multicenter, 18 F-FDG PET-based radiomics for the classification of lung cancer histological subtypes using a machine learning model. Retrospective data included 178 lung cancer cohorts, comprising 117 adenocarcinomas and 61 squamous cell carcinomas from three different centers. PET DICOM image data was preprocessed with 3D ROI segmentation of the lung tumor and healthy liver, followed by the extraction of 111 radiomic features. Subsequently, Z-Score, Quantile, and ComBat were applied to generate three different harmonized datasets. Feature distribution was analyzed, and the top ten features were selected using recursive feature elimination. An eXtreme gradient boosting model was built on each dataset, and performance was assessed using accuracy, precision, sensitivity, specificity, and AUC with a 95% confidence interval. Variations in radiomic feature distribution and feature selection were observed after applying different harmonization methods. During validation of the trained model, AUC improved from 0.556 [95% CI 0.551-0.563] in the unharmonized data to 0.719 [95% CI 0.710-0.720], 0.952 [95% CI 0.951-0.954], and 0.996 [95% CI 0.995-0.996] in Z-Score, Quantile, and ComBat harmonized data, respectively, for classifying adenocarcinoma and squamous cell carcinoma subtypes. The study indicates that feature selection was affected by the different harmonization methods. The ComBat method was shown to significantly enhance the performance of AI-assisted PET radiomics.

A naevus is a benign melanocytic skin tumour made up of naevus cells, characterised by variations in size, shape, and colour. Understanding naevi is essential due to their significant role as markers for the risk of developing melanoma. This study focused on creating a visual representation called a manifold that illustrates the distribution of two types of naevi: suspicious and non-suspicious. The research aimed to classify real naevi using generative adversarial networks (GANs), while also generating realistic synthetic samples and interpreting their distribution through a variational manifold. This inquiry holds promise for applying data-driven methods for early melanoma detection by identifying distinct features linked with suspicious naevi. Our variational autoencoder auxiliary classifier generative adversarial network (VAE-ACGAN) for suspicious naevi revealed a manifold with outstanding performance, including specificity, sensitivity, and area under the curve (AUC) scores, particularly representing suspicious naevi. These models surpassed various deep learning frameworks in key performance metrics while producing a manifold that indicated a significant distinction between the two categories in the resultant image, yielding high-quality and life-like representations of naevi. The results highlight the potential application of GANs in expanding data sets and enhancing the effectiveness of deep learning algorithms in dermatology. Accurate identification and categorisation of naevi could facilitate early melanoma detection and deepen our understanding of these skin lesions through an interpretable clustering method based on visual similarities.

Alzheimer's Disease (AD) is a chronic neurological disorder that impairs the cognitive and behavioral abilities of older people. Early detection and treatment are crucial for minimizing the progression of the disease. Electroencephalogram (EEG) makes it possible to investigate the brain activities linked to various forms of disabilities experienced by individuals with AD. Nevertheless, the EEG signals are non-linear and non-stationary in nature making it difficult to retrieve the concealed information from the EEG signals. Therefore, a Fourier Decomposition Method (FDM) and Hilbert Transform (HT) based EEG signals analysis (FHESA) method is developed in this paper for the automated detection of AD. The FHESA method aims to efficiently analyze the EEG data to identify the important brain regions vulnerable to AD, and to assess the impact of various EEG channels for the timely and early detection of AD. The proposed FHESA method is divided into three primary stages. The first stage deals with the decomposition of the EEG signals into a finite number of Fourier Intrinsic Band Functions (FIBFs). In the second stage, HT is applied to all FIBFs to obtain instantaneous amplitude, frequency, and phase, that are then used to construct feature vectors. In the last stage, various Machine Learning (ML) algorithms are used to classify these feature vectors for efficient AD detection. Two distinct data sets are employed to assess the effectiveness of the proposed FHESA method. The outcome demonstrates that with dataset-I and dataset-II, the proposed methodology can detect AD with 98% and 99% accuracy, respectively. The performance of the proposed FHESA method is compared to other state-of-the-art methods used for AD detection. The promising results show that the proposed FHESA method can assist neurological experts in identifying and utilizing EEG signals for AD detection.

To assess the clinical outcome of single-arc coplanar dynamic conformal arc (C-DCA) in three-fraction hypofractionated stereotactic radiotherapy (3F-HSRT) for single very small brain metastasis (BM; gross tumor volume [GTV] ≤ 1 cm³) from non-small cell lung cancer (NSCLC) and to compare treatment planning with single-arc coplanar volumetric modulated arc therapy (C-VMAT). From December 2019 to May 2023, twenty NSCLC patients with single BM treated with 3F-HSRT (24-39 Gy/3f) using C-DCA were enrolled in this study. Each plan was replanned using C-VMAT, and relevant planning indices were compared. Clinical outcome was evaluated following C-DCA treatment. C-VMAT yielded a higher homogeneity index (1.41 vs. 1.16, p < 0.001) and GTV D_98% (38.10 Gy vs. 32.72 Gy, p = 0.008), with slightly smaller normal brain tissue (NBT) V_23Gy, and V_21Gy. However, C-DCA offered 31.37% lower monitor units (p = 0.008) and 36.55% shorter beam on time (p = 0.007) while achieving a significantly higher gamma passing rate for the 2%/1 mm criterion (p = 0.001). As of October 2023, the median follow-up time was 9.2 months. The intracranial disease control rate was 70%, with a median intracranial progression-free survival of 11.4 (95% CI 4.5-18.3) months and a 1-year intracranial control rate of 45.4%. The intracranial local disease control rate was 95%. Only two irradiated lesions progressed at the end of follow-up. The cerebral radiation necrosis rate of all patients was 5%. For small BM, C-DCA provided nearly equivalent target coverage and NBT sparing to the C-VMAT while maintaining higher delivery efficiency and accuracy. C-DCA HSRT also provided good local control and limited toxicity.

In the application of wrist-based Photoplethysmography (PPG) devices for health monitoring, assessing the quality of PPG signals is essential for accurately monitoring cardiovascular parameters. However, the wrist-based PPG signal is susceptible to motion and light interference in practical applications. A machine learning-based signal quality assessment algorithm for wrist PPG signals was proposed to improve the accuracy and reliability of the monitoring data. The algorithm's performance was evaluated on two datasets: the publicly available Wearable and Clinical Signals (WCS) dataset, containing 3,038 wrist-based PPG segments collected from 18 volunteers using an Empatica E4 device; our LAB dataset, comprising 2,426 wrist-based PPG segments acquired from 12 volunteers under varied interference conditions via a custom-developed wearable watch system. Data pre-processing encompassed denoising and normalization, followed by the extraction of 11 mathematical statistical features in time and frequency domains based on pulse wave morphology and 2 features based on template matching (Euclidean Distance and Correlation Coefficient). The classifier, constructed using the LightGBM algorithm, achieved high performance under rigorous leave-one-subject-out cross-validation (LOSO-CV) on the WCS dataset (accuracy = 92.6%, precision = 96.6%, recall = 89.8%, F1-score = 91.4%, AUC = 0.925) and the LAB dataset (accuracy = 96.1%, precision = 98.1%, recall = 95.2%, F1-score = 96.6%, AUC = 0.941). The results show that the machine learning algorithm for wrist-based PPG signal quality assessment, combining the mathematical statistical features in time and frequency domains and the template matching features, can effectively enhance the performance of signal quality assessment, and provides a powerful tool for improving the accuracy of wearable devices in cardiovascular health monitoring.

Film quality assurance software is an integral component of patient-specific quality assurance for various radiation techniques where high degrees of geometrical and dosimetric accuracy are required. Evaluating the accuracy of film quality assurance software products has relied on various techniques ranging from comparative analyses, measurements with phantoms and other detectors, along with confidence from industry-standard peer reviews. The aim of this work was to determine if a series of synthetically created film and DICOM images can be used to test the accuracy of certain patient specific quality assurance metrics used in film quality assurance software packages. The synthetic images have been engineered to simulate radiographic film scanned TIFF images and treatment planning system exported DICOM files. Each pair of images were designed to test a particular component of the software's ability to process curve fitting, dosimetric differences, distance to agreement, percentage threshold and gamma analysis. It was found that synthetic film could simulate radiographic scanned films and treatment planning system DICOM planes and provide the physicist with empirical data on the accuracy of the mentioned metrics. The series of tests also assisted the physicist in identifying optimal calibration models, validating geometric and dosimetric variations, and offering insights into potential differences in lower dose penumbras.

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.