Search Results (13,216)

Access to reliable medical images is essential for clinical training. To address this need, this paper presents an extended version of MIRAGE, a multimodal retrieval and generation system that utilizes a shared latent space to process medical queries by retrieving real images from the ROCO dataset, generating synthetic scans, and providing LLM-based clinical descriptions alongside dual-concept visual comparisons. To overcome previous computational limits and the lack of clinical validation, we introduce three core enhancements: first, an Auto-$\alpha$ module to dynamically weight visual and textual similarities; second, the integration of LCM-LoRA to accelerate synthetic image generation; and third, an automated clinical auditor based on Gemini 2.5 Flash. Experimental results demonstrate that Auto-$\alpha$ improves retrieval accuracy for heterogeneous queries, reaching $38.83$% Top-1 Recall over a 65,419-image gallery and outperforming nine fusion baselines evaluated under a unified configuration, with a controlled ablation attributing most of this gain to learning the weight rather than merely making it query-adaptive, while the LCM-LoRA module reduces computational costs by a factor of 12.5× in CPU environments, with a blinded radiologist evaluation confirming only a small drop in clinical quality. Furthermore, the clinical auditor achieves a 0.805 Pearson correlation against an expert radiologist, effectively correcting the systematic overestimation of traditional CLIP scores. Finally, the optimized platform is publicly deployed on Hugging Face. Full article

Fashion trend forecasting is crucial for proactive supply chain management, sustainable production, and personalized marketing. However, accurately predicting fine-grained fashion trends across diverse user groups remains challenging, as conventional numerical models rely solely on historical popularity signals and cannot explicitly exploit the high-level semantic context underlying trend evolution. To address this limitation, we propose a Dual-Stream Driven Fashion Trend Forecasting (DDFTF) framework that integrates numerical forecasting with textual semantic reasoning. The numerical stream employs a multi-scale patch Transformer with metadata-aware feature fusion to capture temporal dynamics, while the semantic stream converts historical trends into structured textual descriptions, extracting interpretable semantic representations of global trends and turning points. A residual fusion module combines both streams by treating semantic signals as complementary corrections to numerical predictions. Extensive experiments on the FIT and GeoStyle benchmarks demonstrate that DDFTF consistently outperforms state-of-the-art methods, achieving up to a 14.6% relative MAE reduction in long-term forecasting. Ablation and qualitative analyses further show that the performance gains arise from meaningful semantic information rather than increased model capacity, while also improving the interpretability of fashion trend prediction. Our code is publicly available. Full article

Background/Objectives: Hybrid EEG-fNIRS emotion recognition aims to accurately identify an individual’s emotional state by analyzing neurophysiological signals and constitutes an important research direction in affective brain–computer interfaces and human–computer interaction. In recent years, EEG-fNIRS emotion recognition has advanced from handcrafted feature extraction and shallow fusion to deep learning and graph-based modeling. However, most existing methods rely on predefined fixed frequency-band partitioning and second-order graph structures that only support pairwise connections, making it difficult to accommodate inter-subject frequency variability and to characterize high-order brain network relationships such as multi-channel synergistic activation within a frequency band and cross-frequency coupling. Methods: To address these issues, this paper proposes an EEG-fNIRS emotion recognition framework based on a Spatial-Frequency Hypergraph Neural Network (SF-HGNN). First, a Dynamic Frequency Band Decomposition module is designed to achieve adaptive optimization of the EEG and fNIRS frequency bands; second, a Multi-scale Temporal Convolution module extracts temporal features at different time scales; third, a Spatial-Frequency Adaptive Hypergraph Convolution module is constructed to model intra-band cross-channel spatial synergy and channel-wise cross-frequency coupling; and finally, a Cross-Modal Attention Fusion mechanism achieves high-order interaction between the complementary information of the two modalities. The proposed method was validated on the public ENTER dataset comprising 50 participants and four emotion categories (sadness, happiness, fear, and calm). Results: Experimental results show that SF-HGNN achieves accuracies of 82.94% and 68.85% in subject-dependent and subject-independent experiments, respectively; ablation studies and visualization analyses further verify the effectiveness of each module and the interpretability of the model. Conclusions: Future work will focus on validation with larger-scale data and improving cross-subject domain generalization. Full article

Hyperspectral image classification (HSI) requires a model to distinguish subtle spectral differences while preserving the spatial structure of land-cover regions. CNN-based methods are effective for local spectral–spatial extraction, but their limited receptive fields can weaken broader context modelling. Transformer-based methods improve long-range dependency modelling, yet fixed patch partitioning may reduce their sensitivity to fine local structures. To address these limitations, this study proposes the Dual-Stream Spatial–Spectral Network with Nested Attention (DSSN), which separates local spectral–spatial feature extraction from multi-scale spatial-context modelling before adaptive fusion. The DSSN combines a cascaded 3D-CNN spectral stream, a nested Transformer spatial stream with pixel-level and patch-level interactions, and a channel-attention-based adaptive fusion module. Experiments on Indian Pines, Pavia University and Salinas show DSSN achieves overall accuracies of 98.11%%, 99.88% and 99.82%, respectively, outperforming other baselines. The ablation experiments confirm that each major component contributes to the final performance. Although the model requires more parameters and longer inference time than several compared baselines, its inference time remains at the millisecond level. These results suggest that decoupled spatial–spectral representation and adaptive multi-scale fusion can improve hyperspectral image classification under the evaluated benchmark settings. Full article

Remote fault-diagnosis services in dynamic edge–cloud environments depend on timely monitoring-data upload, remote inference, and result delivery, making their communication layer sensitive to variations in available bandwidth, link latency, and packet loss rate. This study addresses the network-state forecasting layer that supports such services rather than the fault-classification model itself. We propose Horizon-Aware Transformer–Ridge Fusion (HATR-Fusion), which combines a nonlinear Transformer expert with a low-variance ridge-regression expert for joint short- and long-horizon forecasting. Historical available bandwidth, link latency, packet loss rate, mobility, and offered load are used as inputs. Preprocessing statistics are estimated using the training set only, and validation-calibrated convex fusion weights are frozen before test inference. Experiments on controlled synthetic trajectories from eight links sampled at 1-min intervals, using 10 neural-network initialization seeds and ten baselines including DLinear and iTransformer, show that HATR-Fusion reduces mean absolute error (MAE) relative to the standalone Transformer by 6.94–8.31% over the 10-min horizon and by 3.02–4.40% over the 60-min horizon, with all six paired improvements remaining significant after Holm correction. Against iTransformer, HATR-Fusion is significantly more accurate for short-horizon bandwidth and latency, whereas iTransformer is significantly more accurate for long-horizon latency and packet loss; short-horizon packet loss and long-horizon bandwidth are not significantly different after Holm correction. The six-task mean normalized mean absolute error (NMAE) is 0.07106 for HATR-Fusion and 0.07047 for iTransformer, indicating comparable overall accuracy with task-dependent differences between the two methods. Ablation results show complementary short- and long-range contributions from ridge regression and Transformer, while input-quality sensitivity analysis identifies a limitation of the fixed fusion weights under corrupted or missing history. The conclusions are therefore restricted to scenarios with relatively stable input quality and distribution shifts comparable to those evaluated in this study. Full article

Mirror segmentation aims to identify mirror pixels from a single RGB image, yet remains challenging because mirrors provide weak intrinsic texture cues and their appearance is dominated by scene-dependent reflections under varying illumination and viewpoints. While recent models improve performance by leveraging contextual contrast, symmetry priors, frequency/spectral cues, or additional modalities (e.g., depth), many cross-cue or symmetry-aware designs still rely on direct spatial-domain fusion, such as concatenation, addition, or attention. Such fusion can amplify reflection-induced high-frequency variations and lead to leakage, shape distortion, and unstable boundaries. In this paper, we propose a symmetry-aware mirror segmentation framework that stabilizes cross-branch interaction via a frequency-domain cross-conditioned fusion mechanism. We build a dual-path Siamese encoder using the original image and its horizontally flipped counterpart, and introduce Heat Conduction Operator-based Cross Fusion (HCOCF), which performs heat-conduction-inspired spectral attenuation in the DCT domain. Unlike conventional fusion, HCOCF generates a nonnegative cross-conditioned attenuation coefficient map from the opposite branch and applies it to the DCT coefficient grid of the target branch. This produces a DCT-domain attenuation mask that controls the spectral refinement strength of each target feature stream, enabling global context propagation while suppressing unstable reflection-induced high-frequency responses without aggressive direct feature mixing. For multi-scale decoding, we adapt the cross-scale decoder of the baseline symmetry-aware architecture by replacing simple addition with conditional feature aggregation, which refines the HCOCF-enhanced features and improves boundary recovery. Extensive experiments on MSD, PMD, and RGBD-Mirror demonstrate competitive performance against representative supervised mirror segmentation methods. In particular, our RGB-only model achieves 88.47% IoU on MSD and 73.72% IoU on PMD, and remains competitive on RGBD-Mirror without using depth input. Full article

This study addresses the challenges of sparse long-range point clouds, complex background interference, and inconsistent localization quality in 3D object detection for intelligent driving in open-pit mines. A camera–LiDAR multimodal detection method based on Voxel R-CNN is proposed. We introduce a Multimodal Focal Sparse Voxel Enhancement module that combines Focal Sparse Convolution with shallow visual features to guide voxel importance prediction and selective sparse propagation. An Intersection over Union (IoU)-Aware Quality and Geometry Refinement Head is further designed to improve the localization accuracy and ranking reliability of 3D proposals. Experimental results show that the proposed method achieves BEV mAP@0.40 and 3D mAP@0.40 values of 63.43% and 61.23%, respectively, outperforming the strongest comparison method, MambaFusion, by 5.37 and 7.61 percentage points. In the 60–80 m range, the average translation error is reduced to 0.41 m, while the inference speed reaches 27 FPS. These results demonstrate that the proposed method improves the detection and localization of distant sparse objects in complex open-pit mine environments while maintaining real-time inference capability. Full article

Objects in images captured by unmanned aerial vehicles (UAVs) are often small in scale, weak in texture, and frequently occluded. Meanwhile, complex backgrounds can produce local responses similar to those of real targets, increasing the risk of missed detections for small objects and false detections in background regions. To address these challenges, this paper proposes DPFS-YOLO, a small-object detection method designed for complex aerial scenes. Built upon YOLOv8n, the proposed method integrates detail enhancement, foreground selection, and semantic guidance into a collaborative optimization framework, aiming to improve small-object feature representation while suppressing spurious background responses. First, the Dual-Path Edge Fusion (DPEF) module combines explicit edge priors, Gaussian smoothing constraints, and a learnable detail modeling branch to enhance object boundaries and local texture representations. Second, the Foreground-Guided Detail Selection (FGDS) module filters the enhanced detail responses through foreground gating, preserving target-relevant information while weakening false activations caused by complex backgrounds. Finally, the Semantic-Aware Guided Fusion for Tiny Object Detection (SAG-tiny) branch adds a P2 high-resolution detection layer and fuses shallow spatial details with top–down high-level semantic information before semantic-guided refinement, enabling small-object details to be constrained by semantic context during detection. Experiments on the VisDrone2019-DET, HIT-UAV, and TinyPerson datasets demonstrate that the proposed method effectively alleviates missed detections of small objects and reduces false-detection risk under complex background conditions. Full article

The demand for energy-efficient deep learning models for autonomous driving highlights the potential of alternative number systems such as Posits. Posits are designed to improve precision, space efficiency, and dynamic range over traditional floating-point representations. However, the adoption of Posits is hindered by a lack of implementation tools. Therefore, we introduce Posits4Torch, a toolchain that integrates Posit quantization into PyTorch, enabling hardware acceleration of Posit arithmetic on field-programmable gate arrays (FPGAs). We then present a case study applying Posits4Torch to CenterFusion, a 3D object detection model benchmarked on the nuScenes dataset. Our results show the FPGA implementation achieved a 915× speedup over CPU-emulated Posit arithmetic for an 8-bit Posit-quantized CenterFusion model, though host-to-FPGA data transport overheads make it $109\times$ slower than GPU floating-point execution. Moreover, a degradation of up to 9.8% in Mean Average Precision (mAP) and 7.7% in the nuScenes Detection Score (NDS) was observed. For 8-bit integer quantization, this degradation was up to 13.5% in mAP and 10.7% in NDS. These results underscore the trade-offs in low-precision computation. Finally, our work bridges the gap between software deep learning frameworks and hardware acceleration for Posit-based models. Future work aims to improve the hardware acceleration pipeline. Full article

To improve depth reconstruction accuracy of streak tube imaging LiDAR (STIL) under weak echo and low-contrast conditions in complex scattering environments, this paper proposes a hierarchical reliability-guided multispectral polarization reconstruction framework (MSP-STIL). The proposed method addresses measurement uncertainty in multi-wavelength and multi-polarization observations by constructing a progressive reliability modeling strategy, which evolves from polarization stability to statistical uncertainty and finally to signal strength enhancement. First, a Dual-channel Polarization Contrast (Pc) is introduced to evaluate local scattering stability and suppress fringe peak degradation caused by polarization distortion. Second, SNR is employed to model the uncertainty of depth measurements across different wavelength channels. Finally, echo intensity is incorporated as a confidence refinement factor to further enhance high-quality signals. Based on this hierarchical modeling strategy, an adaptive inverse-variance weighting scheme is developed to achieve robust multi-wavelength depth fusion. The results show that the proposed method outperforms equal-weight and single-feature methods in all test regions. Compared with the non-weighted method, the MAE, RE, MSE, RMSE, and STD are reduced by approximately 10.95%, 12.02%, 25.44%, 13.68%, and 15.58% on average, respectively. Under low signal-to-noise conditions (simulated by controlled noise levels) and long-distance detection scenarios, the proposed method still maintains low reconstruction errors and effectively suppresses depth fluctuations, demonstrating good noise resistance and distance robustness. In addition, the maximum contrast of the RGB image constructed through weighted fusion increases from 5.0065 to 5.8300, verifying the effectiveness of the proposed method in low-contrast complex scenes. Overall, the proposed MSP-STIL multi-feature joint weighting method shows clear advantages in reconstruction accuracy, robustness, and scene adaptability. Full article

Nuclear energy production requires robust shielding solutions to ensure the safety of personnel and equipment across a wide range of reactor technologies. Shield for Fusion (S4F) is a new development in nuclear analysis models, and codes a computational framework for optimizing radiation shielding design in fusion reactor applications using OpenMC neutronics simulations and genetic algorithms (GAs). The framework addresses four-layer shielding cabinets with six commercially viable materials, exploring 10,000 possible configurations through a template-based approach with density correction factors. The genetic algorithm efficiently converges to optimal solutions in approximately 100 simulations by balancing competing objectives: neutron flux attenuation, dose minimization, material cost, and weight. Compared to uniform sampling, the GA demonstrates superior performance by identifying high-performing parameter regions with improved attenuation and reduced cost. The framework includes an inverse design tool for rapid retrieval of configurations matching target performance criteria requested by external users, currently available for selected neutron source spectra. Results demonstrate effective multi-objective optimization for fusion shielding applications, with the open-source code available for adaptation to different simulation setups. Full article

Background/Objectives: Posterior cervical decompression and fusion (PCDF) carry perioperative risks and increases postoperative healthcare utilization (HU). Traditional prediction models emphasize comorbidities and surgical factors, yet social determinants of health (SDHs) are known to also affect clinical outcomes. We applied machine learning (ML) to integrate SDHs and clinical variables in predicting 90-day readmission and HU after PCDF. Methods: We conducted a retrospective, single-institution machine learning analysis of adult patients undergoing single or multilevel PCDF (2003–2023). Models were designed using 88 clinical variables and five census-derived Social Vulnerability Index (SVI) scores. Outcomes were 90-day readmission and HU (the unweighted sum of 18 post-discharge components including urgent visits, invasive procedures, non-routine testing, and imaging). Models were trained on 50 repeated 80/20 train–test splits with training-only preprocessing and hyperparameter tuning. Results: Among 1015 patients (mean age of 64.7, 59.0% male, 97.2% with multilevel fusions), 90-day readmission was 15.5% and mean HU score was 13.6 ± 9.0. Regarding readmission, the clinical-only logistic regression model was the best predictor (AUROC 0.65). For HU, clinical-only random forest performed best (MAE: 4.80, R²: 0.326). Adding SVI to the matched clinical-plus-SDH models did not improve prediction for either 90-day readmission or healthcare utilization. Length of stay, year of surgery, and several SVI measures were prominent contributors across both SHAP analyses. Conclusions: ML models integrating SDH and clinical factors available by index discharge modestly predicted readmission and HU after PCDF. However, adding SVI did not meaningfully increase their overall performance, and the models require external validation before clinical use. Full article

Degenerative lumbar disorders are commonly treated with lumbar interbody fusion, but comparative evidence among ALIF, OLIF, and XLIF remains limited. This prospective multicenter observational cohort study compared perioperative outcomes, complications, and patient-reported outcomes (PROMs) after single- or multi-level ALIF, OLIF, or XLIF in 242 consecutive adults (108, 75, and 59, respectively) treated at Spanish centers. Age and surgical indications differed significantly across groups (p < 0.05), reflecting indication-based approach selection. Primary outcomes were intra- and postoperative complications through 6 months; secondary outcomes included operative time, blood loss, visual analog scale (VAS) pain scores, and the Oswestry Disability Index (ODI). XLIF had the shortest operative time (median, 63 min) and lowest blood loss. Overall intra- and postoperative complication rates were 3.3% and 7.8%, with no significant overall differences among techniques, although patterns differed: vascular injuries were associated with ALIF, cage subsidence was more frequent after XLIF, neurological deficits occurred with similar frequency in the XLIF and OLIF groups, and OLIF had the lowest intraoperative rate. All groups showed significant VAS and ODI improvements, with greater 6-month gains after ALIF and OLIF than after XLIF. These approaches are effective but have distinct risk–benefit profiles, supporting individualized selection. Full article

In precision agriculture, fruit segmentation serves as a fundamental visual prerequisite for orchard robotic operations, including precision management, automated harvesting and yield estimation. For unripe passion fruits, complex scene interference, high fruit–leaf similarity and foliage occlusion in practical orchard scenes easily cause missed detection and over-segmentation in existing models. To address these challenges, we build a multi-scene unripe-passion-fruit dataset named ZKMPF. Based on the UNet architecture, we propose Environment-Adaptive and Prompt-Fusion UNet (EAPF-UNet). First, EAPF-UNet embeds an Environmental Adapter into the encoder, which adjusts feature parameters to mitigate complex scenes interference. Then, it incorporates a Localization Multi-Scale Fusion Module (LMSM) and Refinement Multi-Scale Fusion Module (LMFM) to achieve accurate localization and refinement of unripe passion fruits and address scale variations. Finally, it designs the Multi-Dimensional Prompt Fusion module that integrates color, geometry and texture priors to improve the feature discriminability between unripe passion fruits and background. We conduct experiments on our self-built dataset, comparing EAPF-UNet with eight other segmentation models. Evaluated against eight segmentation models, EAPF-UNet obtains mDice of 85.51% and mIoU of 77.16% across six metrics and achieves competitive segmentation results within this dataset’s test scenes. Full article

Predicting driver-state evolution in extra-long tunnel corridors remains challenging because of prolonged spatial confinement and repeated lighting transitions. This study uses a statistical human–vehicle composite, the comprehensive driving index (CDI), as a reproducible quantitative target for predictive auditing. In this study, multimodal information denotes synchronized ocular, physiological, vehicle-motion, and environmental sensor signals; the objective is to quantify their incremental predictive value rather than introduce a new fusion architecture. Fully nested leave-one-driver-out cross-validation with a prespecified 120 s unsupervised initialization estimated all preprocessing, scaling, PCA, model-selection, and calibration parameters from training data only. The five components explained 60.36% of target variance. In the original-range 30 s task (4835 evaluation windows), history-only ridge regression achieved an RMSE of 0.08294 and an R² of 0.166, while directly tuned AR achieved an RMSE of 0.08261 and an R² of 0.168. On the common 4259-window sample, expanded ridge and AR achieved RMSEs of 0.08123 (R² 0.187) and 0.08153 (R² 0.182). Adding coarse scene information produced an ΔRMSE = +0.00002 (95% CI −0.00026 to 0.00027), whereas external environmental summaries produced an ΔRMSE = −0.00019 (95% CI −0.00037 to −0.00003). HistGradientBoosting did not improve performance. The primary contribution is a leakage-controlled predictive-audit framework for screening candidate information sources before deployment decisions. Full article