Image Processing & Applications

Musculoskeletal disorders pose a significant occupational health challenge, impairing worker well-being and operational efficiency, particularly in manufacturing. The risks of improper postures require effective identification and mitigation strategies. Current assessment methods rely on subjective, qualitative measures and lack robust numerical metrics, especially in hand pose estimation. This study introduces an advanced machine learning algorithm designed for the identification and classification of hand poses using inertial gloves. By tracking hand positions, the proposed approach facilitates objective risk assessments during work activities. Experimental findings validate the system's efficacy, indicating its potential to improve ergonomic analysis and enhance workplace safety.

Keywords: hand posture recognition , Industry 5.0 , inertial gloves , machine learning algorithm

Background and objective: Despite the availability of several commercial solutions for predicting the soft tissue outcomes of maxillofacial surgeries, none have proven sufficiently reliable for routine clinical use. This study proposes a 3D methodology for predicting soft tissue displacement following maxillofacial surgery without relying on mechanical modeling, unlike most existing approaches. Methods: Pre- and post-operative Cone Beam Computed Tomography scans of patients with class III malocclusion were collected. Tailored image processing and volume reconstruction techniques were applied to semi-automatically generate 3D soft tissue models. Cephalometric landmarks were identified to perform a geometrical similarity analysis among patients with the same malocclusion class undergoing the same surgical procedure. Vectorial displacement maps were generated to capture the soft tissue changes from pre- to post-operative and were then applied to the pre-operative of test patients to predict soft tissue outcomes. Euclidean distances were calculated between predicted and real post-operative positions, and the Wilcoxon signed-rank test was conducted to assess statistical differences between predicted and real landmark coordinates. Results: Error maps indicated that approximately 70 % of predicted facial points had errors below 2.5 mm, while around 10 % ranged between 2.5 mm and 3 mm. Statistically significant differences (p < 0.05) were observed only for the gonion and cheilion. Conclusion:. The findings support the validity of the geometrical similarity analysis and the vectorial displacement map approach. The simplicity and promising accuracy of the proposed method encourage further investigations across different surgical procedures. Additionally, integrating this methodology into surgical planning could offer a viable alternative to commercial solutions.

Keywords: 3D soft tissue prediction , Dental malocclusion , Displacement vectorial map , Orthognathic surgery , Soft tissue displacement

Work-related musculoskeletal disorders (WMSDs) are a leading concern in industrial ergonomics, often stemming from sustained non-neutral postures and repetitive tasks. This paper presents a vision-based framework for real-time, frame-level ergonomic risk classification using a lightweight Vision Transformer (ViT). The proposed system operates directly on raw RGB images without requiring skeleton reconstruction, joint angle estimation, or image segmentation. A single ViT model simultaneously classifies eight anatomical regions, enabling efficient multi-label posture assessment. Training is supervised using a multimodal dataset acquired from synchronized RGB video and full-body inertial motion capture, with ergonomic risk labels derived from RULA scores computed on joint kinematics. The system is validated on realistic, simulated industrial tasks that include common challenges such as occlusion and posture variability. Experimental results show that the ViT model achieves state-of-the-art performance, with F1-scores exceeding 0.99 and AUC values above 0.996 across all regions. Compared to previous CNN-based system, the proposed model improves classification accuracy and generalizability while reducing complexity and enabling real-time inference on edge devices. These findings demonstrate the model's potential for unobtrusive, scalable ergonomic risk monitoring in real-world manufacturing environments.

Keywords: attention mechanism , computer vision , deep learning , ergonomic risks , human-centered manufacturing , posture recognition , work-related musculoskeletal disorders

Respiratory rehabilitation is essential for patients after cardiac surgery to avoid postoperative pulmonary complications. However, the majority of rehabilitation programs are self-administered at home, which often results in low compliance. This research work aims to develop a smartphone application to support patients during respiratory rehabilitation sessions with a three-ball incentive spirometer. The developed application leverages the smartphone camera and a keypoints detection algorithm to track spirometer exercises, guiding patients through the session. The app also stores quantitative data, making them available to medical staff on a telemedicine web platform. The application relies on the YOLOv8-pose model to detect seven spirometer keypoints, allowing for accurate tracking of ball positions and corresponding inspiration rates. We trained and optimized three YOLOv8-pose models (Nano, Small, and Medium) using a custom dataset and four image input sizes. Each model was evaluated on a test set of real-world scenarios. YOLOv8-pose Nano, with an input size of 192 pixels, was selected for app deployment, achieving a recall of 98.49%, precision of 98.51%, mAP50 of 98.38%, mAP50-95 of 85.26%, and average inference time on a consumer smartphone of 279±73ms. These results demonstrate the potential of using a lightweight YOLOv8-pose model for real-time monitoring and compliance improvement of respiratory rehabilitation.

Keywords: Incentive Spirometer , Keypoint Detection , Respiratory Rehabilitation , Telemonitoring , YOLO

Anterior cruciate ligament reconstruction (ACL-R) has a high success rate, but a subset of patients report unsatisfactory outcomes, highlighting the need for improved predictive tools. Recent studies have shown that bony morphology plays an important role in both the risk of ACL injury and post-reconstruction outcomes. This study aims to develop a framework for understanding the relationship between morphology and patient satisfaction. We explore the use of machine learning (ML) models to predict patient satisfaction following ACL reconstruction based on preoperative bone morphology. Using 3D models from high-resolution Magnetic Resonance (MR) imaging, 21 morphological parameters of the femur and tibia were analysed, including linear, surface and volumetric measurements. These were correlated with patient satisfaction as assessed by the Knee injury and Osteoarthritis Outcome Score Global (KOOS Global) questionnaire, administered six months after surgery. To streamline the analysis, automated systems based on statistical shape models and deep learning were employed for landmark placement and segmentation. A dataset of 45 patients was used to train and evaluate three ML models: Random Forest, Logistic Regression and Gradient Boosting. Dimensionality reduction identified 10 key morphological features and Gradient Boosting outperformed the other models, achieving accuracy of 0.89, precision of 0.90 and recall of 0.89. The results demonstrate the potential of combining advanced imaging, automated measurement and ML for personalized prediction of patient satisfaction. The predictive model shows that primarily femoral features, along with intercondylar notch volume and tibial measurements, play a critical role in determining patient satisfaction. This approach may guide surgical planning, patient counselling, intervention strategies and rehabilitation. Although promising, further research with larger datasets, additional parameters and advanced ML techniques is needed to improve model robustness and clinical applicability.

Keywords: 3D modeling , Anterior cruciate ligament , Machine learning , Morphology , Prediction

Lip-reading technology, based on visual speech decoding and automatic speech recognition, offers a promising solution to overcoming communication barriers, particularly for individuals with temporary or permanent speech impairments. However, most Visual Speech Recognition (VSR) research has primarily focused on the English language and general-purpose applications, limiting its practical applicability in medical and rehabilitative settings. This study introduces the first Deep Learning (DL) based lip-reading system for the Italian language designed to assist individuals with vocal cord pathologies in daily interactions, facilitating communication for patients recovering from vocal cord surgeries, whether temporarily or permanently impaired. To ensure relevance and effectiveness in real-world scenarios, a carefully curated vocabulary of twenty-five Italian words was selected, encompassing critical semantic fields such as Needs, Questions, Answers, Emergencies, Greetings, Requests, and Body Parts. These words were chosen to address both essential daily communication and urgent medical assistance requests. Our approach combines a spatiotemporal Convolutional Neural Network (CNN) with a bidirectional Long Short-Term Memory (BiLSTM) recurrent network, and a Connectionist Temporal Classification (CTC) loss function to recognize individual words, without requiring predefined words boundaries. The experimental results demonstrate the system's robust performance in recognizing target words, reaching an average accuracy of 96.4% in individual word recognition, suggesting that the system is particularly well-suited for offering support in constrained clinical and caregiving environments, where quick and reliable communication is critical. In conclusion, the study highlights the importance of developing language-specific, application-driven VSR solutions, particularly for non-English languages with limited linguistic resources. By bridging the gap between deep learning-based lip-reading and real-world clinical needs, this research advances assistive communication technologies, paving the way for more inclusive and medically relevant applications of VSR in rehabilitation and healthcare.

Keywords: 3D convolutional neural network , automatic speech recognition , deep learning , Lip-reading , visual speech decoding

Ensuring worker safety is a basic concern in high-risk areas like construction sites, where accurate and efficient human identification is essential. Accurate identification of personnel at worksites is crucial for reducing security hazards. Contemporary approaches exhibit low identification rates, localization inaccuracies, and elevated latency, hence jeopardizing safety. This paper presents SECURITY YARDS, a framework utilizing cloud computing and the Internet of Things (IoT) for human identification through gait recognition. Our approach utilizes Speeded Up Robust Features (SURF) and Convolutional Neural Networks (CNNs) to extract and classify gait characteristics, providing a more nuanced and adaptable representation. A Support Vector Machine (SVM) at the Fog level optimizes classification, improving precision and processing efficacy. Final decision-making, data storage, and monitoring are conducted on the cloud, ensuring scalability and real-time safety management. The findings indicate that SECURITY YARDS markedly enhances accuracy and lowers latency relative to traditional biometric identification systems.

Keywords: Cloud computing , Deep Learning , Internet of Things , Machine Learning

Background: Adherence to home-based respiratory rehabilitation is a significant challenge for cardiac surgery patients. Few studies have attempted to address this problem. Methods: This research introduces a digital solution to improve adherence to respiratory rehabilitation by offering virtual supervision. It leverages the YOLOv11 pose detection algorithm to track the Voldyne® 2500 incentive spirometer and monitor rehabilitation sessions. Usability and accuracy were evaluated with nine healthy volunteers, combining objective performance metrics with the System Usability Scale (SUS) survey. Results: The models demonstrated high accuracy in spirometer tracking and efficient inference times. The app achieved an average SUS score of 83%, indicating good usability, though refinements to the monitoring algorithms are recommended. Conclusion: This low-cost, non-invasive solution shows potential for clinical use. Future efforts will focus on enhancing usability and accuracy to prepare the app for clinical trials.

Keywords: Cardiac Rehabilitation , Respiratory Therapies , Spirometry , YOLO Network

Extended reality (XR) technologies are no longer peripheral innovations but emerging cornerstones of human–technology interaction across critical sectors. This article takes the position that engineering and healthcare represent the most mature and strategically relevant domains for XR adoption, given their safety-critical nature, intensive training requirements, and strong alignment with the human-centric visions of Industry 5.0 and Healthcare 5.0. We synthesize evidence from product design, manufacturing, training, and patient care to demonstrate how XR is reshaping workflows, skills, and therapeutic practices. Beyond surveying applications, we argue that the future of XR depends on its integration with artificial intelligence, digital twins, and multisensory feedback, converging into systems capable of perceiving, reasoning, and adapting to complex physical and human environments. We contend that widespread adoption will remain limited without open standards, validated protocols, and robust evaluation frameworks addressing safety, interoperability, and data governance. By framing XR as both a technological enabler and a societal imperative, this position article calls for coordinated action among researchers, practitioners, and policymakers to realize XR's role in building sustainable, personalized, and participatory innovation ecosystems.

Keywords: digital industry , extended reality , human-centered innovation , human–computer interfaces/interactions , immersive technology , virtual and augmented reality environments , virtual prototyping , XR for training , XR in healthcare , XR in manufacturing

The rapid advancements in computer processing, algorithmic development, and the availability of large-scale datasets have positioned Artificial Intelligence (AI) as a valuable tool across multiple domains, including Medicine. In the field of Nuclear Medicine neuroimaging, with Positron Emission Tomography (PET), AI has demonstrated significant potential in improving diagnostic accuracy for neurodegenerative cognitive disorders. This is especially relevant for the early diagnosis, preclinical detection, and prediction of disease progression in Alzheimer's disease (AD), the most prevalent form of cognitive impairment in individuals over 65 years of age. This narrative review aims to synthesize current advances, explore future directions, and highlight outstanding challenges in the application of Artificial Intelligence to PET imaging for the clinical management of Alzheimer's disease, with particular focus on three key modalities: 18F-FDG PET, Amyloid PET, and Tau PET.

Keywords: 18F-FDG PET , Alzheimer's disease , Amyloid PET , Artificial Intelligence , Tau PET

Background: Differentiating Parkinson's disease (PD) from essential tremor (ET) is often challenging, especially in early or atypical cases. Dopamine transporter (DAT) single-photon emission computed tomography (SPECT) with 123I-Ioflupane supports diagnosis, and semi-quantitative tools such as DaTQUANT® and BasGanV2™ provide objective measures. This study compared their diagnostic performance when integrated with supervised machine learning. Methods: We retrospectively analysed 123I-Ioflupane SPECT scans from 169 patients (133 PD, 36 ET). Semi-quantitative analysis was performed using DaTQUANT® v2.0 and BasGanV2™ v.2. Classification tree (ClT), k-nearest neighbour (k-NN), and support vector machine (SVM) models were trained and validated with stratified shuffle split (250 iterations). Diagnostic accuracy was compared between the two software packages. Results: All classifiers reliably distinguished PD from ET. DaTQUANT® consistently achieved higher accuracy than BasGanV2™: 93.8%, 93.2%, and 94.5% for ClT, k-NN, and SVM, respectively, versus 90.9%, 91.7%, and 91.9% for BasGanV2™ (p < 0.001). Sensitivity and specificity were also consistently higher for DaTQUANT® than BasGanV2. Class imbalance (PD > ET) was addressed using Synthetic Minority Over-sampling Technique (SMOTE). Conclusions: Machine learning analysis of 123I-Ioflupane SPECT enhances differentiation between PD and ET. DaTQUANT® outperformed BasGanV2™, suggesting greater suitability for AI-driven decision support. These findings support the integration of semi-quantitative and AI-based approaches into clinical workflows and highlight the need for harmonised methodologies in movement disorder imaging.

Keywords: 123I-Ioflupane , artificial intelligence , BasGanV2™ , DaTQUANT® , dopamine transporters , machine learning , movement disorders , Parkinson's disease , rehabilitation , SPECT

Seabed mapping represents a very useful tool for seascape characterization and benthic habitat study and requires advanced technologies for acquiring, processing, and interpreting remote data. Particularly, acoustic instruments, such as high-resolution swath bathymetry sounders (i.e., multibeam echosounder: MBES), allow us to recognize, identify, and map the extension of benthic habitats without applying invasive mechanical procedures. Bathymetry and backscatter (BS) data are crucial to perform modern habitat mapping. Although the acquisition and processing of bathymetric data follow standardized procedure (e.g., Hydrographic Organization guidelines), and recent studies have proposed recommendations for backscatter acquisition and processing, a broadly validated methodological approach, integrating geomorphometric analysis for benthic habitat mapping, is still lacking. In this work, a new approach for benthic habitat mapping, with a focus on coralligenous bioconstructions, was developed using the open-source software QGIS. This methodology, tested within the Isola Capo Rizzuto Marine Protected Area (Calabria, Italy), is designed to be freely reproducible by researchers working in the field of marine ecosystem monitoring and conservation. Through the proposed mapping procedure, it is possible to (i) identify benthic habitats in selected study areas by combining bathymetry and BS data with geomorphological indices performed in QGIS and (ii) quantitatively define the 2D and 3D distribution of coralligenous bioconstructions in terms of surface covered, thickness, and volume. Moreover, the statistical analysis of quantitative morphometric data allowed for comparison of geometric characteristics of different coralligenous morphotypes. The obtained results, combined with improvement of minimally invasive sampling and geobiological–geochemical characterization, can contribute to the development of protocols aimed at monitoring marine bioconstructed ecosystems, many of which are protected by national and international regulations due to their importance for Mediterranean biodiversity preservation, and plan actions for their protection and persistence.

Efficient thermal management is essential for the reliability and performance of traction inverters. However, direct optimization via Computational Fluid Dynamics (CFD) is often impractical due to the high dimensionality of the design space and the high computational cost of each simulation. To overcome this limitation, a surrogate-based optimization framework is developed to enhance the thermal and hydraulic performance of an automotive traction inverter cooling system. The methodology integrates CFD, deep neural networks (DNNs), and a multi-objective evolutionary algorithm. A simplified representation of the ACEPACKTM DRIVE power module is employed to generate an extensive dataset through automated, GPU-accelerated CFD simulations, making data generation computationally feasible while avoiding the prohibitive cost of direct optimization. A DNN surrogate model is trained to accurately predict pressure drop and heated-wall temperature, achieving mean relative errors below 3% and 1%, respectively. This surrogate model then guides a Non-Dominated Sorting Genetic Algorithm III in the optimization of key geometric parameters, including pin-fin diameter, spacing, height, wall clearance, as well as of physical parameter such as the surface roughness of the pin-fins. CFD-based validation of the Pareto-optimal designs, performed on the full inverter geometry, indicates reductions of up to 25% in pressure drop and approximately 2% in junction temperature. These results suggest that the proposed methodology promises robustness and generalizability, showing good potential for further application in data-driven thermal design optimization.

Keywords: CFD , Deep neural network , NSGA-III , Optimization , Power module

The automotive development of Electric Vehicles (EVs) is accelerating due to environmental concerns and technological advances. The complexity and performance demand of EVs require the use of Silicon-Carbide (SiC) technologies for the related efficiency, high-temperature resistance, and robust dynamic switching. To ensure the high performance of electric vehicles, the rigorous screening of the delivered SiC devices plays a pivotal role. Defects in SiC production can cause failures that directly affect the performance of electric vehicle engines, particularly in the traction inverter subsystem. To identify these defects early, advanced visual screening through optical microscopy and X-ray approaches has been proposed by leading car makers. These screening methodologies require advanced technical expertise and are still prone to significant errors, even when performed manually by experienced operators. To address these inefficiencies, the authors propose an ensemble deep learning pipeline for performing a robust automated visual inspection of SiC devices. By combining Multi-Head Attention blocks with adaptive input data distortion compensation through Jacobian regularization within convolutional architectures, the proposed model leverages global context and local feature maps, enhancing accuracy and robustness in multimodal defects detection. The proposed combined approach, tested on ACEPACK/TPACK DRIVE SiC power modules provided by STMicroelectronics, achieved an average accuracy of approximately 93% in both methodologies.

Keywords: Deep Learning , Defects analysis , Silicon-Carbide

Accurate monitoring of the gate–oxide health of Silicon-Carbide (SiC) MOSFET devices is pivotal for condition-based maintenance of electrified vehicles, high-frequency converters and consumer electronic applications. A MOSFET parameter closely related to the device oxide interface is the threshold voltage. The SiC MOSFETs threshold voltage (Vth) drifts with interface oxide–trap accumulation and it is therefore a direct early-marker of the device degradation. The Vth drift monitoring is then significantly relevant in the mentioned application fields. This work proposes a non-intrusive, single-shot intelligent pipeline that infers Vth solely from an input data-snapshot of the steady-state MOSFET static behavior based on current-to-voltage I – V family curves sampled during routine gate driving. We introduce HyperDeep, a novel Hyperbolic Möbius Deep architecture that embeds each I – V input snapshot in the hyperbolic Poincaré ball, processes it through ad-hoc designed residual Möbius layers followed by hyperbolic self-attention blocks. An adaptive exponential Lipschitz regularization block compensates process-spread-induced statistical drift in SiC devices, preserving mapping stability across SiC MOSFET's wide threshold-voltage window. The proposed HyperDeep system was trained on large dataset that includes both synthetic and real data obtained through Dynamic Gate Stress (DGS) methodology applied to SiC MOSFET GEN3 devices delivered by STMicroelectronics. The proposed deep pipeline attains in average of 0.015 MSE (Mean Squared Error) in retrieving actual Vth, outperforming compared state-of-the-art architectures.

Keywords: Deep learning , gate driving , hyperbolic geometry , MOSFET , residual lifetime

Silicon Carbide (SiC) power devices are highly sought in modern power electronics for their superior performance; however, yield is severely impacted by crystal-growth defects often termed killer defects. Fast and accurate post-growth inspection of SiC wafers is therefore critical to identify such defects and improve manufacturing outcomes. We propose an improved YOLO-based one-stage detector that integrates a Non-Local Block + Gradient Reversal Layer (GRL) module into the YOLO backbone between the Spatial Pyramid Pooling (SPP) and BottleneckCSP layers. The Non-Local blocks enable the CNN to leverage global context in the microscopic images, improving its ability to recognize subtle defect patterns across the wafer. Meanwhile, the GRL-based domain adaptation allows the model to learn domain-invariant features, addressing distribution shifts between different datasets or imaging conditions. The proposed architecture, comprising these components, is evaluated on a dataset of SiC wafer optical micrographs containing various defect types to demonstrate its effectiveness (e.g., dislocations and stacking faults). The proposed method achieves mAP@0.5=95% and mAP@0.5:0.95=94.5%. These improvements demonstrate the efficacy of combining attention and adversarial domain adaptation in the YOLO architecture. The enhanced model detects SiC defects with higher accuracy and robustness to varying imaging conditions, marking a significant step forward in automated wafer defect inspection.

Keywords: Deep Learning , Defects analysis , Non-local block , Silicon-Carbide

Real-time rendering is increasingly used in augmented and virtual reality (AR/VR), interactive design, and product visualisation, where materials must prioritise efficiency and consistency rather than the extreme accuracy required in offline rendering. In parallel, the growing demand for personalised and customised products has created a need for digital materials that can be generated in-house without relying on expensive commercial systems. To address these requirements, this paper presents a low-cost digitisation workflow based on photometric stereo. The system integrates a custom-built scanner with cross-polarised illumination, automated multi-light image acquisition, a dual-stage colour calibration process, and a node-based reconstruction pipeline that produces albedo and normal maps. A reproducible evaluation methodology is also introduced, combining perceptual colour-difference analysis using the CIEDE2000 ((Formula presented.)) metric with angular-error assessment of normal maps on known-geometry samples. By openly providing the workflow, bill of materials, and implementation details, this work delivers a practical and replicable solution for reliable material capture in real-time rendering and product customisation scenarios.

Keywords: digitalisation workflow , material scanner , photometric stereo , real-time rendering , texture evaluation

This paper introduces a novel triangular encoding technique specifically designed for identifying geometric correspondences in structured light three-dimensional (3D) imaging. The proposed method, based on a hybrid approach between pseudo-dot arrays and M-arrays, overcomes limitations of existing encoding strategies by using a triangular pattern that directly encodes information in a 2D grid. The paper presents the theoretical framework for the encoding, a pseudo-algorithm for its application to 2D images, and a discussion of the algorithm's capabilities and limitations. An interesting potential of using this pattern is its ability to reduce the post-processing time required to generate 3D models, as it can inherently determine the connectivity of the vertices in the projected mesh. The proposed triangular encoding approach is shown to be particularly effective for dynamic 3D scene reconstruction and for applications requiring robust information encoding, such as QR codes. Additionally, the paper outlines the potential for future work in developing a decoding strategy and further refining the encoding system using multiple states, such as RGB colors.

Keywords: 3D imaging , combinatorial geometry , encoding techniques , pattern recognition , structured light

Classification is a crucial task for reintroducing end-of-life fabrics as raw materials in a circular process, thus reducing reliance on dyeing processes. In this context, this review explores the evolution of automated and semi-automated colour classification methods, emphasizing the transition from deterministic techniques to advanced methods, with a focus on machine learning, deep learning, and particularly Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs). These technologies show potential for improving accuracy and efficiency. The results highlight the need for enriched datasets, deeper AI integration into industrial processes, and alignment with circular economy objectives to enhance sustainability without compromising industrial performance. Tested against a case study, the different architectures confirmed the state-of-the-art statements demonstrating that they are effective in classification, with better performance reached by CNN-based methods, which outperforms other methods in most colour families, with an average accuracy of 86.1%, indicating its adaptability for this task. The adoption of the proposed AI-based colour-classification roadmap could be effective in reducing dyeing operations, lower costs, and improve sorting efficiency for textile SMEs.

Keywords: circular economy , CNNs , colour assessment , textiles