Search Results (3,262)

Real-time automatic detection of fire and smoke is a critical component of modern safety systems in surveillance, industrial, and environmental-protection applications, where conventional sensor-based systems show fundamental limitations. In this paper, a comparative analysis of four deep-learning architectures—DETR, Faster R-CNN (ResNet-50-FPN), Faster R-CNN (MobileNetV2), and RetinaNet—was conducted on a heterogeneous corpus of 67,765 annotated images originating from four different datasets. All models shared the same data-preparation, augmentation, and evaluation pipeline, while each was trained with the default configuration of its reference implementation; the comparison therefore reflects each architecture as it is typically deployed rather than a comparison under a single unified training budget. Faster R-CNN with the ResNet-50-FPN backbone achieved the highest accuracy (mAP@0.50 = 78.7% on the Indoor set; 73.8% on the SmokeAndFire set) and the highest mean mAP@0.50 across all datasets (48.7%), obtained with an inference speed of 76.9 FPS and a latency of 13.5 ms on NVIDIA RTX 4090 hardware, which makes it a promising candidate for real-time fire-detection systems on comparable hardware. The main contribution of this work is a unified evaluation of four representative object-detection architectures across four heterogeneous fire-and-smoke datasets. The study further quantifies the performance asymmetry between fire and smoke detection through a normalized morphological asymmetry index, reflecting the consistently lower detection accuracy achieved for smoke owing to its diffuse and semi-transparent appearance, and provides practical guidelines for selecting an architecture according to real-time deployment requirements. Because each configuration was trained once with a fixed random seed and all speed measurements were obtained on a single desktop GPU, the reported differences are interpreted descriptively; their statistical validation, cross-dataset evaluation, and measurement on embedded hardware are identified as future work. Full article

Background: Internet of Things (IoT) deployments face complex network dependencies and persistent data limitations that constrain machine learning (ML) intrusion detection systems (IDS). Objective: To synthesise peer-reviewed machine learning IDS research for IoT, focusing on dependency-driven failure propagation and chronic data scarcity, positioned against 2024–2025 EU regulatory requirements (NIS2, the Cyber Resilience Act). Eligibility criteria: Peer-reviewed empirical studies proposing or evaluating a machine learning or deep learning IoT intrusion detection method published in English from January 2018 (limited pre-2018 exceptions for seminal works). Information sources: IEEE Xplore, SpringerLink, Elsevier ScienceDirect, Scopus, Web of Science, and Google Scholar, searched on 12 February 2025. Risk of bias: Each candidate was scored against four criteria (objectives clarity, methodological soundness, reproducibility, IoT-security relevance); studies scoring at least 3 out of 4 were retained. Screening and scoring were performed by one reviewer, with a second reviewer independently checking 20 percent of records. Synthesis methods: Narrative thematic synthesis; heterogeneous metrics and incompatible datasets across studies precluded quantitative meta-analysis. Included studies: Of 427 records identified, 52 studies initially met inclusion criteria; a post hoc independently validated reconstruction of individual QA1–QA4 scores subsequently found that six did not meet the threshold or topical eligibility criteria, yielding a final 46-study corpus. Main findings: Generative adversarial networks (GANs) dominate dataset augmentation work, graph-based communication analysis addresses dependency modelling, and methods based on transformers or federated learning emerge from 2023 onward. Certainty of evidence: No formal grading (GRADE) applies to this narrative synthesis. Confidence in the corpus composition is high, following independent QA1–QA4 validation, while confidence in the thematic findings is moderate given single-reviewer screening and judgment-based classification. Conclusions: We identify three recurring gaps: real-time detection under resource constraints, dependency-aware detection, and regulatory compliance. Closing these gaps requires detection methods that treat IoT security as a networked and regulated system rather than an isolated device classification problem. Registration: Open Science Framework, 10.17605/OSF.IO/NMAK4 (registered retrospectively). No external funding supported this review. Full article

Objective: This systematic review evaluated the clinical application of autologous tissue grafts for chin augmentation performed with or without genioplasty. The primary outcomes included clinical and aesthetic improvement, graft stability and integration, resorption, complications, patient satisfaction, and the need for secondary procedures. Methods: The review was prospectively registered in OSF and conducted in accordance with the PRISMA 2020 statement. PubMed, Scopus, Embase, Web of Science, and WorldCat were searched using terms related to genioplasty, chin advancement, and autologous grafting materials, including bone, adipose tissue, cartilage, dermal tissue, and tooth-derived grafts. Eligible studies included original clinical publications involving human patients and reporting outcomes following chin augmentation with an autologous tissue graft, with or without genioplasty. Study selection and data extraction were conducted independently according to predefined eligibility criteria. Methodological quality was assessed using the appropriate Joanna Briggs Institute critical appraisal tools. Because of substantial clinical and methodological heterogeneity across the included studies, no meta-analysis was performed, and the findings were instead synthesized qualitatively. Results: Seventeen publications were included, comprising predominantly retrospective studies, case series, case reports, and technique-oriented clinical reports, with only one prospective randomized comparative trial; the overall level of evidence was therefore low, and comparative data across graft types remained limited. The evaluated materials comprised autologous adipose tissue, dermal grafts, iliac crest bone, costal cartilage and costochondral grafts, coronoid process bone, external oblique line corticocancellous bone, mandibular bone harvested during orthognathic surgery, a third-molar tooth graft, and an osteocartilaginous nasal hump graft. Most studies reported improvements in chin projection, facial profile, symmetry, or lower facial proportions. The available evidence suggests that autologous bone and cartilage grafts may provide integration and structural support, with limited clinically evident resorption reported; however, these observations derive from limited and heterogeneous evidence. Soft-tissue grafts improved chin contour but showed less predictable volume maintenance. Dermal graft resorption reached approximately 35% after 12 months, while fat grafting was associated with soft-tissue relapse and occasional secondary lipofilling. Serious graft-related complications were not frequently reported; however, adverse-event reporting was inconsistent, preventing reliable estimation of their incidence. Reported events included infections, temporary sensory disturbances, contour irregularities, and isolated graft removals. The certainty of the findings was limited by heterogeneous study designs, predominantly small or uncontrolled samples, variable outcome measures, and inconsistent follow-up. Conclusions: Autologous tissue grafts represent potentially effective options for chin augmentation when the graft source and surgical technique are selected according to the type and extent of the deformity. The available evidence suggests that bone and cartilage grafts may provide structural support, whereas adipose and dermal tissues may be considered for moderate soft-tissue augmentation; however, these conclusions are based on limited and heterogeneous evidence. However, the available evidence does not establish the superiority of autologous grafts over sliding genioplasty or alloplastic implants. Registration: Open Science Framework. Full article

Background/Objectives: Lateral flow assays (LFAs) are widely used point-of-care diagnostic devices due to their low cost, portability, and ease of use. However, most LFAs provide only qualitative results, limiting their utility for applications requiring continuous biomarker monitoring. This study introduces DeepBand, a deep learning-enabled multi-stage framework designed to automate the continuous quantification of analyte concentrations from unstandardized smartphone-captured lateral flow assay (LFA) images. Methods: A publicly available dataset containing 672 COVID-19 LFA images corresponding to four analyte concentrations (0.0, 1.8, 3.7, and 7.4 ng) was analyzed. A multi-stage pipeline was developed consisting of: (1) YOLOv11-based object detection to isolate the LFA cartridge from background artifacts, (2) a custom computer vision algorithm to identify and crop the test and control bands, and (3) a custom convolutional neural network (CNN) trained as a supervised regression model to predict continuous analyte concentrations. Data augmentation, hyperparameter optimization, and 5-fold cross-validation were used to improve model robustness. Results: The YOLOv11 model achieved approximately 99% mAP50 and 93.96% mAP95 for cartridge detection. Initial CNN models exhibited systematic underprediction of higher concentrations due to target imbalance; replacing mean squared error with Huber loss substantially improved performance, resulting in a final 20% held-out test set RMSE of 0.0292 ng. Analysis of HSV image channels demonstrated that the saturation-channel test-to-control intensity ratio was strongly correlated with analyte concentration (r = 0.94), consistent with the Beer–Lambert law governing LFA signal formation. Channel ablation studies confirmed the saturation channel as the most informative feature, while saliency mapping showed that the model primarily focused on biologically relevant test and control line regions. Conclusions: The proposed deep learning-enabled workflow, DeepBand, successfully integrates object detection, image processing, and CNN-based regression to provide automated quantitative interpretation of LFA results from smartphone images. Furthermore, the observed agreement between model behavior and Beer–Lambert theory suggests that the network learns biologically meaningful signal characteristics, supporting its potential for quantitative point-of-care diagnostics and longitudinal disease monitoring. Full article

Also known as spherical or omnidirectional video, 360-degree video has become increasingly prevalent in autonomous vehicles (AVs), virtual reality (VR), augmented reality (AR), and immersive media applications. However, its compression poses unique challenges due to the projection from a spherical surface onto a 2D plane. Common projections like Equirectangular Projection (ERP) introduce significant geometric distortions, causing straight-line motions on the sphere to appear as curved trajectories in the 2D domain. Standard motion estimation in codecs like H.266 (Versatile Video Coding, VVC) relies on translational (linear) motion vectors, leading to poor prediction accuracy, large residual errors, and inflated bitrates for 360-degree video content. This paper proposes the implementation of Spherical Motion Vector (SMV) mapping in the pre-encoder stage. By performing motion vector calculation directly on the spherical coordinate system (θ, φ) before mapping to the 2D pixel grid (x, y), SMV enables accurate tracking of object motion across projection boundaries and warped regions. This approach minimizes residual data and improves overall compression efficiency. This paper details the mathematical foundations, integration with H.266, implementation considerations, and simulated performance gains. The proposed method builds on prior work in rotational and geodesic motion models while introducing pre-encoder spherical preprocessing for broader compatibility. Full article

Electric vehicle (EV) adoption is accelerating rapidly, increasing pressure on public charging infrastructure, urban energy systems, and network-expansion planning. However, many existing studies examine charging demand forecasting, station network characteristics, or infrastructure siting as separate problems, limiting their ability to translate observed charging behaviour into coordinated planning decisions. To address this gap, this study proposes a multi-domain machine learning framework that integrates spatio-temporal charging demand, station-level infrastructure characteristics, land-use information, and graph-based network relationships for EV charging prediction and infrastructure siting. The analysis uses 8,544,695 public charging transactions recorded at 8553 stations across Beijing during January and July 2025. Charging sessions are aggregated into an hourly station panel and modelled using seasonal-naïve and historical mean baselines, ridge regression, gradient-boosted trees, graph-augmented gradient boosting, and a gated recurrent unit network. Model reliability is assessed through temporal validation, feature ablation, spatial holdout testing, cross-season transfer analysis, explainability, and non-parametric statistical comparison. The results show that graph-augmented gradient boosting achieved the strongest RMSE and variance-explained performance, with an RMSE of 53.80 kWh and $R^2=0.734$, while the gated recurrent unit produced the lowest MAE of 23.82 kWh. Graph neighbour features yielded only a marginal and statistically non-significant forecasting improvement, indicating that the station network is structurally informative but predictively redundant once temporal history is available. For infrastructure expansion, NSGA-II achieved the highest Pareto front hypervolume and identified a knee-point solution of 156 additional chargers, reducing unmet demand by 43.4%. These findings demonstrate that integrated forecasting, graph analysis, and multi-objective siting can provide more defensible and operationally relevant evidence for city-scale EV charging planning. Full article

Background: Large language models (LLMs) can produce clinically plausible recommendations that are not adequately supported by authoritative evidence. Objectives: We introduce GuidelineGuard, a modular multi-agent retrieval-augmented generation pipeline in which a separate Auditor verifies claim–sentence support before a recommendation is surfaced. Methods: The original evaluation used a 73-sentence guideline corpus and GG-Bench-60, with replication across three open-weight backbones. In response to reviewer concerns about benchmark size and selective evaluation, we added a source-traceable GG-Bench-200 stress test and the complete 500-case held-out PQA-L test split of PubMedQA. The revision experiments compare single-pass RAG, a paired multi-agent no-Auditor ablation, and GuidelineGuard; the paired runner is designed to share the Planner–Retriever–Clinician draft so that the Auditor is the only intervention. Checkpoint verification confirmed an identical observable pre-audit state for all 200 GG-Bench cases and 496/500 PubMedQA cases; four PubMedQA cases were regenerated after quota-interrupted resumption and were correct commitments in both arms. Because the originally used hosted Llama endpoints became unavailable after the initial experiments, the expanded runs use openai/gpt-oss-20b for generation and openai/gpt-oss-120b for the Auditor. Results: On GG-Bench-200, single-pass RAG achieved 0.970 operational accuracy, while the no-Auditor and GuidelineGuard arms achieved 0.955 and 0.925, respectively. GuidelineGuard committed on 186/200 cases (coverage 0.930) and was correct on 185/186 commitments (selective accuracy 0.995); all 186 commitments cited at least one gold evidence identifier. Relative to the paired no-Auditor arm, the gate rejected six otherwise-correct commitments and no incorrect commitment. On PubMedQA-500, single-pass RAG achieved 0.644 operational accuracy at 0.950 coverage, the no-Auditor arm 0.638 at 0.896 coverage, and GuidelineGuard 0.550 at 0.736 coverage. Selective accuracy increased across those operating points from 0.678 to 0.712 to 0.747. Within the 496 PubMedQA cases with verified-identical observable pre-audit state, the gate rejected 36 incorrect and 45 correct pre-audit commitments, demonstrating both error enrichment and a substantial false-rejection cost. Conclusions: The expanded results support GuidelineGuard as a selective claim–evidence verification mechanism, not as a universally more accurate generator. Its value is the explicit, auditable coverage–risk trade-off; the appropriate verification threshold is task- and cost-dependent and requires prospective clinical validation. Full article

Human hands play a central role in manipulation, communication, and physical interaction, making their accurate digital reconstruction a long-standing challenge in computer vision and graphics. Reliable modeling of hand pose, hand shape, and interaction is essential for applications ranging from immersive virtual and augmented reality to robotics, activity understanding, and human–machine interfaces. Over the past decade, research has evolved from isolated single-hand pose estimation toward increasingly holistic frameworks that jointly reconstruct two hands and the objects they manipulate. This review provides a comprehensive overview of hand reconstruction methods, tracing the progression from classical model-based approaches and early learning-driven pipelines to modern systems capable of two-hand interaction modeling and category-agnostic hand–object reconstruction. We structure the surveyed literature according to fundamental algorithmic paradigms, encompassing model-based formulations, convolutional and graph-based learning methods, transformer-based architectures, and emerging neural implicit and Gaussian representations. This review is tailored for researchers new to the field and follows a chronological and conceptual organization that elucidates how key design choices have shaped current capabilities and limitations. We analyze persistent challenges that hinder the widespread adoption of hand reconstruction methods in practical applications, including articulation complexity, generalization to unseen objects, and computational efficiency. Finally, this paper provides a forward-looking perspective on emerging research directions, highlighting trends toward real-time, category-agnostic, and semantically meaningful hand reconstruction systems for future human-centered computing. Full article

Background/Objectives: Evidence comparing cement- and screw-retained implant-supported restorations remains heterogeneous, particularly in augmented posterior maxillae. This study evaluated the association between definitive retention mode and peri-implant marginal bone level at a minimum 5-year follow-up after maxillary sinus floor elevation. Methods: This retrospective clustered cohort included 248 patients and 528 bone-level implants; radiographic measurements were available for 522 implants from 247 patients. Marginal bone level was assessed by CBCT immediately after implant placement and at least 60 months later. The primary adjusted analysis used generalized estimating equations with patient-level clustering and robust standard errors. Results: Mean marginal bone level at final follow-up was 0.133 mm for cement-retained and 0.105 mm for screw-retained restorations; the unadjusted difference was not statistically significant (Mann–Whitney U = 31,366; p = 0.064). In the primary GEE model, which excluded provisional retention because of its strong collinearity with definitive retention, definitive screw retention was not associated with marginal bone level at follow-up (B = −0.0028 mm; 95% CI: −0.0165–0.0108; p = 0.684). The time × definitive retention interaction was also not significant (B = −0.0069 mm; 95% CI: −0.0187–0.0049; p = 0.249). Conclusions: Definitive retention mode was not significantly associated with marginal bone level in the primary adjusted analysis, and longitudinal marginal bone change did not differ significantly between retention strategies. Sensitivity analyses indicated that estimates were influenced by the strong collinearity between provisional and definitive retention. These observational findings do not demonstrate superiority or equivalence of either retention strategy. Full article

Objective: This study aimed to evaluate the long-term degradation behavior, biostimulatory effects, and biocompatibility of a novel poly-L-lactic acid-block-polyethylene glycol/hyaluronic acid (PLLA–b–PEG/HA) composite filler for soft tissue augmentation. Methods: PLLA–b–PEG/HA microsphere properties were characterized via scanning electron microscopy (SEM), X-ray diffraction (XRD), Fourier-transform infrared spectroscopy (FTIR), nuclear magnetic resonance hydrogen spectroscopy (¹H NMR), thermogravimetry (TG) and differential scanning calorimetry (DSC). A 104-week in vivo rabbit model was established to systematically observe filler degradation and tissue responses. Ultrasound monitoring, histological staining, ELISA and RT-PCR were performed to assess volumetric changes, inflammatory reactions and collagen synthesis-related signaling. Results: Physicochemical property tests demonstrated that PLLA–b–PEG retains the fundamental physicochemical properties of pristine PLLA while exhibiting enhanced hydrophilicity. B-ultrasound demonstrated a presented uniform in vivo distribution without displacement or diffusion over time, confirming steady and predictable degradation. SEM verified progressive morphological degradation and porous evolution of the microspheres. The filler induced a mild, balanced inflammatory microenvironment with early expression of both pro-inflammatory (IL-12, TNF-α) and anti-inflammatory (IL-4) cytokines, which resolved gradually over time. Sustained TGF-β upregulation persisted throughout the 104-week observation period, driving continuous neocollagenesis and prominent neoelastogenesis, thereby achieving favorable and long-term tissue remodeling with excellent biocompatibility. Conclusions: The PLLA–b–PEG/HA composite filler exhibits controllable degradation properties and homeostatic regulatory effects, along with outstanding long-term biosafety and tissue integration capacity. As an ideal biostimulatory filler for soft tissue augmentation, it can effectively facilitate the regeneration of high-quality functional extracellular matrix rich in collagen fibers and elastic fibers, and holds promising clinical prospects for natural and long-lasting soft tissue filling applications. Full article

Background/Objectives: Diabetic retinopathy is a major cause of preventable vision loss worldwide, making early and accurate disease grading crucial for timely treatment. Although convolutional neural network (CNN)- and transformer-based architectures have demonstrated promising performance for retinal image analysis, comprehensive comparisons within a unified experimental framework remain limited. This study systematically compares representative standard and lightweight CNN- and transformer-based architectures for multi-class DR grading. Methods: Six ImageNet-pretrained deep-learning models, including ResNet50, EfficientNet-B0, MobileNetV2, Vision Transformer (ViT), Swin-Tiny, and Swin Transformer, were evaluated on the APTOS 2019 retinal fundus image dataset under a unified experimental configuration with consistent preprocessing, data augmentation, training, and evaluation settings. All models were fine-tuned and evaluated independently over five runs with different random seeds. Their performance was assessed using accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), Quadratic Weighted Kappa (QWK), per-class analysis, computational efficiency, and statistical analysis. Results: Transformer-based models generally achieved higher mean classification performance than the evaluated CNN-based models. Swin-Tiny achieved the highest mean accuracy (82.3%), macro F1-score (64.4%), weighted F1-score (82.1%), and QWK (89.8%) across the five runs. Among the CNN-based models, EfficientNet-B0 achieved the strongest overall classification performance, whereas MobileNetV2 provided the lowest computational complexity. The results also highlighted differences in learning behavior and computational requirements across the evaluated architectures. Repeated experiments demonstrated stable performance across different random seeds, supporting the reliability of the proposed evaluation. Conclusions: Overall, this study provides a comprehensive comparison of representative CNN- and transformer-based architectures under consistent experimental settings and offers practical guidance for selecting suitable deep learning models for automated diabetic retinopathy screening. Full article

Background/Objectives: Accurate delineation of the optic disc and optic cup in retinal fundus photographs is a prerequisite for automated glaucoma screening. While encoder–decoder segmentation models have advanced considerably, the contribution of upstream preprocessing to segmentation accuracy, and the stability of that contribution across repeated training runs, remain insufficiently characterized. Methods: Five preprocessing pipelines, baseline, Contrast Limited Adaptive Histogram Equalization (CLAHE), Region of Interest (ROI) cropping, ROI+CLAHE, and CLAHE with heavy augmentation, were benchmarked under a fixed EfficientUNet++ model with an EfficientNet-B7 encoder on three publicly available fundus datasets (REFUGE, ORIGA, and Drishti-GS). Every configuration was retrained under three independent random seeds (42, 15, and 89) to assess run-to-run variability. Seed-level standard deviations accompany every reported mean and define the confidence limit on each ranking. Results: On REFUGE, CLAHE with augmentation (Config 5) achieved the strongest mean Dice (disc $0.9523\pm 0.0017$; cup $0.8348\pm 0.0018$). On ORIGA, all five configurations clustered within 0.0067 disc Dice; ROI+CLAHE (Config 4) was marginally ahead on disc ($0.9681\pm 0.0002$) and augmentation led on the cup ($0.8873\pm 0.0024$). On Drishti-GS, all five configurations converged successfully once optimizer and loss settings were corrected; the near-total failures seen in earlier single-run experiments reflected a configuration problem, not the small (81-image) training set. Conclusions: CLAHE applied to full-resolution images is the single most consistently beneficial preprocessing choice across all three datasets. ROI+CLAHE showed a small, initialization-stable advantage on ORIGA, but ROI crop centres were derived from ground-truth centroids, an oracle localization setting, and these results should not be interpreted as achievable by a fully automated pipeline. Data augmentation showed a consistent reduction in initialization sensitivity on small datasets and may be beneficial as a default strategy. Full article

Background/Objectives: Dental implants are a reliable treatment for tooth loss, but identifying the implant brand when patient records are unavailable remains a clinical challenge that complicates prosthetic repair and complication management. This study aimed to develop and evaluate a deep learning-based system for automated identification of dental implant brands from panoramic and periapical radiographs. Methods: In this retrospective study, anonymized radiographs containing implants of twelve known brands were obtained from the archives of Necmettin Erbakan University Faculty of Dentistry. A two-stage pipeline was employed: a YOLOv11 detector first localized and cropped the implant regions, after which an EfficientNetV2-M convolutional neural network, fine-tuned via transfer learning, classified the implant brand. Class imbalance was addressed through offline and online data augmentation. Results: On the held-out test set of 531 implant crops spanning twelve brands, the classifier achieved an overall accuracy of 96.23% (95% CI 94.5–97.7%), a macro-averaged F1-score of 0.953, and a macro-averaged ROC-AUC of 0.991; the complete pipeline evaluated end to end on detector-predicted crops reached 96.0% match-conditional implant-level accuracy, corresponding to a precision-aware end-to-end identification F1-score of 88.2% (precision 81.6%, recall 96.0%) when all predicted boxes, including false detections, were counted. Grad-CAM analysis, including misclassified and low-confidence cases, indicated that predictions were based on clinically meaningful implant morphology. Conclusions: These findings indicate that the proposed two-stage approach provides accurate and interpretable implant brand identification, supporting its potential as a clinical decision support tool. Full article

Background/Objectives: Repurposing existing drugs may represent a promising strategy for effective cancer therapy. This study was the first to investigate the augmented antitumor therapeutic effect achieved by co-incorporating the NSAID Niflumic acid (NIF) with ceramides into PEGylated cerosomes (NIF-loaded PEG-CERs) in a novel platform that targets specifically the MAPK-ERK signaling pathway and miR-21-5p modulation. Methods: The prepared formulae were statistically optimized utilizing a full factorial design and the optimal formula (C5) was further incorporated into a topical gel and evaluated for ex vivo rat skin permeation, and tested in vivo in a subcutaneous solid Ehrlich carcinoma (SEC) mice model. Results: The optimal formula (C5) showed tubular elongated morphology with higher EE% (96.71 ± 0.0), lower vesicular size (VS) and PDI values, 292.95 ± 0.78 nm and 0.47 ± 0.0 respectively. A high ZP value (−37.5 ± 0.57 mV) was in accordance with stability results showing good stability of the optimal formula (C5). Permeability studies exhibited 2.02-fold higher skin permeation compared to pure NIF gel. A significant decrease in tumor volume and marked improvement in survival rate in SEC mice were confirmed by downregulation of EGFR, ERK1, ERK2, and miR-21-5p expression. Furthermore, an increase in total antioxidant capacity and caspase-3 levels was observed, accompanied by significant suppression in cyclin D1, MMP-2, COX-2, and MDA levels. Finally, histopathological analysis revealed the superior antitumor effect of C5 gel together with immunohistochemical assay showing the lowest BCL-2-positive staining, indicating the restoration of physiological apoptotic balance. Conclusions: Based on the previous findings, NIF-loaded PEG-CERs offer augmented therapeutic potential for efficient topical skin cancer management in an SEC mice model. Full article

Reliable fault diagnosis of autonomous underwater vehicles (AUVs) remains challenging because fault-sensitive patterns can be obscured by measurement noise, while several fault categories may exhibit similar responses and ambiguous decision boundaries. To address these issues, this study proposes an attention-enhanced deep residual shrinkage network with adaptive boundary-aware hard feature augmentation (AE-DRSN-ABHFA). First, channel and temporal attention mechanisms are integrated into a deep residual shrinkage network (DRSN) to jointly suppress noise-related responses and emphasize fault-sensitive information across sensor and temporal dimensions. Second, the proposed adaptive boundary-aware hard feature augmentation (ABHFA) mechanism identifies ambiguous boundary samples using class prototypes and boundary-aware scores, while class-level difficulty feedback is used to determine the class-processing order during augmentation. Directional pseudo-features are then introduced during training to enrich the feature distribution near difficult decision boundaries. A joint objective incorporating pseudo-feature classification and batch-hard triplet constraints is further employed to improve intra-class compactness and inter-class separability. The main comparative, noise-robustness, and core ablation experiments were repeated using five independent random seeds. Under the current experimental setting, AE-DRSN-ABHFA achieves a mean clean-condition Accuracy of 95.27 ± 1.17% and a Macro-F1 of 0.9498 ± 0.0128. Across the four evaluated noisy conditions, the proposed model obtains an average Accuracy of 85.37 ± 2.76%, an average Macro-F1 of 0.8253 ± 0.0317, and an average worst-class Accuracy of 75.39 ± 5.98%, demonstrating favorable overall and class-level diagnostic performance under the evaluated noisy operating conditions. These results demonstrate the potential of the proposed framework for noise-robust multi-class AUV fault diagnosis, although further validation on different AUV platforms and real sea-trial data remains necessary. Full article