Physical and Engineering Sciences in Medicine最新文献

Access to medical imaging data is crucial for research, training, and treatment planning in medical imaging and radiation therapy. However, ethical constraints and time-consuming approval processes often limit the availability of such data for research. This study introduces DICOMator, an open-source Blender add-on designed to address this challenge by enabling the creation of synthetic CT datasets from 3D mesh objects. DICOMator aims to provide researchers and medical professionals with a flexible tool for generating customisable and semi-realistic synthetic CT data, including 4D CT datasets from user defined static or animated 3D mesh objects. The add-on leverages Blender's powerful 3D modelling environment, utilising its mesh manipulation, animation and rendering capabilities to create synthetic data ranging from simple phantoms to accurate anatomical models. DICOMator incorporates various features to simulate common CT imaging artefacts, bridging the gap between 3D modelling and medical imaging. DICOMator voxelises 3D mesh objects, assigns appropriate Hounsfield Unit values, and applies artefact simulations. These simulations include detector noise, metal artefacts and partial volume effects. By incorporating these artefacts, DICOMator produces synthetic CT data that more closely resembles real CT scans. The resulting data is then exported in DICOM format, ensuring compatibility with existing medical imaging workflows and treatment planning systems. To demonstrate DICOMator's capabilities, three synthetic CT datasets were created: a simple lung phantom to illustrate basic functionality, a more realistic cranial CT scan to demonstrate dose calculations and CT image registration on synthetic data in treatment planning systems. Finally, a thoracic 4D CT scan featuring multiple breathing phases was created to demonstrate the dynamic imaging capabilities and the quantitative accuracy of the synthetic datasets. These examples were chosen to highlight DICOMator's versatility in generating diverse and complex synthetic CT data suitable for various research and educational purposes, from basic quality assurance to advanced motion management studies. DICOMator offers a promising solution to the limitations of patient CT data availability in medical physics research. By providing a user-friendly interface for creating customisable synthetic datasets from 3D meshes, it has the potential to accelerate research, validate treatment planning tools such as deformable image registration, and enhance educational resources in the field of radiation oncology medical physics. Future developments may include incorporation of other imaging modalities, such as MRI or PET, further expanding its utility in multi-modal imaging research.

The lower image contrast of megavoltage computed tomography (MVCT), which corresponds to kilovoltage computed tomography (kVCT), can inhibit accurate dosimetric assessments. This study proposes a deep learning approach, specifically the pix2pix network, to generate high-quality synthetic kVCT (skVCT) images from MVCT data. The model was trained on a dataset of 25 paired patient images and evaluated on a test set of 15 paired images. We performed visual inspections to assess the quality of the generated skVCT images and calculated the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Dosimetric equivalence was evaluated by comparing the gamma pass rates of treatment plans derived from skVCT and kVCT images. Results showed that skVCT images exhibited significantly higher quality than MVCT images, with PSNR and SSIM values of 31.9 ± 1.1 dB and 94.8% ± 1.3%, respectively, compared to 26.8 ± 1.7 dB and 89.5% ± 1.5% for MVCT-to-kVCT comparisons. Furthermore, treatment plans based on skVCT images achieved excellent gamma pass rates of 99.78 ± 0.14% and 99.82 ± 0.20% for 2 mm/2% and 3 mm/3% criteria, respectively, comparable to those obtained from kVCT-based plans (99.70 ± 0.31% and 99.79 ± 1.32%). This study demonstrates the potential of pix2pix models for generating high-quality skVCT images, which could significantly enhance Adaptive Radiation Therapy (ART).

The impairment of inter-muscular coordination and changes in frequency components are two major pathological symptoms associated with knee injuries; however, an effective method to simultaneously quantify these changes has yet to be developed. Moreover, there is a need to propose a reliable automated system for identifying knee injuries to eliminate human errors and enhance reliability and consistency. Hence, this study introduces two novel inter-muscular coordination features: Dynamic Time Warping (DTW) and Dynamic Frequency Warping (DFW), which integrate time and frequency characteristics with a dynamic matching procedure. The support vector machine classifier and two types of dynamic neural network classifiers have also been used to evaluate the effectiveness of the proposed features. The proposed system has been tested using a public dataset that includes five channels of electromyogram (EMG) signals from 33 uninjured subjects and 28 individuals with various types of knee injuries. The experimental results have demonstrated the superiority of DFW and cascade forward neural network, achieving an accuracy rate of 92.03% for detection and 94.42% for categorization of different types of knee injuries. The reliability of the proposed feature has been confirmed in identifying knee injuries using both inter-limb and intra-limb EMG channels. This highlights the potential to offer a trade-off between high detection performance and cost-effective procedures by utilizing fewer channels.

The prevalence of Attention-Deficit/Hyperactivity Disorder among children is rising, emphasizing the need for early and accurate diagnostic methods to address associated academic and behavioral challenges. Electroencephalography-based analysis has emerged as a promising noninvasive approach for detecting Attention-Deficit/Hyperactivity Disorder; however, utilizing the full range of electroencephalography channels often results in high computational complexity and an increased risk of model overfitting. This study presents a comparative investigation between a proposed multi-headed deep learning framework and a traditional baseline single-model approach for classifying Attention-Deficit/Hyperactivity Disorder using electroencephalography signals. Electroencephalography data were collected from 79 participants (42 healthy adults and 37 diagnosed with Attention-Deficit/Hyperactivity Disorder) across four cognitive states: resting with eyes open, resting with eyes closed, performing cognitive tasks, and listening to omniarmonic sounds. To reduce complexity, signals from only five strategically selected electroencephalography channels were used. The multi-headed approach employed parallel deep learning branches-comprising combinations of Bidirectional Long Short-Term Memory, Long Short-Term Memory, and Gated Recurrent Unit architectures-to capture inter-channel relationships and extract richer temporal features. Comparative analysis revealed that the combination of Long Short-Term Memory and Bidirectional Long Short-Term Memory within the multi-headed framework achieved the highest classification accuracy of 89.87%, significantly outperforming all baseline configurations. These results demonstrate the effectiveness of integrating multiple deep learning architectures and highlight the potential of multi-headed models for enhancing electroencephalography-based Attention-Deficit/Hyperactivity Disorder diagnosis.

In radiotherapy treatment planning, the extrapolation of computed tomography (CT) values for low-density areas without known materials may differ between CT scanners, resulting in different calculated doses. We evaluated the differences in the percentage depth dose (PDD) calculated using eight CT scanners. Heterogeneous virtual phantoms were created using LN-300 lung and - 900 HU. For the two types of virtual phantoms, the PDD on the central axis was calculated using five energies, two irradiation field sizes, and two calculation algorithms (the anisotropic analytical algorithm and Acuros XB). For the LN-300 lung, the maximum CT value difference between the eight CT scanners was 51 HU for an electron density (ED) of 0.29 and 8.8 HU for an extrapolated ED of 0.05. The LN-300 lung CT values showed little variation in the CT-ED/physical density data among CT scanners. The difference in the point depth for the PDD in the LN-300 lung between the CT scanners was < 0.5% for all energies and calculation algorithms. Using Acuros XB, the PDD at - 900 HU had a maximum difference between facilities of > 5%, and the dose difference corresponding to an LN-300 lung CT value difference of > 20 HU was > 1% at a field size of 2 × 2 cm². The study findings suggest that the calculated dose of low-density regions without known materials in the CT-ED conversion table introduces a risk of dose differences between facilities because of the calibration of the CT values, even when the same CT-ED phantom radiation treatment planning and treatment devices are used.

The use of machine learning has seen extraordinary growth since the development of deep learning techniques, notably the deep artificial neural network. Deep learning methodology excels in addressing complicated problems such as image classification, object detection, and natural language processing. A key feature of these networks is the capability to extract useful patterns from vast quantities of complex data, including images. As many branches of healthcare revolves around the generation, processing, and analysis of images, these techniques have become increasingly commonplace. This is especially true for radiotherapy, which relies on the use of anatomical and functional images from a range of imaging modalities, such as Computed Tomography (CT). The aim of this review is to provide an understanding of deep learning methodologies, including neural network types and structure, as well as linking these general concepts to medical CT image processing for radiotherapy. Specifically, it focusses on the stages of enhancement and analysis, incorporating image denoising, super-resolution, generation, registration, and segmentation, supported by examples of recent literature.

Alzheimer's disease (AD) is a neurodegenerative disorder that challenges early diagnosis and intervention, yet the black-box nature of many predictive models limits clinical adoption. In this study, we developed an advanced machine learning (ML) framework that integrates hierarchical feature selection with multiple classifiers to predict progression from mild cognitive impairment (MCI) to AD. Using baseline data from 580 participants in the Alzheimer's Disease Neuroimaging Initiative (ADNI), categorized into stable MCI (sMCI) and progressive MCI (pMCI) subgroups, we analyzed features both individually and across seven key groups. The neuropsychological test group exhibited the highest predictive power, with several of the top individual predictors drawn from this domain. Hierarchical feature selection combining initial statistical filtering and machine learning based refinement, narrowed the feature set to the eight most informative variables. To demystify model decisions, we applied SHAP-based (SHapley Additive exPlanations) explainability analysis, quantifying each feature's contribution to conversion risk. The explainable random forest classifier, optimized on these selected features, achieved 83.79% accuracy (84.93% sensitivity, 83.32% specificity), outperforming other methods and revealing hippocampal volume, delayed memory recall (LDELTOTAL), and Functional Activities Questionnaire (FAQ) scores as the top drivers of conversion. These results underscore the effectiveness of combining diverse data sources with advanced ML models, and demonstrate that transparent, SHAP-driven insights align with known AD biomarkers, transforming our model from a predictive black box into a clinically actionable tool for early diagnosis and patient stratification.

This study introduces a novel optimization framework for cranial three-dimensional rotational angiography (3DRA), combining the development of a brain equivalent in-house phantom with Figure of Merit (FOM) a quantitative evaluation method. The technical contribution involves the development of an in-house phantom constructed using iodine-infused epoxy and lycal resins, validated against clinical Hounsfield Units (HU). A customized head phantom was developed to simulate brain tissue and cranial vasculature for 3DRA optimization. The phantom was constructed using epoxy resin with 0.15-0.2% iodine to replicate brain tissue and lycal resin with iodine concentrations ranging from 0.65 to 0.7% to simulate blood vessels of varying diameters. The phantom materials validation was performed by comparing their HU values to clinical reference HU values from brain tissue and cranial vessels, ensuring accurate tissue simulation. The validated phantom was used to acquire images using cranial 3DRA protocols, specifically Prop-Scan and Roll-Scan. Image quality was assessed using Signal-Difference-to-Noise Ratio (SDNR), Dose-Area Product (DAP), and Modulation Transfer Function (MTF). Imaging efficiency was quantified using the Figure of Merit (FOM), calculated as SDNR²/DAP, to objectively compare the performance of two cranial 3DRA protocols. The task-based optimization showed that Roll-Scan consistently outperformed Prop-Scan across all vessel sizes and regions. Roll-Scan yields FOM values ranging from 183 to 337, while Prop-Scan FOM values ranged from 96 to 189. Additionally, Roll-Scan (0.27 lp/pixel) delivered better spatial resolution, as indicated by higher MTF 10% value than Prop-Scan (0.23 lp/pixel). Most notably, Roll-Scan consistently detecting 2 mm vessel structures among all regions of the phantom. This capability is clinically important in cerebral angiography, which is accurate visualization of small vessels, i.e. the Anterior Cerebral Artery (ACA), Posterior Cerebral Artery (PCA), and Middle Cerebral Artery (MCA). These findings highlight Roll-Scan as the superior protocol for brain interventional imaging, underscoring the significance of FOM as a comprehensive parameter for optimizing imaging protocols in clinical practice. The experimental results support the use of the Roll-Scan protocol as the preferred acquisition method for cerebral angiography in clinical practice. The analysis using FOM provides substantial and quantifiable evidence in determining the acquisition methods. Furthermore, the customized in-house phantom is recommended as a candidate to optimization tools for clinical medical physicists.

This study investigates how deep learning (DL) can enhance ovarian cancer diagnosis and staging using large imaging datasets. Specifically, we compare six conventional convolutional neural network (CNN) architectures-ResNet, DenseNet, GoogLeNet, U-Net, VGG, and AlexNet-with OCDA-Net, an enhanced model designed for [¹⁸F]FDG PET image analysis. The OCDA-Net, an advancement on the ResNet architecture, was thoroughly compared using randomly split datasets of training (80%), validation (10%), and test (10%) images. Trained over 100 epochs, OCDA-Net achieved superior diagnostic classification with an accuracy of 92%, and staging results of 94%, supported by robust precision, recall, and F-measure metrics. Grad-CAM ++ heat-maps confirmed that the network attends to hyper-metabolic lesions, supporting clinical interpretability. Our findings show that OCDA-Net outperforms existing CNN models and has strong potential to transform ovarian cancer diagnosis and staging. The study suggests that implementing these DL models in clinical practice could ultimately improve patient prognoses. Future research should expand datasets, enhance model interpretability, and validate these models in clinical settings.