Early detection of lung cancer is crucial as it significantly improves survival rates by facilitating timely and effective treatment. Lung cancer often begins as bronchial lesions developing along the airway walls. Bronchoscopy is the minimally invasive method of choice for detecting such lesions. Currently, three complementary bronchoscopic video modalities have been utilized for this purpose: white-light bronchoscopy (WLB), narrow-band imaging (NBI), and autofluorescence bronchoscopy (AFB). Unfortunately, current practice forces the clinician to manually examine each video source and later interactively correlate the results of these exams to make final lesion decisions. Because of the lack of effective tools for multimodal endoscopic video analysis, this proves to be an extremely time-consuming, error-prone process, making it impractical for common clinical use. To address this problem, we propose a multimodal video analysis and synchronization system that enables efficient analysis of multimodal bronchoscopic videos for early cancer lesion detection. The system provides methods for planning and guiding a straightforward multimodal airway exam through the major airways. Subsequent video processing methods then draw on deep-learning-based techniques to identify candidate single-mode bronchial lesions. Next, a synchronization/registration pipeline registers all bronchoscopic video data to a reference 3D airway tree model derived from a patient's X-ray computed tomography (CT) scan. This finally facilitates interactive graphical visualization and interaction with all processed multimodal data. Results with lung cancer patient studies indicate the system's promise for efficient, effective video analysis and lesion detection.
Emotional recognition is a fundamental task towards the true understanding of the instantaneous brain state. Promising results using electroencephalography (EEG) for emotion detection have been presented, yet most approaches are constrained to laboratory settings with limited ecological validity. To address this, we employed an immersive curved-screen projection environment, chosen for its ability to preserve ecological validity while minimizing EEG signal artifacts commonly induced by virtual reality headsets. We present a novel EEG dynamic functional connectivity (dFC) framework for emotional recognition by using the Neural Gas algorithm for capturing the rapid EEG alterations in an immersive projection environment. EEG responses from 33 participants were collected during video stimulus presentations featuring five emotional environments: calm beach, calm nature, shark attack, rollercoaster, and night walk. The EEG acquisition employed a Unicorn Hybrid Black 8-channel system, while the dFC is computed by using weighted Phase Lag Index (wPLI) through a shifting window-based approach. The symbolic time series are built through Neural Gas, and chronnectomics are then calculated. Statistically significant differences were observed in the brain state flexibility across emotional conditions, with calm conditions showing enhanced flexibility (Nature: 0.93, Beach: 0.93) compared to stress conditions (Shark Attack: 0.75, Night Walk: 0.77). The flexibility index emerged as the primary discriminator of emotional states, with all stress versus calm comparisons achieving statistical significance $(p<0.001)$. Mean dwell time and occupancy entropy provided complementary insights into network stability and state diversity. This advanced EEG DFG study in an immersive real-world environment demonstrates superior discrimination of emotional states compared to traditional laboratory settings, establishing new methodological standards for ecological emotion recognition research.
In response to an infection, lymph nodes may swell, indicating heightened activity as the body mobilizes its defenses. This study examined artificial intelligence approaches for detecting and localizing cervical lymph nodes during palpation, capturing hand palpation and gesture analysis, thereby facilitating early disease identification. The Anatomical study states that between one-fifth and one-sixth of all the nodes in the body are located on either side of the neck. This study investigated the detection of lymph nodes in real-time video acquisition using YOLO-v8. The research leveraged data from the Mugshots images datafig 1set from Kaggle and GitHub, focusing on both the accuracy of detection and the response time of the model. The YOLO-v8 model offers the optimal balance between accuracy and response time, making it the most suitable choice for lymphatic nodes detection in real-time video acquisition. YOLO-v8 was evaluated on the collected data, achieving a global detection precision of 89.3 % with a response time of 20 ms. During video acquisition, the palpation hand is tracked using Mediapipe, and the corresponding palpation gesture is subsequently evaluated by MobileNetV3. This approach allows for efficient and automated analysis of palpation techniques.
Depression is a mental disorder that can lead to self-harm or suicidal thoughts if left untreated. Clinicians face the challenge of determining the most effective therapeutic approach to depression. Selective Serotonin Reuptake Inhibitors are widely prescribed depression therapies, but their response rate is only around 50 %, which is relatively low compared to the treatment success of other mental diseases. To address this issue, a novel classification framework is introduced to build a computer-aided decision system that predicts the outcome of depression therapies. This proposed system utilizes novel systematic extraction and selection of time-domain features. Our methodology is not only effective for EEG subject classification but also widely applicable to similar EEG studies benefitting researchers working in cognitive, affective, and clinical neuroscience. In this 30 -subject pilot study, the multi-channel EEG signals are denoised using low and high-pass filters. Subsequently, feature extraction models are applied to the channels to extract the generalized pattern of the EEG data. Overlap coverage for moving segments was varied from $\mathbf{0} \boldsymbol{\%}$ to $\mathbf{7 5 \%}$, and four feature selection algorithms were evaluated. To avoid bias in the results, the classification models are trained and validated using a leave-one-subject-out (LOSO) crossvalidation strategy. To prevent data leakage during feature selection, the test subject in each fold was excluded from the dataset prior to feature ranking. Randomized experiments repeated 30 times indicate that models utilizing under 50 features consistently reach an average balanced accuracy of 95%. In conclusion, this novel approach demonstrates its effectiveness through highly accurate and nonbiased classification results. The top-ranking features, dominated by energy and variability metrics from right frontal (F8), central (Cz), and occipital (O2) channels, highlight the critical role of fronto-central and posterior cortical dynamics in characterizing depression-related EEG biomarkers.
As service animals play a vital role in supporting individuals with physical, sensory, or cognitive impairments, communication between humans and their animal companions becomes needed. Emotional and behavioral signals conveyed through animal vocalizations often remain difficult to interpret without specialized tools. This paper investigates the effectiveness of deep learning architectures for real-time interpretation and classification of dog vocalizations into emotional and behavioral states such as alertness, aggression, and contentment. The developed system uses a fully automated approach and utilizes Mel-frequency cepstral coefficients (MFCCs) for feature extraction and a convolutional neural network (CNN) for classification. The model is trained using a custom dataset of dog vocalizations divided into multiple categories. To make the system user-friendly and context-aware, a real-time interface is developed where users can record audio on command and receive spoken feedback. The system can deliver the interpretation in multiple languages, including English and Romanian, through a speech synthesis engine, in order to enhance accessibility and usability across diverse user groups. This porTable, assistive tool aims to support individuals who rely on service animals by improving communication, offering emotional reassurance, and strengthening the human-animal bond. In particular, the system is designed to assist people with physical, sensory, or cognitive disabilities by providing a clearer understanding of service animal behavior and emotional cues. By integrating bio-signal processing with accessible interface design, the proposed system contributes to the development of intelligent, context-aware technologies for assistive bioengineering applications. The system achieved a test accuracy of 94.12% on the three-class classification task (dog bark, growl, grunt), outperforming previous approaches and demonstrating generalization capabilities. These findings demonstrate that the proposed system can facilitate communication between individuals with disabilities and their service animals.
The temporomandibular joint (TMJ) is a complex anatomical and biomechanical structure critical to mastication, speech, and overall craniofacial function. This review consolidates finite element analyses of the temporomandibular joint published in the last fifteen years. The content of over one hundred studies is organized into three topics: (a) Orthodontic treatment, which reshapes pressure and tension in the area; (b) craniofacial deformities—chiefly mandibular projection; and (c) joint dysfunctions, in which temporary or permanent disc displacement and disc perforation significantly redefine contact pressure. Together, these models clarify how mechanical responses differ with treatment, anatomy, and pathology. The study also reviews treatment approaches, including occlusal splints and prosthetic joint replacements, in the context of biomechanical insight provided by finite element (FE) analysis. Prompted by the frequent occurrence of TMJ-related injuries among cyclists and motorcyclists, this study presents a concise synthesis of epidemiology, injury mechanisms, and current human body modeling gaps. It contributes to the understanding of TMJ biomechanics and supports the design of more effective diagnostic and therapeutic interventions.
Deep neural networks often require large-scale, accurately labeled datasets to perform well, but in practice the labels are frequently corrupted by noise in medical imaging, especially instance-dependent noise. In this work, we propose a novel framework to address instance-dependent label noise by integrating three key components: (i) self- supervised pretraining using SimCLR to learn robust, noise-agnostic feature representations; (ii) an iterative pseudo-label refinement strategy employing a stage-wise consensus mechanism to progressively correct mislabeled samples; and (iii) a softmax-weighted crossentropy loss that dynamically down-weighs uncertain predictions. We validate our approach on benchmark datasets such as CIFAR10 and CIFAR-100 corrupted with synthetic noise at 20 %, 30 % and 50 % levels, demonstrating significant improvements over state-of-the-art methods. We further validated our method on Chest X-rays medical imaging datasets.
Heart failure is one of the leading causes of death among cardiovascular diseases. The detection of heart failure requires invasive and time-consuming procedures, and it is limited by the availability of specialized facilities. Non-invasive procedures are necessary for the early and accurate prediction of heart failure, and voice is considered one of the successful biomarkers for disease prediction. The accurate prediction of heart failure requires multiple voice recordings and diverse voice patterns to reflect the physiological changes associated with the disease. The high-dimensional feature vector extracted from different voice recordings is used in machine learning for heart failure prediction. In our research, we carefully selected 5 different voice recordings and identified the most frequent and important features during multiple runs with varying numbers of feature constraints. For feature selection, we used a genetic algorithm to select features that maximize the prediction accuracy from the pool of features for all tasks. We introduced a multi-task feature balancing strategy in a genetic algorithm for feature selection and mutation. This strategy ensures the equal contribution of all the voice recordings and promotes diversity of features. Using a multi-task feature balancing-based genetic algorithm, the most frequent features are ranked and reported. The results show that we achieve high accuracy with a low-dimensional feature vector compared to a highdimensional feature vector for all tasks.
This work presents a 2D human model and its use in a novel gait analysis framework, which only requires a stereo camera to produce impressive results for the inverse kinematics, inverse dynamics, as well as ground reaction forces during gait. The model is designed to resemble the human body in the sagittal plane, with anatomical landmarks used as keypoints in the inverse kinematics calculations that yield accurate estimates of the joints' motion during gait. The gait dynamics are formulated in compact form, allowing the simultaneous estimation of internal joint torques, as well as ground reaction forces via the solution of a fully-defined system of linear algebraic equations. The proposed framework offers an affordable alternative to costly gait analysis systems, and can have various applications in robotics and in biomechanics.
Diffuse Axonal Injury (DAI) is a primary cause of prolonged disability and high mortality among Traumatic Brain Injury (TBI) patients. Pathophysiological disorders and behavioral implications of DAI can be evaluated through a proper understanding of TBI mechanism. In this study, we examine the translational acceleration-induced DAI during direct head impacts using a validated high fidelity finite element human head model. The segmented brain model characterizes the DAI mechanism and reliably determines the stress-strain distribution across white matter. Intracranial pressure distribution is not affected by the inhomogeneous nature of cerebral tissue. However, deep white matter contours at the white-grey junction contribute substantially to the focal lesion formation across intracerebral regions.
Accurate classification of lung cancer subtypes from histopathological images is challenging due to limited labelled data and high visual similarity between classes. This limitation is especially critical in clinical diagnostics, where timely and accurate classification can impact treatment decisions. To address this, LungCLR presents a comprehensive evaluation of a two-stage framework that combines contrastive self-supervised learning with EfficientNet-B3 for effective feature extraction. In the first stage, the SimCLR framework is employed to pretrain the encoder on unlabeled histopathology images, enabling it to learn meaningful representations by distinguishing subtle morphological variations. A projection head is incorporated to optimise the NT-Xent loss during training. In the second stage, a lightweight classification head is attached and fine-tuned using small labelled subsets, as few as $\mathbf{1 0 0}$ samples per class. Finally, partial end-to-end fine-tuning is applied to further enhance performance. LungCLR is evaluated on the LC25000 dataset, which includes three tissue categories: benign, adenocarcinoma, and squamous cell carcinoma. Using the full dataset, the proposed model achieves an impressive accuracy of 99.97 %, outperforming previous state-of-the-art methods. Importantly, even under low data conditions, the model performs robustly, reaching 90.03% accuracy with only 100 labelled samples per class. These results highlight the effectiveness of established contrastive learning techniques when carefully applied in a clinically relevant, lowdata setting.
Accurate detection of clinically significant prostate cancer (csPCa) remains a major challenge in prostate MRI interpretation. Radiomics provides a noninvasive approach by extracting quantitative descriptors from imaging data, such as Apparent Diffusion Coefficient (ADC) maps. In this study, we systematically benchmarked radiomics-based machine learning pipelines across two validation scenarios: (1) repeated cross-validation on a multicenter dataset, and (2) nested cross-validation with external testing on the public PROSTATEx dataset. We evaluated eight feature selection methods, fifteen classifiers, and the impact of ComBat harmonization. The best-performing pipeline-Recursive Feature Elimination (RFE) with 20 features and a Random Forest classifier, without ComBat-achieved an internal AUCPR of $0.880 \pm 0.061$ and an external AUC-PR of 0.717, with an F1-score of 0.748. Although ComBat improved calibration (F1score) in some cases, its effect on external discrimination was limited. Notably, we found that selecting 20 features yielded optimal generalization performance, supporting the “rule of 10 ” guideline for feature-to-sample ratio. These findings highlight the importance of rigorous feature selection and classifier design in developing robust and generalizable radiomics models for csPCa detection.
In this study, we present preliminary results from the evaluation of an automated real-time stress detection approach during cognitively demanding human-computer interaction. The proposed method relies on heart rate variability (HRV) analysis derived from photoplethysmography (PPG) data collected via a wireless wearable sensor. A personalized HRV threshold is established during a Stroop task of escalating difficulty and subsequently used for binary classification between calm and stressed states. The stress-detection approach is evaluated through an experimental protocol employing a novel dynamic biofeedback serious game (SG) and continuous post-game self-annotations of perceived stress levels. Fourteen individuals participated in the study, with three excluded due to sensor-related issues. The collected data were analyzed in two directions-absolute values and ordinal-centric measuresexamining associations between the detected state and userreported annotations. Statistically significant differences between calm and stressed game segments were observed for the mean ($\mathbf{p}=0.002$), median ($\mathbf{p}=0.003$), and trapezoidal interval $(\mathbf{p}=\mathbf{0. 0 0 2})$ of the annotation values. Ordinal analysis further confirmed this relationship, with positive Spearman rank correlations for the same features ($\mathbf{p}=\mathbf{0. 0 0 2}$) and consistent directionality in ten of eleven participants. These preliminary findings underscore the potential of the proposed thresholding approach and motivate further refinement of rule-based stress detection systems.
Early diagnosis of Alzheimer's disease (AD) is essential for timely intervention and effective care. This paper examines handwriting analysis as an accessible and non-invasive way of early detection by focusing on the comparison of raw handwriting images and tabular features. Two data sets were considered, the DARWIN handwriting dataset, which contains raw images and tabular data from pen movement, and the Alzheimer's disease dataset (ADD), which presents the patient's history and cognitive assessments in tabular format. A range of classification methods including machine learning (Random Forest, SVM, and XGBoost) was tested on tabular data from the two datasets, a deep learning Swin Transformer for image classification, and a multimodal approach that integrated both. Random Forest outperformed other models on DARWIN tabular data $(83.03 \% \pm 1.18)$, while XGBoost was the best on ADD $(83.53 \% \pm 3.44)$. The Swin Transformer also performed consistently on handwriting images $(80.02 \% \pm 0.87)$, capturing features associated with stroke tremors and fluency, as well as other visual aspects of the dataset. A late fusion model incorporating both modalities achieved the highest overall accuracy of $89.15 \% \pm 1.73$, showing that the image and the tabular features produce a complementary diagnostic value. These results indicate that handwriting includes fine neuromotor features related to early AD that can surpass clinical conventional data. We also present ablation studies on task order with respect to image training and end-to-end multimodal learning. These findings provide further evidence of the benefits of modular fusion in situations where data is restricted. Handwriting samples have the potential to become a useable and scalable resource in AD screening because of their low cost, ease of collection, and acquisition logistics, which even accommodate home-based settings.
The number of artificial intelligence (AI) systems developed for skin lesion classification has grown rapidly, notably driven by the release of publicly available datasets, such as those from the ISIC Archive. While this accessibility has allowed many research groups, often without a medical background, to train deep learning algorithms under ideal conditions, these often fail to maintain performance in real-world clinical environments where variations in image quality reveal limited generalizability. To address this gap, we propose a reinterpretation of the traditional lesion classification task by prioritizing clinical urgency and categorizing lesions into three priority levels rather than assigning a specific diagnosis. We developed a customized ConvNeXt-Tiny model, implemented specific techniques to handle class imbalance, and trained it on a combination of two public datasets, ISIC 2019 and Hospital Italiano de Buenos Aires, and a large private real-world dataset from Virgen Macarena University Hospital (Seville, Spain). In parallel, an exhaustive investigation of existing AI systems for skin lesions was also carried out, and two representative models were finally selected for adaptation to the task. Although both systems performed effectively in their original diagnostic contexts, a considerable drop in accuracy was observed when evaluated on our threelevel framework. In contrast, our proposed system demonstrates significant performance and specifically achieves improvements in accuracy from 13 % to 21 % at the evaluated priority levels. These results highlight the importance of adapting AI tools to realistic triage tasks and confirm the potential of our method as a possible practical assistance tool in dermatology.
In medical ultrasound, segmentation of the pubic symphysis and fetal head is critical for automating fetal and maternal assessments, yet it remains challenging due to anatomical variability, low contrast, and imaging artifacts. In this work, we present DAUNet, an efficient and lightweight segmentation model designed for compute efficient clinical use in ultrasound workflows. The architecture builds upon the UNet backbone and introduces two key innovations: Deformable V2 Convolutions for capturing non-rigid anatomical boundaries, and SimAM-an attention mechanism that enhances spatial saliency without adding parameters. We validate DAUNet on the FHPS dataset, which involves the simultaneous detection of fetal head and pubic symphysis in transperineal ultrasound images. The model achieves competitive segmentation accuracy compared to state-of-the-art methods while maintaining a substantially smaller parameter footprint. Robustness experiments further demonstrate that DAUNet retains high performance under partial visibility, highlighting its potential for deployment in edge-based ultrasound systems. The proposed approach supports the broader goal of enabling efficient, and accessible AI for maternal-fetal healthcare.
Glaucoma is one of the leading causes of irreversible blindness worldwide, underscoring the need for accurate and accessible diagnostic solutions. In this study, we propose a scalable glaucoma classification pipeline that leverages deep features extracted from sensor-acquired retinal fundus images using pretrained architectures-Swin Transformer V2, ConvNeXt V2, and EfficientNet V2. These models are employed as fixed feature extractors, and the concatenated representations are subsequently used to train and evaluate classical classifiers, including Random Forest, Multilayer Perceptron, and Support Vector Machine (SVM). Among these approaches, the SVM achieved the most balanced performance, which was further improved through hyperparameter optimization. Experiments conducted on the public RIM-ONE DL dataset demonstrated that the optimized SVM attained an accuracy of 0.957, an F1-score of 0.938, and an AUC-ROC of 0.978. These results highlight the effectiveness of integrating transformer-based and CNN-based deep representations with conventional machine learning models, offering a practical and resource-efficient framework for automated glaucoma classification.
Targeted drug delivery (TDD) has emerged as a promising strategy in cancer therapy, with nanoparticles playing a critical role in enhancing delivery precision and minimizing off-target toxicities. Leveraging protein ion channels for selective transport presents a novel approach to improve delivery specificity, particularly in cancer cells. In this study, we developed a molecular model to investigate the translocation of bare gold nanoparticles through the Transient Receptor Potential Vanilloid type 1 (TRPV1) ion channel embedded in both normal and cancer membranes. We simulated the interaction of nanoparticles with both the wild-type TRPV1 structure and a cancer-associated G563S mutant, embedded in lipid bilayers designed to reflect the distinct compositions of normal and cancer cell membrane models. Steered molecular dynamics (SMD) simulations were used to analyze the transport behavior across these systems. Our results demonstrate that gold nanoparticles interact more favorably with cancer-mimicking membranes, especially in the presence of the hyperactive TRPV1 mutant. These findings suggest that cancer-specific alterations in membrane compositions and channel conformations may facilitate selective nanoparticle transport, offering potential insights for developing TRPV1-targeted delivery strategies.
This study presents the validation of the Greek translation of the Godspeed Questionnaire Series (GQS) through an interactive workshop. Data were collected from 94 participants who assessed seven different wearable robotic devices in person and online. The validation examined psychometric characteristics including internal consistency reliability and factor structure preservation. The Greek translation successfully adapted semantic differentials while achieving equivalent performance compared to the original English version. Cross-language equivalence was maintained across most dimensions. All subscales demonstrated significant discriminative ability between robotic devices, validating the Greek GQS as a reliable instrument for human-robot interaction research and user assessment in wearable robotics applications.
The accurate and interpretable detection of neurodegenerative disorders remains a critical challenge in clinical neuroscience. NeuroXAI is introduced as a high-performance and explainable artificial intelligence (XAI)-driven deep learning framework for the electroencephalography (EEG)-based classification of Alzheimer's disease (AD) and Parkinson's disease (PD). The framework integrates a hybrid temporal convolutional network with disease-specific attention mechanisms to capture multi-granular spatial-temporal representations, achieving 98.79 % accuracy for PD and 90.64 % for AD on benchmark datasets, with robust generalization under leave-one-subject-out cross-validation. Clinical interpretability is supported through a comprehensive XAI pipeline that combines gradient-based saliency, integrated gradients, and perturbation-based occlusion analyses, which expose channel- and time-specific activation patterns consistent with known motor and cognitive biomarkers. Functional connectivity and nonlinear signal complexity measures, including fractal dimension and entropy, further validate disease-specific neural disruptions captured by the model. By uniting high predictive performance with clinically meaningful explanations, NeuroXAI offers a robust and deployment-ready solution for EEG-based neurological disorder screening, advancing the adoption of XAI in real-world diagnostics.
