Jan Kubicek

Deep learning has become a key enabling technology for detecting security-relevant events in visual surveillance data acquired from CCTV systems, UAV platforms, and other imaging sensors. However, despite substantial progress in benchmark performance, the operational deployment of such systems remains challenging due to dataset bias, domain shift, limited robustness, edge-computing constraints, and a lack of operationally meaningful evaluation metrics. This structured narrative review synthesises research published between 2021 and 2026, complemented by selected seminal studies. It is guided by predefined research questions and a documented, purposive search and selection strategy, and examines deep learning approaches for weapon detection, violence detection, anomaly recognition, person-related security tasks, perimeter monitoring, and UAV-based surveillance. The analysis covers major architectural paradigms, including YOLO-based detectors, RT-DETR, Vision Transformers, spatiotemporal networks, anomaly-detection frameworks, and edge-optimised models, with particular attention to dataset characteristics, model robustness, adversarial vulnerabilities, multimodal sensing, model compression, and regulatory considerations associated with the EU AI Act. As structuring contributions, the review proposes a multi-dimensional taxonomy that links security-event categories to architecture class, evaluation metric, sensor modality, and EU AI Act risk level, together with a critical analysis of architecture-specific failure modes under operational conditions. It further identifies recurring limitations that hinder real-world deployment: across representative studies, evaluation remains dominated by accuracy-oriented metrics such as mAP, F1-score, and FPS, whereas operational aspects including detection latency, false-alarm burden, and deployment robustness are insufficiently addressed. To bridge this gap, the review proposes Time-to-Detection (TTD) as a complementary operational evaluation framework, and highlights persistent challenges related to realistic datasets, demographic and domain biases, adversarial resilience, privacy-preserving learning, and federated deployment. The findings indicate that future research should prioritise standardised operational benchmarking, TTD-aware evaluation, robust and explainable models, realistic security datasets, efficient edge-AI deployment, and regulation-aware system design. Addressing these challenges will be essential for translating advances in deep learning into reliable and trustworthy security applications.

Over four decades of retinal vessel segmentation (RVS) research (1982–2025) are synthesized, tracing developments across classical image processing, traditional machine learning, and deep learning paradigms. A cross‐paradigm taxonomy is introduced that links computational methods with clinical interpretability and model generalization, addressing gaps identified in earlier surveys. A comparative meta‐analysis of 428 studies catalogs preprocessing and postprocessing techniques, segmentation architectures, and loss functions and assesses reported performance on benchmark datasets (DRIVE, STARE, CHASE_DB1, HRF) using 20 standardized evaluation metrics. Architectural trends, including residual‐enhanced U‐Nets, attention‐guided models, graph convolutional networks, and transformer‐based architectures (SGAT‐Net, MTPA‐UNet), correlate with measurable improvements in vessel continuity, topology preservation, and clinical reliability. Emerging interpretability strategies, including attention mapping, saliency visualization, and feature attribution, are examined for their ability to enhance clinical transparency and trust in automated RVS systems. Persistent challenges include cross‐dataset generalization, annotation scarcity, and limited clinical deployment of explainable models. A forward‐looking roadmap emphasizes domain adaptation, semi‐supervised and weakly‐supervised learning, and efficient architectures for clinical translation. By integrating quantitative evidence with a clinically grounded taxonomy, a novel perspective on RVS is provided, serving as a comprehensive reference for research and clinical adoption in precision ophthalmology.

Emotion recognition from biological signals provides valuable complementary information that may support clinical monitoring, symptom tracking, and assessment of affective processes, although it is not used as an independent diagnostic method. Despite its potential, dataset usability is hindered by challenges in accessibility, licensing, and inclusion of neutral stimuli. Existing datasets primarily focus on either AI training or eliciting emotional responses, both crucial for advancing the field. This review systematically analyzes current datasets and offers recommendations for their improvement. Using the PRISMA methodology, six databases were searched with predefined keywords, yielding 2,679 records. After filtering, 43 studies were included. Datasets were assessed based on four criteria: inclusion of neutral stimuli, downloadability, licensing for commercial use, and emotion assessment methods. Analyzed modalities include visual, auditory, physiological, and multimodal datasets. Methodological rigor determined their relevance to emotion recognition research. Findings reveal significant limitations in dataset usability, such as restrictions to academic use, inconsistent inclusion of neutral stimuli, and variability in assessment methods. Multimodal datasets show promise but require refinement for real-time applications in clinical and technological contexts. This review highlights the urgent need for open-access datasets and standardized methodologies to drive innovation and practical applications in emotion recognition technologies.

Pancreatic tumor segmentation from computed tomography (CT) images remains a challenging task due to limited annotated datasets, pronounced anatomical variability, and the high computational demands of state-of-the-art deep learning models, which collectively hinder their routine clinical adoption. This study proposes a clinically oriented end-to-end framework that bridges methodological advances in deep learning with practical deployment by enabling adaptive segmentation under realistic data growth scenarios. Rather than introducing a novel segmentation architecture, the framework integrates existing convolutional and transformer-based models within a lightweight graphical user interface (GUI) and employs a recursive augmentation strategy as a simulation mechanism to emulate the incremental availability of annotated clinical data over time. Multiple candidate architectures were first evaluated using cross-validation, after which representative lightweight and high-capacity models were selected for recursive augmentation. The framework was subsequently evaluated using both CNN-based architectures, such as 3D U-Net, and transformer-based models, such as VT-UNet-B, on multiple large-scale public datasets. Across all experiments, the proposed recursive augmentation consistently improved segmentation performance relative to baseline training, yielding relative Dice Similarity Coefficient (DSC) gains in the range of approximately 4–11% before reaching architecture-dependent saturation. Lightweight CNNs exhibited earlier saturation with smaller but consistent improvements, whereas transformer-based models benefited more substantially from incremental data expansion. By embedding segmentation models into an interactive GUI that supports real-time visualization and expert-driven refinement, the proposed framework emphasizes deployment feasibility, adaptability, and continuous performance improvement. The results outline a practical pre-clinical pathway toward resource-aware pancreatic tumor segmentation in real-world healthcare environments.

Recent advances in non-invasive blood pressure (BP) monitoring are transforming the field, offering alternatives to traditional cuff-based methods that are often limited by discomfort and intermittent measurement. Following PRISMA guidelines, this review systematically analyses studies published between 2020 and 2024, categorizing current BP measurement techniques into cuff-based, cuffless, and contactless methods. Special attention is given to emerging technologies such as photoplethysmography (PPG), electrocardiography (ECG), bioimpedance, radar sensing, and remote PPG (rPPG), as well as the integration of artificial intelligence (AI) for signal processing and BP estimation. The review evaluates the comparative performance, clinical relevance, and readiness for real-world application across modalities. While cuffless and contactless approaches show significant promise for continuous and user-friendly monitoring, their accuracy often depends on calibration, multimodal data fusion, and robust algorithm design. Ethical considerations related to unobtrusive monitoring are also discussed. This work highlights current limitations, identifies future research opportunities, and supports the development of next-generation, AI-driven BP monitoring systems suitable for both clinical and consumer use.

Accurate measurement of spleen volume is essential for the diagnosis of splenomegaly. While Computed Tomography (CT) is among the most reliable imaging modalities for this task, manual segmentation of the spleen is labor-intensive and impractical for routine clinical workflows. Automatic segmentation methods provide a more viable alternative for clinical deployment. In recent years, 3D Convolutional Neural Network (CNN) models have been widely used for this purpose due to their high segmentation accuracy. However, their computational and memory demands make them less suitable for real-time applications on edge devices with limited processing capabilities. To address these limitations, we introduce SVNC-Net (Spleen Volume and Neighborhood Convolutional Network) for efficient 3D spleen segmentation from CT scans. Rather than developing an entirely new architecture from scratch, SVNC-Net builds upon the U-Net framework with targeted architectural optimizations for efficiency. In SVNC-Net, each CT slice is processed independently using 2D convolutions. In its architecture, depthwise separable convolution is used to significantly reduce computational complexity and memory usage. To evaluate its performance and efficiency, a comparative analysis was conducted against well-known CNN-based models, including UPerNet, EMANet, CCNet, SegNet, and ShuffleNet. This evaluation was performed on two publicly available datasets used together for the first time in the literature. The promising results achieved from the comparative analysis verified that SVNC-Net is highly suitable for real-time applications and resource-constrained environments. Additionally, we explore post-training compression techniques such as pruning and quantization, which further enhance the model’s compactness and inference speed. These findings contribute to the ongoing efforts to develop efficient 2D deep learning models for 3D organ segmentation, particularly in resource-constrained clinical scenarios.