Search Results (371)

Node-level cargo-volume forecasting in logistics sorting networks requires modeling temporal dynamics together with directed inter-node dependencies and planned topology perturbations. This study addresses 1-h-ahead forecasting under a partial-information boundary, where historical node volumes, the pre-change network structure, and planned route-topology changes are available before prediction, whereas continuous post-change dynamic edge weights and realized post-change graph states are unavailable. We propose a perturbation-aware framework that represents the sorting system as a directed network and integrates temporal features, pre-change structural descriptors, topology-change encodings, perturbation-response proxies, and similarity-assisted support for data-scarce nodes within a unified forecasting protocol. A shared random forest backbone is used only to assess the incremental value of these representations. Experiments on 57 sorting centers show that temporal dynamics dominate under stable-network conditions. Under topology perturbation, topology-change signals reduce test weighted absolute percentage error (WAPE) from 18.10% to 17.11%, and perturbation-response proxies further reduce it to 16.91%. For data-scarce nodes, similarity support reduces test WAPE from 29.43% to 26.68%, with consistent gains under 10%, 20%, and 30% sample-retention settings. These results suggest that the framework provides an interpretable and information-admissible representation strategy for node-level forecasting in directed networked systems. Full article

In large-scale state-owned enterprise audit and compliance tasks, ensuring that similar violations receive consistent disciplinary decisions is essential for procedural fairness and institutional credibility. However, existing retrieval methods face three major challenges: lexical matching methods fail to recognize semantically equivalent violation descriptions, general-purpose semantic encoders lack knowledge of inspection-specific terminology and regulatory distinctions, and retrieved precedents are often not directly transformed into actionable disciplinary references. To address these problems, this paper proposes InspectCL, a domain-enhanced contrastive learning and Retrieval-Augmented Generation framework for similar case retrieval, validated on audit data from a provincial power grid company. First, to provide task-specific supervision that is unavailable in existing benchmarks, we construct InspectCase, a de-identified dataset of 4200 audit and compliance cases across 12 violation categories, with expert-validated positive pairs and hard negative pairs. Second, to overcome the weak domain awareness of generic encoders, we design a domain-enhanced contrastive learning model. Specifically, terminology-masking augmentation improves robustness to specialized inspection expressions, regulatory semantic injection incorporates disciplinary rules to distinguish factually similar but legally different cases, and hierarchical contrastive optimization strengthens both case-level similarity learning and category-level boundary separation. Third, to convert retrieved precedents into practical decision support, the Top-K similar cases are used as evidence for a large language model to generate structured disciplinary recommendation summaries, including violation classification, penalty references, applicable regulations, and rectification measures. Experimental results on InspectCase show that InspectCL substantially outperforms BM25, BERT-base, SimCSE, and Legal-BERT baselines, achieving 56.9% ± 0.7% Recall@5 and an 87.6% ± 0.4% Penalty Consistency Score (PCS). These results demonstrate that the proposed problem-driven modules jointly improve semantic retrieval accuracy and disciplinary decision consistency, offering a practical reference for similar power-grid audit scenarios, with broader applicability to be validated in future cross-domain studies. Full article

In this paper, we propose and evaluate a feature distillation technique for object detection under poor visibility conditions, and we analyze its impact when deployed on an FPGA platform. We demonstrate via extensive experiments how different detection architectures generalize across scenes, and we infer that a scale-permuted feature extraction is the ideal choice for detection tasks in unconstrained environments with an 11–12% gain. As verified by the experiments, image enhancement often fails to provide significant detection gains. We hence introduce a joint training in a scale-permuted student network that learns dehazed features from a dual teacher network without an explicit dehazing step. The student learns to replicate not only the teacher outputs but also the decision-making process of the teacher by using attention transfer. Although the overall goal is to produce a real-time system capable of providing driving assistance in challenging scenarios, the FPGA implementation of a scale-permuted network is the first of its kind. To achieve effective implementation of the model in FPGA technology, a high-level synthesis approach and model compression techniques are employed to obtain a deployment with a good trade-off between quality and memory footprint metrics. We develop two distilled models using the joint feature distillation technique and show that these perform better in poor visibility scenes when compared to other detectors with similar size or even bigger sizes in some cases. Our 8.5 M model shows an mAP gain of almost 1% compared to YOLOv10-M with 15 M parameters, on the Cityscapes Hazy dataset. On night images from the BDD dataset, our 8.5 M model shows an approximate mAP gain of 4% compared to YOLO26-S with 9.5 M parameters. We further perform cross-domain testing with the DriveIndia dataset to show that our models generalize well beyond the distillation distribution and can be used for generic driving scenarios. Full article

Precise segmentation of kidney tumors in medical images is crucial for diagnosis, treatment planning, and prognosis assessment. In this work, we present a newly proposed hybrid deep learning model that combines the merits of U-Net and the Swin Transformer architectures in order to enhance the segmentation performance. Although U-Net has great spatial localization ability thanks to the encoder–decoder structure, which works in a hierarchical way, it is still difficult to capture global context well. The Swin Transformer instead captures long-range dependencies and assists in local detail extraction, while attention pooling might also smear fine boundary details. This motivates our hybrid integration. To attempt to resolve these issues, we extend U-Net with the Swin Transformer blocks in the backbone encoder path in order to efficiently perform multi-scale semantic feature extraction while preserving structural consistency. We trained and cross-validated the model on the publicly available Kidney Tumor Segmentation Challenge 2021 (KiTS21) dataset with extensive data augmentation as well as custom loss functions to address class imbalance and boundary obscureness. Experiments demonstrated that it achieved better performance when compared with the solo models, seeking a similar multi-task learning objective on not only U-Net and the Swin Transformer but also other baseline architectures in terms of the average Dice similarity coefficient (average DSC), intersection over union score (IoU) and Hausdorff distance. The proposed model achieved a Dice similarity coefficient (DSC) of 0.91, an IoU of 0.87, a PR-AUC of 0.89, and an overall voxel-wise accuracy of 98%, demonstrating robust and precise kidney tumor segmentation across varying tumor sizes and shapes. Moreover, the integrated solution is more robust and generalizes better, particularly in challenging cases with diverse anatomical variations. These findings demonstrate the power of Transformer-based hybrid models for medical image segmentation. Our results have positive implications for the design of computer-aided diagnostic systems and their association with other prevalent medical imaging tasks besides organ-specific or pathology-focused tasks. Full article

Background and Objectives: Although the anterior mandible is generally considered a safe region for implant placement, injury to the medial lingual foramen (MLF) may result in significant vascular complications. Accurate identification of this structure is challenging due to its small size, low volumetric representation, and anatomical variability. This study aimed to evaluate the anatomical characteristics of the MLF using cone-beam computed tomography (CBCT) and to develop and validate a deep learning-based approach for its automated detection and segmentation. Materials and Methods: A total of 106 CBCT scans were retrospectively analyzed to assess the morphology and position of the MLF. Manual pixel-wise annotations of the complete canal trajectory were performed on sagittal slices and used to train convolutional neural network models based on a U-Net-derived framework. Multiple configurations, including multi-class, binary, two-dimensional, and three-dimensional approaches, were evaluated. Given the extremely limited volumetric representation of the MLF, severe class imbalance represented a major challenge during model training and evaluation. Model performance was assessed using the Dice similarity coefficient, precision, recall, and Hausdorff distance. External validation was performed on an independent dataset of 10 CBCT scans. Results: The MLF was identified in all patients, with a single canal observed in 63% of cases. The sagittal-plane binary segmentation model achieved the best performance, with a test Dice score of 0.79, precision of 0.88, and recall of 0.73. External validation demonstrated a Dice score of 0.81, precision of 0.89, and recall of 0.71. The 95th percentile Hausdorff distance was 2.6 mm, and the mean center-point localization error was 1.2 mm. The model correctly detected the MLF in 90% of external cases. Conclusions: Deep learning-based segmentation of the MLF is feasible and may support automated localization assistance during preoperative CBCT assessment. Performance was influenced by the alignment between the annotation strategy and model input, highlighting an important consideration for small-structure segmentation. Further validation on larger multicenter datasets is required before clinical implementation can be considered. Full article

Background/Objectives: This study explored parents’ perspectives regarding digital media use in children and adolescents with neurodevelopmental disorders (NDs) and examined how these views vary according to family and clinical characteristics. Methods: Data were collected from an Italian survey involving 352 families. Items assessed the perceived effects of digital devices on child development and parenting, awareness of screen time guidelines, and use of time- and content-limiting tools. Quantitative analyses were complemented by a reflexive thematic analysis of open-ended responses describing how digital media influenced parenting. Results: Parents expressed divergent attitudes towards digital media, with broadly similar proportions reporting positive, neutral, and negative views regarding both child development and parenting. More favourable views were associated with greater perceived benefits for children and were more frequent among parents of children with more severe functional disabilities. About half had discussed screen use with health professionals, and most were aware of existing guidelines. Thematic analysis identified six themes related to digital parenting: educational means (digital devices as tools for communication, learning, and socialisation), entertainment (screens as a source of leisure or behavioural management), reward (digital media used as reinforcement), screen time as a “necessity” (technology as an integral and sometimes rehabilitative part of daily life), negative effects on the child (concerns about detachment, reduced social interaction, and mood dysregulation), and parental behaviour and attitudes (reflecting the emotional burden of regulation and broader beliefs about digital media). Conclusions: Parents of children with NDs navigate digital media use through a complex balance of perceived risks and benefits. Findings highlight the need for family-centred guidance and assistive technology approaches that promote digital inclusion while addressing parental stress and regulatory challenges. Full article

The increasing demand for better fruit flavor and eating quality has driven the need for rapid and non-destructive assessment of internal attributes to support fruit grading and precision supply. Visible–near-infrared hyperspectral imaging (VNIR-HSI) provides rich spectral–spatial information for evaluating sweetness in citrus fruit, but its practical use is constrained by high spectral dimensionality, redundancy, and system cost. Here, we propose a reconstruction-assisted, attention-guided band-selection framework for non-destructive prediction of soluble solids content (SSC) in Shimen honey mandarins. The framework integrates spectral–spatial attention, probability-based differentiable band selection, and full-band reconstruction into a unified end-to-end architecture, enabling compact and informative band learning. Using 952 samples, the model selected 56 informative bands from the original 176-band hyperspectral data and achieved competitive SSC prediction on the test set (RMSE = 0.63 °Brix, R² = 0.80) while maintaining high-fidelity reconstruction of the full-band hyperspectral cube from the compact input (peak signal-to-noise ratio, PSNR = 36.47 dB; structural similarity index, SSIM = 0.89). These findings support the proposed framework as a methodological proof of concept for non-destructive citrus quality evaluation, indicating that substantial spectral compression can be achieved under the current VNIR setting while largely preserving predictive performance. The selected bands may provide candidate spectral regions for future compact citrus-quality sensing systems. Full article

Cranial reconstruction is a critical task in computer-assisted surgery, requiring both high geometric accuracy and computational efficiency for patient-specific implant design. While recent deep learning approaches, particularly 3D UNet-based models, have demonstrated promising performance, most studies primarily focus on architectural modifications, with limited attention to the systematic impact of data preparation and training strategies on reconstruction quality. In this study, we present a comprehensive data-centric investigation of key factors influencing the performance of a baseline 3D UNet for cranial defect reconstruction. Specifically, we analyze the effects of data preprocessing (denoising), dataset organization (ordered versus randomized training), defect morphology diversity, convolutional kernel size, and loss function design under controlled experimental conditions. Experiments were conducted on 250 complete skulls (NRRD format) from the MUG500+ dataset, with synthetically generated defects across multiple anatomical regions. From these volumes, a total of 3750 training samples were generated, including: (i) 1250 noisy samples with diverse defect morphologies, (ii) 1250 denoised samples with ellipsoidal defects, and (iii) 1250 denoised samples with multiple defect types. The results demonstrate that data-centric and training-related factors have a substantial impact on model performance, in several cases exceeding the influence of architectural design. In particular, denoising significantly improves boundary stability and reduces geometric error, while incorporating diverse defect morphologies enhances generalization to unseen shapes. Additionally, ordered training contributes to more stable convergence, and an optimal kernel size of (3 × 3 × 3) achieves the best trade-off between accuracy and computational efficiency. A hybrid Dice and boundary loss further improves boundary precision compared to conventional Dice loss. The optimized configuration achieves a Dice Similarity Coefficient of 0.94 and a Hausdorff Distance of 3.8 mm, with an average inference time of 0.004 s per case. These results demonstrate that data-centric optimization can be as influential as, or even more impactful than, architectural design in cranial defect reconstruction. The findings provide practical and reproducible guidelines for developing efficient, robust, and clinically applicable deep learning-based systems for patient-specific cranial implant design. Full article

Motor imagery electroencephalography (MI-EEG) decoding remains challenging in cross-subject scenarios due to pronounced inter-subject variability and signal non-stationarity, which often lead to performance degradation on unseen subjects. Existing prototype-based and domain adaptation methods typically rely on global feature alignment or single-level class representation, limiting their ability to capture both channel-wise spatial variability and high-level semantic structure. To address these limitations, we propose a dual-stage prototype representation framework for cross-subject MI-EEG decoding. The framework models spatial and semantic variability in a hierarchical manner by introducing channel prototypes and feature prototypes, enabling more consistent representations across subjects. Furthermore, a prototype-guided pairwise similarity learning strategy is employed to enhance intra-class compactness and inter-class separability in the embedding space. To mitigate cross-subject distribution shifts, we integrate a lightweight statistical perturbation method (StyleMix) with Wasserstein-based domain alignment, helping reduce subject-specific distribution variations. Experiments on the BCI Competition IV 2a and 2b datasets show that the proposed method achieves competitive performance under the evaluated target-assisted few-shot setting, reaching average accuracies of 79.12% and 87.31%, respectively, and improving over the strongest baseline by up to 2.99 percentage points. Full article

Video compression is central to large-scale video delivery, where better rate–distortion efficiency directly reduces bandwidth and storage cost. A practical way to improve efficiency is to encode a low-resolution video stream with a standard codec and restore high-resolution details with a learned super-resolution model at the decoder. However, prior SR-assisted compression methods usually update the reconstruction model at fixed temporal intervals, which can waste bitrate when those update boundaries do not match actual scene changes. In this paper, we present SASVC, a scene-adaptive super-resolution video compression framework for offline codec-augmented compression. SASVC detects scene changes using frame-wise grayscale differences, updates only compact adapter modules when a content transition is observed, and compresses the resulting model updates with chained differencing, quantization, and entropy coding. In this way, the method reduces unnecessary model-stream overhead while preserving scene-specific reconstruction fidelity. Experimental results on both long-form and short-form datasets show that SASVC consistently outperforms SRVC-style baselines and conventional codec-based alternatives under the Bjontegaard delta rate based on peak signal-to-noise ratio (BD-rate/PSNR) criterion. Complementary rate–distortion (RD) comparisons in terms of structural similarity index measure (SSIM) and Video Multi-Method Assessment Fusion (VMAF) show the same overall trend, indicating that the gain is not limited to a single distortion metric. Specifically, SASVC achieves BD-rate gains of $-41.33\%$ and $-53.49\%$ on Vimeo and Xiph, respectively, and further reaches $-51.53\%$ and $-39.83\%$ on UVG and MCL-JCV. The decoder also maintains real-time 1080p reconstruction at 125 frames per second (FPS) on an NVIDIA RTX 3080 Ti GPU, indicating that scene-aligned model updates can improve compression efficiency while keeping decoder-side deployment practical. Full article

Accurate reservoir dam detection in arid regions is challenging because of spectral similarity between dams and surrounding backgrounds, indistinct boundaries, and substantial target-scale variation. To address these issues, this study proposes a deep learning object detection with multi-conditional constraint assistance (DLOD&MCCA) framework that combines a dual deep enhancement YOLO network (DDE-YOLO) with a multi-conditional constraint assistance (MCCA) strategy. In DDE-YOLO, visible (VIS) and near-infrared (NIR) imagery are fused to enhance cross-spectral discrimination, while task-oriented architectural refinements improve the representation of dam targets with diverse scales and structural characteristics. Meanwhile, the MCCA strategy constrains the search space to geographically plausible candidate regions, thereby reducing background interference and improving detection efficiency. Experiments conducted on the self-constructed S2-Dam dataset and the public DIOR dataset show that DDE-YOLO achieves mAP50 values of 92.8% and 76.2%, respectively, outperforming existing state-of-the-art (SOTA) methods. Furthermore, regional-scale dam mapping in Xinjiang achieved an accuracy of over 95%, demonstrating the effectiveness and practical applicability of the proposed framework for large-scale reservoir dam detection in arid environments. Full article

Background: Prostate cancer (PCa) imposes a substantial global health burden, with robot-assisted radical prostatectomy (RARP) established as the gold standard for localized disease. While da Vinci^® Xi maintains market dominance, Toumai^® MT-1000 offers a potentially cost-competitive alternative lacking prospective validation. Objective: To evaluate perioperative safety, oncologic quality (primary endpoint: positive surgical margins), early functional recovery (continence), and surgeon learning curve between Toumai^® MT-1000 (T-RARP) and da Vinci^® Xi RARP (DV-RARP) performed in high-volume European practice. Materials and Methods: This is a prospective single-center comparative study carried out at Policlinico Gemelli, Rome (May–November 2025), enrolling 80 patients with localized or locally advanced PCa, elected for radical prostatectomy and casually allocated to receive surgery with Toumai or the da Vinci robotic platform. The primary endpoint was the comparison of positive surgical margin (PSM) rates. Secondary endpoints included the comparison of operative time (skin-to-skin), estimated blood loss, length of hospital stay, 45-day postop outcomes, specifically Clavien–Dindo complications, urinary continence recovery (0–1 pad/day), and IIEF-5 scores. Learning curve was evaluated through the cumulative summation (CUSUM) analysis of operative times and linear regression of operative times (n = 80 cases). The analyses used STATA 19 with two-sided tests at p < 0.05 significance. Results: Baseline characteristics showed balance between cohorts (p > 0.05 for most covariates). Perioperative outcomes proved equivalent: median operative time (OT) was 192.5 min (IQR 165–230) for Toumai^® versus 183.5 min (IQR 147–225) for da Vinci^® Xi (p = 0.38); estimated blood loss (EBL) was 150 mL in both groups (p = 0.87); length of hospital stay (LOS) was 2 days in both groups (p = 0.92). PSM rates were identical at 17.5% (p = 0.79). Continence recovery reached 72.5% versus 80% (p = 0.43). Complications (Clavien–Dindo ≥ II) occurred in 7.5% versus 12.5% of cases (p = 0.45). The CUSUM analysis demonstrated operative time proficiency after only four procedures; operative time regression showed no significant trend (p = 0.38). Conclusions: Toumai^® MT-1000 demonstrates similar performance to da Vinci^® Xi across different RARP quality metrics, with no detectable learning curve for surgeons previously experienced with da Vinci. These findings support a safe integration of cost-effective platforms into clinical practice, pending multicenter randomized confirmation. Full article

Objective: Existing deep learning (DL) studies on brain metastasis have largely focused on algorithm or reader performance in controlled settings, whereas its role in routine teleradiology quality support remains unestablished. We evaluated the report-level impact of DL assistance on brain metastasis interpretation in a real-world teleradiology workflow using dual-sequence MRI. Materials and Methods: In this retrospective study, 600 patients who underwent contrast-enhanced dual-sequence brain MRI during two consecutive 3-month periods before (pre-DL, n = 286) and after (post-DL, n = 314) DL integration into teleradiology workflow were analyzed. Ten board-certified teleradiologists interpreted all the cases with or without DL-generated overlays. Report-level diagnostic metrics were assessed against a consensus reference standard established by faculty neuroradiologists. Subsequently, exploratory case-level stratified sensitivity analyses were performed for metastasis-positive examinations based on lesion multiplicity and the largest lesion size. Teleradiologists’ perceptions were assessed using a post-interpretation survey. Results: Compared with the pre-DL group, the post-DL group showed higher sensitivity (77.7% vs. 90.8%, p < 0.001), specificity (82.3% vs. 90.8%, p = 0.002), accuracy (80.8% vs. 90.8%, p < 0.001), positive predictive value (68.2% vs. 85.7%, p < 0.001), and negative predictive value (88.3% vs. 94.2%, p = 0.011). False-positive and false-negative rates were lower after DL implementation (11.9% vs. 5.7%, p = 0.009; 7.3% vs. 3.5%, p = 0.045). Sensitivity gains were most pronounced for cases with single metastasis (74.6% vs. 91.2%, p = 0.007) and with the largest lesion ≤ 5 mm (74.3% vs. 92.0%, p = 0.004), whereas sensitivity was similar for multiple metastases and for cases with a largest lesion > 5 mm. Survey responses suggested favorable usability and diagnostic support. Conclusions: In this real-world teleradiology workflow, DL implementation was associated with higher report-level diagnostic metrics and fewer false interpretations. DL assistance may help support quality control for brain metastasis interpretation, particularly in more subtle and diagnostically challenging cases, although radiologist judgment remains essential for subtle or borderline lesions. Full article

Short-chain fatty acids (SCFAs) are key biomarkers of gut microbiota activity; however, routine quantification in fecal samples relies largely on chromatography, which is instrument-intensive and throughput-limited chromatography techniques. Herein, we present a rapid machine-learning-assisted electroanalysis platform for SCFAs profiling that integrates a disposable three-electrode planar gold chip with voltammetric fingerprinting and artificial neural network (ANN)-based signal decoupling. To generate orthogonal chemical information and improve the discrimination of structurally similar species, a dual pretreatment strategy combining acid-catalyzed esterification and alkaline dissociation was employed prior to electrochemical analyses. Differential pulse voltammetry (DPV) and cyclic voltammetry (CV) were employed to acquire high-dimensional fingerprints, from which current-, potential-, and area-based descriptors were extracted using a cross-information feature strategy. A hierarchical modeling framework improved total SCFAs prediction by incorporating ANN-predicted propionate and butyrate concentrations as auxiliary inputs. While linear calibration was achievable in standard mixtures, direct linear models performed poorly in real fecal matrices due to strong sample-dependent matrix interference. In contrast, the ANN captured nonlinear relationships among multifeature inputs and suppressed matrix effects. Validation against gas chromatography–mass spectrometry in an independent fecal test cohort (n = 30) demonstrated excellent agreement and low prediction errors, with mean absolute error/root mean square error values of 0.063/0.072 mM (propionic acid), 0.029/0.034 mM (butyric acid), and 0.135/0.202 mM (total SCFAs). The DPV/CV acquisition requires only minutes per sample, whereas pretreatment takes 1~3 h depending on the target route but can be performed in parallel for batch processing; thus, overall throughput is determined mainly by batch pretreatment rather than per-sample instrument time. This electrochemical–ANN workflow provides a portable, high-throughput alternative to chromatography for fecal SCFAs profiling in clinical screening and microbiome research. Full article

Reducing traction energy in urban rail systems while preserving safety, punctuality, and passenger comfort remains challenging. Additionally, route-level tram studies that train deep reinforcement learning (DRL) policies using Operational Train Monitoring Recorder (OTMR) logs and benchmark them across multiple objectives remain limited. This study develops and evaluates a Human-Assisted Deep Reinforcement Learning (HADRL) framework for multi-objective tram control in an OTMR-grounded simulation. Two HADRL agents were trained using a human-assistance action mapping: a standard Proximal Policy Optimisation (PPO) baseline and a recurrent, history-augmented PPO. Their performance was compared against that of four human drivers using indices for speed-limit compliance, schedule deviation, traction energy, jerk-based comfort, and stopping accuracy. These performance measures were aggregated using the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) with both equal and entropy-derived weights. Both HADRL agents reproduce the characteristic accelerate–coast–brake driving pattern, reduce traction energy relative to all human baselines, and achieve near-complete speed-limit compliance, all while remaining within the specified schedule-deviation and comfort thresholds. TOPSIS yields identical rankings under both weighting schemes, with Multi-Objective Tram Operation Non-Stationary Proximal Policy Optimisation (MOTO-NSPPO, a recurrent, history-augmented PPO) ranked first and PPO second. Full article