This study evaluates the influence of connectivity analysis parameters on the ability of weighted Phase Lag Index (wPLI)-based networks, derived from stereoelectroencephalography (SEEG) recordings, to represent the epileptogenic network. It also identifies the parameter combinations that most accurately delineate the epileptogenic zone. By comparing network metrics obtained under different connectivity configurations with surgical resections and postoperative outcomes, we aimed to establish a framework for selecting connectivity analysis parameters in epilepsy surgery planning. This optimized framework could improve the precision of stimulation zone identification, thereby refining resection or ablation strategies and potentially leading to better surgical outcomes and improved quality of life for patients with refractory epilepsy, while complementing clinical judgment. Ictal SEEG recordings from 28 patients with refractory epilepsy were analyzed. Functional networks were generated using wPLI, varying frequency bands, start/end analysis times, and epoch durations. Connectivity degree and betweenness centrality were then compared with surgical outcomes using the R coefficient and Mann-Whitney test. Optimal configurations showed significant overlap between resected nodes and key network nodes in patients with successful surgical outcomes. Both the theta (4–8 Hz) and high-frequency (160–220 Hz) bands were represented among the most effective configurations, most commonly combined with analysis windows beginning at or shortly before ictal onset; an epoch duration of 0.5 s was the most frequent choice among these configurations. Rather than assessing a single connectivity configuration, this study establishes a framework for selecting connectivity analysis parameters in ictal SEEG. The identified configurations generated network representations that were more consistent with surgically validated epileptogenic regions, providing methodological guidance for future network-based surgical planning.
Medical imaging has been transformed by Artificial Intelligence (AI) and Deep Learning (DL). Yet, multi-hospital deployment remains limited by patient privacy concerns, heterogeneous data distributions, and insufficient model interpretability, which affect regulatory approval and clinical trust. This study proposes a regulatory-grade Federated Learning (FL) framework for secure, interpretable, and generalizable collaborative medical imaging. The proposed framework integrates Slicing Window Adaptive Kalman Filtering (SWAKF) for image denoising, Structured Multi-Modal Autoencoder Attention Fusion (SMAAF) for feature representation, and adaptive federated aggregation to address non-IID data across hospitals. Patient privacy is preserved using secure aggregation, differential privacy, and encryption, while Grad-CAM, SHAP, and LIME provide model interpretability. The proposed framework outperformed Vision Transformer, AlexNet, FedAvg, and FedProx on Brain Tumor and Alzheimer's MRI datasets. It achieved 96.1
Advances in image acquisition and computational processing technologies have significantly transformed histological and hematological analysis. However, to maximize the clinical impact of these technologies, it is essential to develop automatic or semi-automatic segmentation techniques. The objective of this work is to develop an automatic multi-class segmentation technique that achieves competitive performance while minimizing computational costs. The proposed technique consists of an unsupervised algorithm that models a neutrosophic image by combining only two features: the luminance channel L* from the CIELUV color space and the HH sub-band of the discrete wavelet transform obtained from the grayscale image. The neutrosophic components are enhanced to promote robust clustering and segmentation, especially in regions of uncertainty. Finally, morphological operations are applied as a post-processing stage. Experimental results demonstrate strong performance in multiclass segmentation, with average standard index values above 90
This study presents a preliminary application of Physics-Informed Neural Networks (PINNs) for computing breast temperature distributions using a simplified three-dimensional geometry. The results were validated against experimental data reported by Gautherie (1980) and numerical results from Das and Mishra (2015). A standard (vanilla) PINN architecture was employed. The mathematical model adopted was the Bioheat Transfer Equation (BHTE), also known as Pennes’ equation, which was solved using a standard PINN implementation. A simplified three-dimensional geometry was considered, in which the tumor was modeled as a sphere. Boundary conditions of the first and third kinds were applied. The PINN successfully reproduces the overall thermal trend across the domain and captures the localized temperature elevation associated with the presence of a tumor. Acceptable levels of error were observed when the computed temperatures were compared with experimental and numerical data reported in two independent studies, Gautherie (1980) and Das and Mishra (2015). PINNs provide a flexible framework for incorporating physical laws and heterogeneous tissue properties, enabling meaningful comparisons with classical thermographic datasets and established numerical simulations. In addition, they show potential for patient-specific parameter estimation. Further improvements may be achieved by combining the PINN architecture with other neural network approaches.
This innovation aims to assist specialists in selecting appropriate treatments to enhance patient outcomes and prolong their lives. Initially, the required 3D MRI is synthetically initiated by the Attention-based Stacked Conditional Generative Adversarial Network (A-SCGAN). The synthetically generated images are used to mitigate class imbalance in the brain tumor detection process. The synthetically generated 3D images are passed to the segmentation phase, leveraging with 3D Trans Dilated Mobile Unet (3D-TDMobileUNet). The segmented images are given to the Adaptive MobilenetV2 (AMV2) model for classifying the brain tumor into various classes. Here, the parameter tuning takes place via Enhanced Exploration-based Carpet Weaver Optimization (EECWO) to enhance tumor classification performance. Experimentation is conducted on synthetically created 3D images and after generating 3D images, the quality of synthetically generated images is compared over the original images to prove the same. The tumors identified in the segmented images are classified into discrete types with the support of the AMV2 model. This proposed model utilizes depth wise separable convolutions, which perform a filtering process to extract relevant features and then produce classified outcomes depends on these features extracted. The developed AMV2 model achieves high classification accuracy, especially when trained on well-prepared datasets that reflect the variations of the MRI images. The accuracy of the developed EECWO-MV2 model reached 93.1
Oral squamous cell carcinoma (OSCC) is the most common oral cancer, with high mortality rates, making early detection crucial. Malignant transformation is preceded by oral potentially malignant disorders (OPMDs), with leukoplakia being the most frequent. Epithelial dysplasia (ED) is a key histological feature for predicting the risk of progression to OSCC. However, its diagnosis relies on subjective visual assessment by pathologists, which is influenced by experience and emotional state. This study aims to develop a Machine Learning (ML), based methodology to assist in ED detection in leukoplakia lesions. The proposed methodology integrates histopathological image analysis with complementary patient data. Epithelial regions were selected for cutout extraction based on pathologist knowledge and combined with risk factors from patient data. A multilayer perceptron artificial neural network (MLP-ANN) was trained and evaluated. Performance was compared with a pre-trained convolutional neural network (ResNet-50V2) using the McNemar test. The MLP-ANN achieved an area under the curve (AUC) of 0.9484. No statistically significant difference was found between the proposed method and the pre-trained CNN. However, the proposed approach was approximately 9.7 times faster in computational time. The proposed ML-based methodology demonstrated competitive classification performance along with improved computational efficiency. Refining the confidence criterion for intermediate classification results may further enhance its diagnostic accuracy, supporting pathologists in ED detection.
Thermal necrosis following bone drilling is a serious threat to orthopedic surgeries as it can cause irreversible damage to bone cells and failure of fracture treatment. Part of the bone temperature rise is due to chip formation while the other part is related to the heat sources of drill bit-hole wall friction as well as chip-hole wall friction. The present study has examined reduction of frictional heating by changing the drill bit design. Drilling tests were performed on 31 states of drill bits with different designs under conditions of rotational speed of 1000 rpm, feed rate of 50 mm/min, and hole depth of 8 mm on the bovine femur. The change in the diameter and length of the different sections of the drill bit had a significant influence on the bone temperature rise; in 21 cases it led to a reduction, while in 6 cases it resulted in an increase in the temperature compared to the result of the standard drill (Tm = 24 °C). Further, using statistical analysis through Minitab software, a statistical model of bone temperature rise was developed based on the drill geometry and optimal values for the drill bit geometry were extracted. The validation test performed on the optimal drill bit revealed an acceptable agreement of its result (Tm = 8.6 °C) with the value predicted by statistical analysis (Tm = 8.8 °C) as well as the possibility of preventing thermal necrosis by applying this new design for the drill bit, compared with other methods of drilling.
Heart rate (HR) measurement is essential for assessing physical fitness and diagnosing various diseases. Photoplethysmogram (PPG) is a convenient method for HR estimation due to its simplicity. However, PPG signals recorded from the wrist are highly susceptible to motion artifacts. Traditional noise cancellation techniques, such as adaptive filtering, are sensitive to the reference signal choice. This paper presents a novel approach using switch-mode decomposed gyroscopic and accelerometer signals as reference inputs for a three-stage Least Mean Square (LMS) filter-based noise cancellation system. The signal decomposition is done using Multivariate Variational Mode Decomposition (MVMD), allowing more flexible noise handling. By using decomposed modes as noise references, this method enables separate treatment of different motion artifact components by adjusting the LMS filter parameters, which is not feasible with raw accelerometer or gyroscope data. The proposed method was tested on a publicly available PPG dataset with recordings from 24 subjects. The method achieved an Average Absolute Error (AAE) of 2.3, a Standard Deviation of Absolute Error of 2.05, and an Average Relative Error of 4.43
Cutaneous melanoma accounts for most skin cancer deaths due to its high metastatic potential, making early detection essential. This study investigates whether color indices can enhance dermatoscopic images and improve benign–malignant classification. Three Browning Indices (Aimonino, Fetuga, Lunadei2) and one Vegetation Index (VI) were combined to form composite chromatic representations of RGB images. A Power-Log Based Histogram Equalization technique was applied for contrast enhancement. Four CNN architectures (VGG19, ResNet50, NASNetMobile, EfficientNetV2B0) were trained on the Javid Melanoma Dataset (10,605 images). Performance was assessed using AUC, F1-score, ROC/PR curves, ablation studies, and statistical significance tests. Chromatic-index preprocessing improved performance across all models. NASNetMobile and EfficientNetV2B0 achieved the best results, with AUCs of 96.20
This study introduces a new method for classifying Diabetic Retinopathy (DR) with enhanced accuracy, focusing on addressing the limitations of existing approaches. DR classification is still a difficult task due to the presence of various complex and overlapping lesions in the retinal images. The primary goal is to develop an extremely effective model that can identify the different phases of diabetic retinopathy (DR) very accurately from the fundus images, therefore reducing the misclassification rates and helping in the very early diagnosis. The focus of our inquiry is on the determination of whether the new Hierarchical Auto-Associative Inception Polynomial Transformer Convolutional Neural Network with Snake Optimizer (HAutoAIPTCNNet + SO) will result in a significantly higher level of accuracy when compared with the existing methods. Fundus images from the EyePACS and DIARETDB1 datasets are processed using Gradient Domain Guided Filtering (GDGF) for enhanced contrast, noise reduction, and normalization. Segmentation of DR-affected regions is achieved with the EfficientNet and Cascaded Visual Attention Network (ENet-CVAN) framework. The classification process then employs the Hierarchical Auto-Associative Inception Polynomial Transformer Convolutional Neural Network (HAutoAIPTCNNet) further refined through the Snake Optimizer (SO). The HAutoAIPTCNNet + SO model, developed in Python, features both hierarchical and polynomial transformations for the precise imaging of retinal characteristics. The suggested HAutoAIPTCNNet + SO approach reached a classification accuracy of 99.8
The aging population has led to an increase in chronic degenerative diseases, such as Alzheimer’s disease (AD) and frontotemporal dementia (FTD). Future projections indicate that this increase will be more pronounced, with cases tripling by 2050. Therefore, it is necessary to identify patterns of brain activity related to dementia to assist in earlier diagnosis. In this study, we propose to compare the performance of Machine Learning (ML) models across different databases in patients with dementia and elderly individuals using EEG data. We investigated various machine learning models using two distinct EEG databases: OpenNeuro (AD, FTD, and healthy controls) and GeronGnosis (mild cognitive impairment - MCI, AD, FTD, and healthy controls), the latter developed by the researchers and presented here for the first time. After extracting different signal attributes, we applied the Bayes Network, Random Tree, Random Forest, and SVM models and compared their best performances. The Random Forest model with 500 trees demonstrated the best classification performance, achieving an accuracy of 98
Performance of centrifugal blood component separators in apheresis is highly dependent on operational parameters such as rotational speed and inlet flow rate. However, fixed-protocol systems limit multi-objective optimization and adaptability to donor variability. To investigate the coupled effects of rotational speed and flow rate on separation performance and to identify optimal operating conditions using a programmable centrifugal separator. A programmable separator was evaluated across rotational speeds of 500–9000 RPM and flow rates of 1–200 mL/min. Experiments were conducted with five replicates per condition (n = 5). Key performance metrics included separation efficiency, plasma purity, platelet collection rate, and temporal efficiency. Data were analyzed using ANOVA and response surface methodology for multi-objective optimization. Maximum separation efficiency (93
Through devices and techniques developed in instrumentation, it is possible to create low-cost systems to diagnose variables related to human movement. These systems can assist in rehabilitating individuals, analyzing pathologies, and enhancing performance. This work presents the development of an inertial sensor network dedicated to characterizing the spatial dynamics of the ankle-foot segment and analyzing asymmetries in human gait. The system consists of four Micro Electro-Mechanical Systems (MEMS) inertial modules strategically positioned on the ankle-foot segment. Each module includes a triaxial accelerometer and gyroscope, and their data are controlled through an Inter-Integrated Circuit (I²C) bus using a microcontroller. Anomalies in human gait and observations of how the ankle-foot segment behaves during the phases of the gait cycle were determined using the system. Factoring the maximum angles a segment can reach on each global axis, the X and Y axes exhibited greater relative deflections due to center of pressure variation and gait asymmetries. Taking into account the XY plane that makes up the transverse region of the foot, the resulting acceleration presented angles close to 25° in relation to the tibial axis during initial contact. Our results indicate that inertial sensors can be applied to determine the spatial dynamics of the ankle-foot segment during gait. The X and Y axes exhibited greater relative deflections .Given the commercial systems high cost, the use of low-cost capacitive sensors offers a good alternative for assessing human movement. This study demonstrated that it is possible to assess the spatial dynamics of the ankle-foot segment during gait using a low-cost system, enabling more physiotherapy clinics to access this technology.
Medical device development (MDD) is a highly intricate process. The dynamic nature of regulatory frameworks further complicates the MDD process. As medical technologies continue to evolve, regulatory systems introduce increasing challenges that must be addressed through the development lifecycle to ensure safe medical devices. This study aims to identify challenges in implementing regulations in medical device development. A bibliometric analysis of 124 papers was conducted to investigate research gaps in the field of medical device development. We analyzed literature from Scopus, PubMed, IEEE Xplore, and Google Scholar using VOSviewer to identify regulatory barriers in medical device development, methods and technologies that address these challenges and assess their integration with design and operational aspects. The analysis reveals underexplored regulatory dimensions such as regulation complexity, costs, innovation, requirement elicitation and collaboration issues. Our findings highlight the need to simplify regulatory compliance using advanced tools which can enhance the safety and efficiency of medical device development processes.
Each year, pneumonia is responsible for more than 2.2 million deaths. Although deep learning has helped to analyze chest X-rays (CXRs), current systems approach segmentation and classification separately. They also use basic combination methods and give clinically inappropriate overconfident predictions. To formulate and verify a pneumonia detection framework with the clinically safe deployment of anatomically informed segmentation, cross-attention fusion, and ensemble calibration. We created a MedSAM-guided Transformer enhanced U-Net segmentation framework combined with cross attention from dual CheXFound branch encoders. Training was done on the RSNA and NIH ChestX-ray14 datasets (n=28,526), and external validation was done on the CheXpert and COVIDx datasets. The ensembles of five models with temperature scaling improved calibration. Key metrics to evaluate the performance of the models include AUROC, expected calibration error (ECE), and radiologist Grad-CAM++ annotations for interpretability validation. On combined test sets (n=6,112), the framework achieved AUROC 0.973 [95
Deep learning models require large and well-prepared datasets to achieve reliable performance. However, publicly available cardiac magnetic resonance (CMR) datasets often lack the necessary uniformity and quality for direct use in training. This paper proposes a comprehensive image preprocessing methodology aimed at improving the performance of deep learning models, specifically a 2D U-Net, for myocardium segmentation. The proposed preprocessing pipeline focuses on two main objectives: (i) enhancing image quality through contrast and brightness adjustment, gradient equalization, anisotropic diffusion filtering, and CLAHE; and (ii) accurately identifying the region of interest (ROI) containing the myocardium using a Hough Transform–based approach. After ROI detection, images are cropped to reduce dimensionality while preserving relevant anatomical structures. The methodology was evaluated using three public CMR datasets, comparing segmentation performance with and without preprocessing. The U-Net model was trained under consistent conditions, and performance was assessed using DICE and Hausdorff metrics. The preprocessing pipeline produced more uniform, higher-quality images and achieved robust ROI localization, with a 100
Real-time Healthcare management assimilates the Internet of Things (IoT) and its allied computing techniques for granting optimal services to patients and users. The data processing and computing techniques inherited in healthcare management help to improve the swiftness and responsive nature of the service providers. To improve the swiftness of large medical data handling in real-time healthcare management, this manuscript presents an interactive healthcare service-responsive model (IHSRM). This model is designed using deep learning to identify the different classes of service requirements and their associated data. Based on the service demands, the available data is processed promptly, reducing the prolonged wait time. Initially, the proposed model distinguishes the patient requirement and extracts the part of large data eligible for computation. The data extraction process is performed with the knowledge of responsive interaction and swift response of the current and previous services. This helps to improve the service reliability for any type of healthcare application. The proposed computing/ processing model is analyzed using the metrics of response failures, data identification, response delay, and service reliability.
Medicare Functional Classification Levels (K-levels) are widely used to grade community mobility in people with transfemoral amputation (TFA), yet their relationship with quantitative gait biomechanics remains unclear. To compare gait kinematics and kinetics between K2 and K3 individuals with unilateral TFA, and between prosthetic and sound limbs. Seventeen adults with unilateral TFA (8 K2, 9 K3) walked on an instrumented treadmill. Joint-angle and joint-moment waveforms (hip, knee, ankle) were compared using Statistical Parametric Mapping (SPM; α = 0.05). Participant-level k-means clustering was performed on swing-phase peaks from both limbs (8 features in total), with silhouette analysis used to select the number of clusters. SPM detected no K2–K3 differences in any joint-angle waveform for either limb. On the prosthetic limb, between-group differences were limited to joint-moment regions: hip moment at 55–58
