Deep Learning Methods Research Articles

Prediction of pore water pressure generation of liquefiable clean sands under cyclic shear loading through deep learning

ABSTRACT In this study, a novel data-driven approach is carried out to predict the pore pressure generation of liquefiable clean sands during cyclic loading. An extensive and comprehensive database of actual stress-controlled cyclic simple shear test results in terms of pore pressure time histories is gathered from a large number of experiments. While the classical machine learning (ML) algorithms help predict the number of liquefaction cycles in a few models, the desired level of accuracy in predicting the actual trend and robustness in pore pressure build-up is only achieved in deep learning (DL) methods. Results indicate that the Long-Short Term Memory (LSTM) working model, employed with Stacked LSTM and the Windowing data processing method, is necessary for making fairly good cyclic pore pressure build-up predictions. This study proposes a model that can ultimately be utilised to predict the pore pressure response of in-situ liquefiable sandy soil layers without resorting to plasticity-based complex theoretical models, which has been the current practice. The robustness achieved in the model reassures the reliability of the study, raising confidence in developing data-driven constitutive models for soils that have the potential to replace conventional plasticity-based theories.

Hyperspectral Imaging and Deep Learning for Quality and Safety Inspection of Fruits and Vegetables: A Review.

Quality inspection of fruits and vegetables linked to food safety monitoring and quality control. Traditional chemical analysis and physical measurement techniques are reliable, they are also time-consuming, costly, and susceptible to environmental and sample changes. Hyperspectral imaging technology combined with deep learning methods can effectively overcome these problems. Compared with human evaluation, automated inspection improves inspection efficiency, reduces subjective error, and promotes the intelligent and precise fruit and vegetable quality inspection. This paper reviews reports on the application of hyperspectral imaging technology combined to deep learning methods in various aspects of fruits and vegetables quality assessment. In addition, the latest applications of these technologies in the fields of fruit and vegetable safety, internal quality, and external quality inspection are reviewed, and the challenges and future development directions of hyperspectral imaging technology combined with deep learning in this field are prospected. Hyperspectral imaging combined with deep learning has shown significant advantages in fruit and vegetable quality inspection, especially in improving inspection accuracy and efficiency. Future research should focus on reducing costs, optimizing equipment, personalizing feature extraction, and model generalizability. In addition, the development of lightweight models and the balance of accuracy, the enhancement of the database and the importance of quantitative research should also be brought to attention. These efforts will promote the wide application of hyperspectral imaging technology in fruit and vegetable inspection, improve its practicability in the actual production environment, and bring important progress for food safety and quality management.

Achieving precision assessment of functional clinical scores for upper extremity using IMU-Based wearable devices and deep learning methods

Stroke is a serious cerebrovascular disease, and rehabilitation following the acute phase is particularly crucial. Not all rehabilitation outcomes are favorable, highlighting the necessity for personalized rehabilitation. Precision assessment is essential for tailored rehabilitation interventions. Wearable inertial measurement units (IMUs) and deep learning approaches have been effectively employed for motor function prediction. This study aims to use machine learning techniques and data collected from IMUs to assess the Fugl-Meyer upper extremity subscale for post-stroke patients with motor dysfunction. IMUs signals from 120 patients were collected during a clinical trial. These signals were fed into a gated recurrent unit network to complete the scoring of individual actions, which were then aggregated to obtain the total score. Simultaneously, on the basis of the internal correlation between the Fugl–Meyer assessment and the Brunnstrom scale, Brunnstrom stage prediction models of the arm and hand were established via the random forest and extremely randomized trees algorithm. The experimental results show that the proposed models can score Fugl-Meyer items with a high accuracy of 92.66%. The R2 between the doctors’ score and the model’s score is 0.9838. The Brunnstrom stage prediction models can predict high-quality stages, achieving a Spearman correlation coefficient of 0.9709. The application of the proposed method enables precision assessment of patients’ upper extremity motor function, thereby facilitating more personalized rehabilitation programs to achieve optimal recovery outcomes.Trial registration: Clinical trial of telerehabilitation training and intelligent evaluation system, ChiCTR2200061310, Registered 20 June 2022-Retrospective registration.

Combining Deep Data-driven and Physics-inspired Learning for Shear Wave Speed Estimation in Ultrasound Elastography.

Shear wave elastography (SWE) provides quantitative markers for tissue characterization by measuring shear wave speed (SWS), which reflects tissue stiffness. SWE uses an acoustic radiation force pulse sequence to generate shear waves that propagate laterally through tissue with transient displacements. These waves travel perpendicular to the applied force, and their displacements are tracked using high-frame-rate ultrasound. Estimating the SWS map involves two main steps: speckle tracking and SWS estimation. Speckle tracking calculates particle velocity by measuring RF/IQ data displacement between adjacent firings, while SWS estimation methods typically compare particle velocity profiles of samples that are laterally a few millimeters apart. Deep learning (DL) methods have gained attention for SWS estimation, often relying on supervised training using simulated data. However, these methods may struggle with real-world data, which can differ significantly from simulated training data, potentially leading to artifacts in the estimated SWS map. To address this challenge, we propose a physics-inspired learning approach that utilizes real data without known SWS values. Our method employs an adaptive unsupervised loss function, allowing the network to train with real noisy data to minimize the artifacts and improve the robustness. We validate our approach using experimental phantom data and in vivo liver data from two human subjects, demonstrating enhanced accuracy and reliability in SWS estimation compared to conventional and supervised methods. This hybrid approach leverages the strengths of both data-driven and physics-inspired learning, offering a promising solution for more accurate and robust SWS mapping in clinical applications.

Deep learning-based hippocampus asymmetry assessment for Alzheimer's disease diagnosis.

The symmetry of the brain hippocampus may be disrupted by natural aging and neurodegenerative diseases. Currently, clinical studies on hippocampus asymmetry are limited to subjective visual evaluation and rough volume measurements, lacking quantitative standards. This paper proposes a quantitative assessment method of the hippocampus asymmetry based on deep learning, named DeepHAA (Deep Learning-based Hippocampus Asymmetry Assessment). The DeepHAA model extracts feature representations of left and right hippocampus structures in MRI images and achieved feature fusion through a cross-attention mechanism. A quantitative assessment method is proposed based on the distance between the multimodal embedding of the input sample and the reference embeddingspace. The experimental dataset of this paper included MRI scans of 199 subjects, including 53 normal cognition (NC), 71 mild cognitive impairment (MCI) and 33 Alzheimer's disease (AD). The experimental results show that DeepHAA model can effectively identify and distinguish the NC, MCI, andAD. The proposed deep learning method integrates asymmetric information about hippocampus structure into the diagnosis of AD and has potential clinical applicationvalue.

DeepGuard: real-time threat recognition using Golden Jackal optimization with deep learning model

Violence recognition in surveillance videos is a vital feature of current security systems. It can improve complete security measures, making it essential in generating safe public places and defending against unforeseen security tasks. Real risk and violence detection in surveillance videos signify a cutting-edge use of deep learning (DL) technologies. By using innovative neural networks, this plan proposes to improve safety by quickly classifying possible threats and occurrences of violence within surveillance footage. The system influences DL models to examine complex visual patterns, differentiating between normal actions and potential safety risks. This real-time detection ability allows instant responses and involvements, donating to a proactive technique in safeguarding public safety and the security of critical infrastructures. Integrating DL methods in surveillance video study showcases the probability for sophisticated, intelligent methods to enlarge classical safety measures, delivering fast and effective threat recognition in different environments. This study presents a DeepGuard model for Real-Time Threat Recognition using Golden Jackal Optimization with Deep Learning. The purpose of the DeepGuard model is to identify violence and non-violence events in the surveillance videos. In the DeepGuard model, an improved ShuffleNetv2 model can be applied to derivate the intrinsic and complex features from the input images. Besides, the Golden Jackal Optimization (GJO) model can be applied for the optimal hyperparameter tuning of the improved ShuffleNetv2 model. The DeepGuard model uses long short-term memory neural networks (LSTM-NNs) for violence detection. The performance evaluation of the DeepGuard model takes place using benchmark datasets. The experimental outcomes of the DeepGuard model obtained optimum performance with 99.00% and 98.63% when related to other techniques in terms of distinct measures.

Lightweight multi-scale convolution with blended feature attention for facial expression recognition in the wild

Abstract Facial expression recognition (FER) has achieved excellent performance in recent years under the controlled scenarios through deep learning methods. However, the accurate recognition of facial expression in the wild conditions with occlusion, pose changes, and uneven lighting still a challenging problem, not to mention the problem of limited computing resources faced by the growing size of proposed network models. To solve these problems, this paper proposes a multi-scale network based on lightweight convolution (MLC-Net), aiming to improve the recognition accuracy of FER in real-world environments while significantly reducing the number of parameters. In MLC-Net, image shallow features are extracted for global and local blocks through pre-extracted blocks. The global feature extraction block uses a mixed washing network as the basis of the Multi-scale Module, reducing the its parameters and computational complexity when extracting different levels of semantic information. Meanwhile, the improved Efficient Lightweight Channel-spatial Attention Module (ELAM) is used to enhance the feature fusion ability of the Multi-scale Module. The local feature extraction block utilizes convolutional groups and Lightweight Spatial Attention Module (LSAM) to extract and enhance local features, guiding the network to pay attention to regions with significant features, and proposes a Local Relationship Transformer (LRT), through which a multi-head attention mechanism is used to establish connections between regions, thus further enhancing the ability to recognize complex expressions. The effectiveness of the proposed MLC-Net is validated on multiple in the wild FER datasets, and the results show that MLC-Net can achieve a good balance between recognition accuracy and network lightweighting, providing a promised solution for practical application of FER.

ProtNote: a multimodal method for protein-function annotation.

Understanding the protein sequence-function relationship is essential for advancing protein biology and engineering. However, less than 1% of known protein sequences have human-verified functions. While deep learning methods have demonstrated promise for protein function prediction, current models are limited to predicting only those functions on which they were trained. Here, we introduce ProtNote, a multimodal deep learning model that leverages free-form text to enable both supervised and zero-shot protein function prediction. ProtNote not only maintains near state-of-the-art performance for annotations in its training set, but also generalizes to unseen and novel functions in zero-shot test settings. ProtNote demonstrates superior performance in prediction of novel GO annotations and EC numbers compared to baseline models by capturing nuanced sequence-function relationships that unlock a range of biological use cases inaccessible to prior models. We envision that ProtNote will enhance protein function discovery by enabling scientists to use free text inputs without restriction to predefined labels - a necessary capability for navigating the dynamic landscape of protein biology. The code is available on GitHub: https://github.com/microsoft/protnote; model weights, datasets, and evaluation metrics are provided via Zenodo: https://zenodo.org/records/13897920. Supplementary Information is available at Bioinformatics online.

Accurate bladder cancer diagnosis using ensemble deep leaning

There are an estimated 1.3 million cases of cancer globally each year, making it one of the most serious types of urinary tract cancer. The methods used today for diagnosing and monitoring bladder cancer are intrusive, costly, and time-consuming. In clinical practice, invasive biopsy followed by histological examination continues to be the gold standard for diagnosis. Bladder cancer biomarkers have been used in expensive diagnostic tests created recently, however their reliability is limited by their high rates of false positives and false negatives. The potential and use of artificial intelligence in urological diseases have been the subject of several research, as interest in artificial intelligence in medicine has grown recently. In this paper, a new bladder cancer model called Ensemble Deep Learning (EDL) will be provided to accurately diagnose patients. Outlier rejection is used to filter data using the interquartile range (IQR) then the image diagnosis. The proposed EDL consists of three deep learning algorithms, which are; Convolutional Neural Network (CNN), Generative Adversarial Network (GAN), and a new deep learning method called Explainable Deep Learning (XDL) that depends on Guided Gradient Weighted Class Activation Map (Guided Grad-CAM). In fact, Guided Grad-CAM enables doctor to understand the diagnose. A new voting mechanism will be used to integrate the results of all three methods to produce the final result to accurately diagnose bladder cancer cases. In fact, the used voting method depends on using majority voting based on two different scenarios according to the results of CNN, GAN, and XDL. If these three methods give the same class category, then the final diagnosis will be this class category. On the other hand, if the three methods give different class category, then the final result will be followed by the accuracy of each class. The proposed EDL model was tested after several trials. The results have proved that EDL model is more efficient and more accurate to diagnose bladder cancer disease. It introduced the highest accuracy results and the lowest error results as well as execution time.

Automated pulmonary nodule classification from low-dose CT images using ERBNet: an ensemble learning approach.

The aim of this study was to develop a deep learning method for analyzing CT images with varying doses and qualities, aiming to categorize lung lesions into nodules and non-nodules. This study utilized the lung nodule analysis 2016 challenge dataset. Different low-dose CT (LDCT) images, including 10%, 20%, 40%, and 60% levels, were generated from the full-dose CT (FDCT) images. Five different 3D convolutional networks were developed to classify lung nodules from LDCT and reference FDCT images. The models were evaluated using 400 nodule and 400 non-nodule samples. An ensemble model was also developed to achieve a generalizable model across different dose levels. The model achieved an accuracy of 97.0% for nodule classification on FDCT images. However, the model exhibited relatively poor performance (60% accuracy) on LDCT images, indicating that dedicated models should be developed for each low-dose level. Dedicated models for handling LDCT led to dramatic increases in the accuracy of nodule classification. The dedicated low-dose models achieved a nodule classification accuracy of 90.0%, 91.1%, 92.7%, and 93.8% for 10%, 20%, 40%, and 60% of FDCT images, respectively. The accuracy of the deep learning models decreased gradually by almost 7% as LDCT images proceeded from 100 to 10%. However, the ensemble model led to an accuracy of 95.0% when tested on a combination of various dose levels. We presented an ensemble 3D CNN classifier for lesion classification, utilizing both LDCT and FDCT images. This model is able to analyze a combination of CT images with different dose levels and image qualities.

STGAT: Spatial–Temporal Graph Attention Neural Network for Stock Prediction

Stock price prediction and portfolio optimization are critical research areas in financial markets, as they directly impact investment strategies and risk management. Traditional statistical methods and machine learning approaches have been widely applied to these tasks, but they often fail to fully capture the complex dynamics of financial markets. Traditional statistical methods typically rely on unrealistic assumptions or oversimplified models, neglecting the nonlinear and high-dimensional characteristics of market data. Additionally, deep learning methods, especially temporal convolution networks and graph attention networks, have been introduced in this area and have achieved significant improvements in both stock price prediction and portfolio optimization. Therefore, this study proposes a Spatial–Temporal Graph Attention Network (STGAT) that integrates STL decomposition components and graph structures to model both temporal patterns and asset correlations. By combining graph attention mechanisms with temporal convolutional modules, STGAT effectively processes spatiotemporal data, enhancing the accuracy of stock price predictions. Empirical experiments on the CSI 500 and S&P 500 datasets demonstrate that STGAT outperforms other deep learning models in both prediction accuracy and portfolio performance. The investment portfolios constructed based on STGAT’s predictions achieve higher returns in real market scenarios, which validates the feasibility of spatiotemporal feature fusion for stock price prediction and highlights the advantages of graph attention networks in capturing complex market characteristics. This study not only provides a robust tool for portfolio optimization but also offers valuable insights for future research in intelligent financial systems.

Smart navigation through a rotating barrier: Deep reinforcement learning with application to size-based separation of active microagents.

We employ deep reinforcement learning methods to investigate shortest-time navigation strategies for smart active Brownian particles (microagents), which self-propel through a rotating potential barrier in a static, viscous, fluid background. The microagent's motion begins at a specified origin and terminates at a designated destination. The potential barrier is modeled as a localized, repulsive Gaussian potential with finite support, whose peak location rotates at a given angular velocity about a fixed center within the plane of motion. We use the advantage actor-critic approach to train microagents for their origin-to-destination navigation through the barrier. By employing this approach, we demonstrate that the rotating potential (as opposed to a static one) enables size-based sorting and separation of the microagents. In other words, microagents of different radii arrive at the destination at sufficiently well-separated average times, facilitating their sorting. The efficiency of particle sorting is quantified by introducing specific separation measures. We also demonstrate how training the microagents in a noisy background, as opposed to a noise-free one, can improve the precision of their size-based sorting. Our findings suggest promising avenues for future research on smart active particles equipped with deep reinforcement learning to navigate complex environments, particularly in microscale applications.

Enhancing Sensor-Based Human Physical Activity Recognition Using Deep Neural Networks

Human activity recognition (HAR) is the task of classifying sequences of data into defined movements. Taking advantage of deep learning (DL) methods, this research investigates and optimizes neural network architectures to effectively classify physical activities from smartphone accelerometer data. Unlike traditional machine learning (ML) methods employing manually crafted features, our approach employs automated feature learning with three deep learning architectures: Convolutional Neural Networks (CNN), CNN-based autoencoders, and Long Short-Term Memory Recurrent Neural Networks (LSTM RNN). The contribution of this work is primarily in optimizing LSTM RNN to leverage the most out of temporal relationships between sensor data, significantly improving classification accuracy. Experimental outcomes for the WISDM dataset show that the proposed LSTM RNN model achieves 96.1% accuracy, outperforming CNN-based approaches and current ML-based methods. Compared to current works, our optimized frameworks achieve up to 6.4% higher classification performance, which means that they are more appropriate for real-time HAR.

AI-Driven Early Detection of Cognitive Decline

Early detection of cognitive decline is critical for initiating timely interventions that can delay or mitigate the progression of neurodegenerative disorders such as Alzheimer’s disease. Traditional diagnostic methods, which rely on episodic clinical evaluations and subjective assessments, often miss the subtle, early signs of impairment. Recent advancements in Artificial Intelligence (AI) offer transformative potential by enabling continuous, objective, and highly sensitive analysis of behavioral, linguistic, physiological, and neurological data. This paper explores the role of AI in revolutionizing the early diagnosis of cognitive decline through the integration of machine learning models, natural language processing, and multimodal data analysis. We examine the key data sources—including speech patterns, gait analysis, neuroimaging, and digital biomarkers—that power AI-driven systems and highlight how these tools surpass conventional approaches in accuracy and scalability. The discussion extends to various AI models such as deep learning and ensemble methods, which can detect subtle patterns indicative of mild cognitive impairment before it becomes clinically apparent. Benefits such as personalized monitoring, remote accessibility, and population-wide screening are considered alongside critical challenges, including data privacy, algorithmic bias, and clinical integration. Ethical considerations and future research directions are emphasized, focusing on the need for transparency, inclusivity, and cross-disciplinary collaboration. By showcasing real-world applications and current limitations, this study provides a comprehensive overview of how AI can support earlier diagnoses and improved care for individuals at risk of cognitive decline. The findings underscore AI's promise as an essential tool in the future of cognitive health.

Age of Information Minimization in Vehicular Edge Computing Networks: A Mask-Assisted Hybrid PPO-Based Method

With the widespread deployment of various emerging intelligent applications, information timeliness is crucial for intelligent decision-making in vehicular networks, where vehicular edge computing (VEC) has become an important paradigm to enhance computing capabilities by offloading tasks to edge nodes. To promote the information timeliness in VEC, an optimization problem is formulated to minimize the age of information (AoI) by jointly optimizing task offloading and subcarrier allocation. Due to the time-varying channel and the coupling of the continuous and discrete optimization variables, the problem exhibits non-convexity, which is difficult to solve using traditional mathematical optimization methods. To efficiently tackle this challenge, we employ a hybrid proximal policy optimization (HPPO)-based deep reinforcement learning (DRL) method by designing the mixed action space involving both continuous and discrete variables. Moreover, an action masking mechanism is designed to filter out invalid actions in the action space caused by limitations in the effective communication distance between vehicles. As a result, a mask-assisted HPPO (MHPPO) method is proposed by integrating the action masking mechanism into the HPPO. Simulation results show that the proposed MHPPO method achieves an approximately 28.9% reduction in AoI compared with the HPPO method and about a 23% reduction compared with the mask-assisted deep deterministic policy gradient (MDDPG).

A hybrid learning network with progressive resizing and PCA for diagnosis of cervical cancer on WSI slides

Current artificial intelligence (AI) trends are revolutionizing medical image processing, greatly improving cervical cancer diagnosis. Machine learning (ML) algorithms can discover patterns and anomalies in medical images, whereas deep learning (DL) methods, specifically convolutional neural networks (CNNs), are extremely accurate at identifying malignant lesions. Deep models that have been pre-trained and tailored through transfer learning and fine-tuning become faster and more effective, even when data is scarce. This paper implements a state-of-the-art Hybrid Learning Network that combines the Progressive Resizing approach and Principal Component Analysis (PCA) for enhanced cervical cancer diagnostics of whole slide images (WSI) slides. ResNet-152 and VGG-16, two fine-tuned DL models, are employed together with transfer learning to train on augmented and progressively resized training data with dimensions of 224 × 224, 512 × 512, and 1024 × 1024 pixels for enhanced feature extraction. Principal component analysis (PCA) is subsequently employed to process the combined features extracted from two DL models and reduce the dimensional space of the feature set. Furthermore, two ML methods, Support Vector Machine (SVM) and Random Forest (RF) models, are trained on this reduced feature set, and their predictions are integrated using a majority voting approach for evaluating the final classification results, thereby enhancing overall accuracy and reliability. The accuracy of the suggested framework on SIPaKMeD data is 99.29% for two-class classification and 98.47% for five-class classification. Furthermore, it achieves 100% accuracy for four-class categorization on the LBC dataset.

Improved deep learning method and high-resolution reanalysis model-based intelligent marine navigation

Large-scale weather forecasting is critical for ensuring maritime safety and optimizing transoceanic voyages. However, sparse meteorological data, incomplete forecasts, and unreliable communication hinder accurate, high-resolution wind system predictions. This study addresses these challenges to enhance dynamic voyage planning and intelligent ship navigation. We propose IPCA-MHA-DSRU-Net, a novel deep learning model integrating incremental principal component analysis (IPCA) with a spatial-temporal depthwise separable U-Net. Key components include: (1) IPCA preprocessing to reduce dimensionality and noise in 2D wind field data; (2) depthwise-separable convolution (DSC) blocks to minimize parameters and computational costs; (3) multi-head attention (MHA) and residual mechanisms to improve spatial-temporal feature extraction and prediction accuracy. The framework is optimized for real-time onboard deployment under communication constraints. The model achieves high accuracy in high-resolution wind predictions, validated through reanalysis datasets. Experiments demonstrated enhanced path planning efficiency and robustness in dynamic oceanic conditions. The IPCA-MHA-DSRU-Net balances computational efficiency and accuracy, making it viable for resource-limited ships. This novel IPCA application provides a promising alternative for preprocessing large-scale meteorological data.

Advancing Spanish Speech Emotion Recognition: A Comprehensive Benchmark of Pre-Trained Models

Feature extraction for speech emotion recognition (SER) has evolved from handcrafted techniques through deep learning methods to embeddings derived from pre-trained models (PTMs). This study presents the first comparative analysis focused on using PTMs for Spanish SER, evaluating six models—Whisper, Wav2Vec 2.0, WavLM, HuBERT, TRILLsson, and CLAP—across six emotional speech databases: EmoMatchSpanishDB, MESD, MEACorpus, EmoWisconsin, INTER1SP, and EmoFilm. We propose a robust framework combining layer-wise feature extraction with Leave-One-Speaker-Out validation to ensure interpretable model comparisons. Our method significantly outperforms existing state-of-the-art benchmarks, notably achieving scores on metrics such as F1 on EmoMatchSpanishDB (88.32%), INTER1SP (99.83%), and MEACorpus (92.53%). Layer-wise analyses reveal optimal emotional representation extraction at early layers in 24-layer models and middle layers in larger architectures. Additionally, TRILLsson exhibits remarkable generalization in speaker-independent evaluations, highlighting the necessity of strategic model selection, fine-tuning, and language-specific adaptations to maximize SER performance for Spanish.

Review of sEMG for Exoskeleton Robots: Motion Intention Recognition Techniques and Applications

The global aging trend is becoming increasingly severe, and the demand for life assistance and medical rehabilitation for frail and disabled elderly people is growing. As the best solution for assisting limb movement, guiding limb rehabilitation, and enhancing limb strength, exoskeleton robots are becoming the focus of attention from all walks of life. This paper reviews the progress of research on upper limb exoskeleton robots, sEMG technology, and intention recognition technology. It analyzes the literature using keyword clustering analysis and comprehensively discusses the application of sEMG technology, deep learning methods, and machine learning methods in the process of human movement intention recognition by exoskeleton robots. It is proposed that the focus of current research is to find algorithms with strong adaptability and high classification accuracy. Finally, traditional machine learning and deep learning algorithms are discussed, and future research directions are proposed, such as using a deep learning algorithm based on multi-information fusion to fuse EEG signals, electromyographic signals, and basic reference signals. A model with stronger generalization ability is obtained after training, thereby improving the accuracy of human movement intention recognition based on sEMG technology, which provides important support for the realization of human–machine fusion-embodied intelligence of exoskeleton robots.

Video-based multi-target multi-camera tracking for postoperative phase recognition.

Deep learning methods are commonly used to generate context understanding to support surgeons and medical professionals. By expanding the current focus beyond the operating room (OR) to postoperative workflows, new forms of assistance are possible. In this article, we propose a novel multi-target multi-camera tracking (MTMCT) architecture for postoperative phase recognition, location tracking, and automatic timestamp generation. Three RGB cameras were used to create a multi-camera data set containing 19 reenacted postoperative patient flows. Patients and beds were annotated and used to train the custom MTMCT architecture. It includes bed and patient tracking for each camera and a postoperative patient state module to provide the postoperative phase, current location of the patient, and automatically generated timestamps. The architecture demonstrates robust performance for single- and multi-patient scenarios by embedding medical domain-specific knowledge. In multi-patient scenarios, the state machine representing the postoperative phases has a traversal accuracy of , of timestamps are generated correctly, and the patient tracking IDF1 reaches . Comparative experiments show the effectiveness of using AFLink for matching partial trajectories in postoperative settings. As our approach shows promising results, it lays the foundation for real-time surgeon support, enhancing clinical documentation and ultimately improving patient care.