Stock price analysis based solely on technical indicators, such as in traditional stock price prediction studies, relies heavily on observations from individual sample markets and fails to consider insights from other modalities, making it imperative to study the implicit interdependence between different modalities. A Candlestick-Integrated Multi-Modal Trading System (CIMTS) is proposed that integrates 1D numerical time-series with 2D spatial candlestick chart-image data. The system analyzes multimodal data by employing Long Short-Term Memory and Vision Transformer models in parallel, identifying graphical signals of stock prices using a late fusion approach to merge the outputs from the two models. The system was tested over an out-of-sample period of approximately 3 years, partitioned from a 12-year dataset of key US and Korean stock indices, including the NASDAQ Composite, S&P500, KOSPI, and KOSDAQ, under different market conditions. The experimental results showed that the CIMTS outperformed single-modality models in terms of aggressiveness (i.e., higher cumulative returns and a greater number of trades) and stability (i.e., lower maximum drawdown and consistent win rates). In particular, the CIMTS outperformed the labeling method that considers next-day movements in the Korean market and the 5-day consecutive trend in the US market. The proposed system achieved cumulative returns of 42.52%, 51.37%, 39.36%, and 95.47% for NASDAQ, S&P500, KOSPI, and KOSDAQ, respectively (payoff ratios: 1.023, 0.877, 1.166, and 1.195; profit factors: 1.364, 1.638, 1.306, and 1.890, respectively), demonstrating consistently improved trading performance over the Buy & Hold benchmark across all tested market regimes.
Background Network Traffic Classification (NTC) is a critical process in network management as it increases visibility to provide enhanced security, optimised resource allocation, and quality of service. However, the escalating volume of network traffic and real-time applications in modern networking is straining traditional processing on high-performance servers. To meet modern requirements for low latency and bandwidth efficiency, network management is increasingly moving computation closer to sources that generate the data, a paradigm referred to as edge computing. This shift relies on resource-constrained devices such as routers, switches, and embedded systems. To address these constraints, researchers proposed lightweight NTC models designed to operate effectively in such environments. However, while these methods offer greater efficiency, they often compromise performance, making the models less effective in practice. Furthermore, despite being lightweight, the throughput of current NTC models remains insufficient to keep up with the high bandwidth demands of modern networks, which continues to limit their practical deployment. Methods Therefore, this article proposes LiteNet, a lightweight model for real-time NTC in a resource-constrained environment to address the disparity between model complexity and edge device limitations. LiteNet adapts the InceptionV1 module that integrates four dedicated parallel branches. Additionally, we integrate DeepSHAP, an Explainable Artificial Intelligence (XAI) technique, to enhance explainability and reduce input features. Furthermore, we leverage model compression techniques, including semi-structured pruning and quantisation, to form a sparse LiteNet architecture that enables lightweight design and real-time classification. Results LiteNet demonstrated F1-scores exceeding 99% across two datasets, ISCXVPN2016 and MalayaNetwork_GT, while having 40% fewer parameters compared to the baseline method. Benchmarking results demonstrated that LiteNet outperformed the state-of-the-art method in terms of performance and efficiency, delivering up to 238% higher throughput and lower resource utilisation on an NVIDIA Jetson Nano. These outcomes demonstrate LiteNet’s strong balance between performance and efficiency, establishing it as a viable lightweight solution for NTC in resource-constrained settings.
In the semiconductor manufacturing process, wafermap defect classification is crucial for ensuring product quality and reliability. Despite recent advances, underrepresented defect types in wafer datasets continue to suffer from poor performance due to severe class imbalance, and models with large parameter counts present high computational costs. To address these challenges, DiffWaferNet, a lightweight differential attention-based network, is proposed to improve the performance of minority defect classes while preserving the accuracy of majority classes. In this study, a Conditional Variational Autoencoder (CVAE) is employed to generate class-consistent synthetic data for underrepresented defect types, effectively balancing the dataset and enhancing model robustness. In the classification phase, DiffWaferNet utilizes a convolutional neural network (CNN) for robust feature extraction and applies a differential attention mechanism to enhance focus on critical regions of wafer maps. This mechanism empowers the model to dynamically learn attention weights that highlight the most relevant features for defect classification. Experimental results demonstrate that DiffWaferNet significantly improves classification performance, particularly for minor defect types, achieving consistent performance across all defect classes. Compared to baseline models, the proposed method ensures both robustness and high overall accuracy in wafer map defect classification. Furthermore, extensive benchmarking on the MixedWM-38K dataset validates the generalization capability of DiffWaferNet, achieving an accuracy of 97.88% and an F1-score of 97.76%. With only 121K parameters and 21.8M FLOPs (Floating-point Operations), the model maintains high computational efficiency, making it well-suited for real-time industrial applications. This research highlights a strong potential for advancing defect classification and establishes a foundation for future studies.
This research work presents an integrated computer vision framework for automated analysis of student behaviour in real classroom environments, offering a practical solution to a pressing issue. A significant problem with existing research on predicting user behaviour is that most methods rely on a single instance image for simplicity. Some works, however, utilise group images/videos, which lag in capturing socially coupled behaviours. Student behaviour in the classroom has fundamental dynamics, and we need a model to understand it in the context of crucial collective alignment. To synchronise the phases of behaviours, we introduce the Kuramoto function from physics as a synchrony regularizer for the deep learning model, allowing it to synchronise the nonlinear complex phases while learning about the group of people. The proposed pipeline combines robust object detection with deep sequential classification to recognise fine-grained behaviours such as writing, listening, reading, or discussing. In the detection stage, You Only Look Once Version 8 (YOLOv8) and Faster Region-based Convolutional Neural Network (Faster R-CNN) models are trained to localise students accurately under occlusion, clutter, and varying classroom conditions. Cropped student regions are then passed to a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) CNN + LSTM + ATTENTION classifier that captures both spatial features and temporal dependencies while highlighting salient frames through an attention mechanism. To further enhance temporal consistency and interpretability, novel regularisers, (i) temporal smoothness for reducing behavioural fluctuations, (ii) Kuramoto synchrony that adopts a non-linear dynamics model of phase alignment to encourage coherent behavioural patterns in a group of students, and (iii) attention entropy helps to prevent degenerate focus on single frame, which are incorporated during training. Experimental results on the UK dataset demonstrate that the pipeline achieves high detection precision (0.961) and recall (0.975) with YOLOv8 and strong behaviour classification performance (validation accuracy: 97.31%), surpassing baseline CNN- or LSTM-only models. This research contributes a scalable, interpretable, and deployment-ready methodology for smart classrooms, reducing reliance on manual observation and enabling real-time behavioural insights to support teaching effectiveness and student engagement monitoring.
Rooftop photovoltaic (PV) systems exhibit significant variability in short-term electricity generation due to differences in panel tilt and azimuth angles. Neglecting this geometric diversity may reduce forecasting accuracy, particularly in distribution networks with heterogeneous rooftop PV installations. This study investigates short-term rooftop PV forecasting using an orientation-aware modeling framework based on a real-scale experimental dataset. Five 300 W monocrystalline PV panels were monitored hourly for approximately 1 year under four different tilt angles and seven azimuth angles, representing 28 distinct orientation configurations. Temporal variables (the cosine of day and hour), meteorological variables, and the cosine of the solar incidence angle were used as input features. Three forecasting model structures (Models I–III), each defined by a different set of input features, were developed. Each model was tested using the Persistence Method (PM), Long Short-Term Memory (LSTM), and Gradient Boosted Regression Trees (GBRT). Among the evaluated models, Model II, which included the cosine of the solar incidence angle, achieved the best overall performance. In this model, GBRT outperformed both LSTM and PM by yielding the lowest mean test root mean square error (RMSE) (13.56 ± 2.18 Wh), the highest mean test R 2 (0.973 ± 0.009), and the lowest mean test mean absolute percentage error (MAPE) (18.04 ± 3.65). The results demonstrate that incorporating physically meaningful orientation-aware features substantially improves forecasting accuracy for heterogeneous rooftop PV systems. These findings highlight the importance of panel orientation in hour-ahead rooftop PV forecasting. Consequently, the proposed approach can improve the accuracy of distributed PV generation forecasts and net demand forecasts at the distribution level compared to traditional aggregation-based approaches relying on general solar radiation data.
Arabic poetry is a cornerstone of Arab cultural and linguistic heritage, yet the computational analysis of spoken Arabic poetry remains critically underexplored. An open question is: how can the meter of a spoken Arabic poem be automatically identified when labeled acoustic data is scarce? Meter identification from audio is more challenging than from text because it must handle both acoustic variability and the complex prosodic rules of Aroud—the classical Arabic science of poetic meter—across all sixteen canonical meters. To the best of our knowledge, no publicly available evaluation covering all sixteen meters from acoustic recordings currently exists, and no standardized benchmark exists for this task. We investigate two methodological approaches: an end-to-end architecture that fine-tunes a pretrained Wav2Vec2 model with a meter classification head, and a novel two-stage integrated framework that chains a high-resource Arabic speech recognition system with a high-resource textual meter classifier, augmented by a domain-specific 4-gram language model trained on Arabic poetry. Both approaches are validated on a controlled baseline set of annotated recordings and a benchmark collected under diverse real-world acoustic conditions. The integrated framework achieves up to 92% accuracy on the benchmark set, demonstrating that combining high-resource systems from adjacent domains can effectively compensate for low-resource constraints. Beyond establishing a strong benchmark result for acoustic Arabic poetry meter classification, this work releases a reusable benchmark that standardizes evaluation for future research in computational Arabic literary heritage.
High-Speed Railway (HSR) bearing health monitoring is essential for ensuring operational safety, reducing unexpected service interruptions, and supporting condition-based maintenance in practical railway systems. In real engineering environments, however, bearing fault diagnosis remains challenging because field vibration signals are usually unlabeled, affected by changing operating conditions, and may contain superimposed fault components rather than isolated single fault patterns. To address these challenges, this study proposes a lightweight and interpretable transfer learning framework for HSR bearing fault diagnosis under limited labeled data and compound fault conditions. The proposed framework integrates data cleaning, resampling to 24 kHz according to the Nyquist sampling theorem, wavelet threshold denoising, and multi-domain feature extraction from time, frequency, and time-frequency domains. The discriminative capability of the extracted 25-dimensional feature representation was verified using Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Random Forest, and Light Gradient Boosting Machine (LightGBM) models, all of which achieved source domain classification accuracies above 99%. To reduce the gap between laboratory single fault data and unlabeled field compound fault data, a feature-level Dirichlet-weighted nonlinear synthesis strategy was developed to generate synthetic composite fault feature vectors from source domain samples. A lightweight fully connected regressor was then trained to estimate proportional fault contributions for target domain samples, enabling probabilistic diagnosis without target domain labels. The effectiveness of domain adaptation was confirmed by t-distributed Stochastic Neighbor Embedding (t-SNE) visualization, which showed clear alignment between source and target feature distributions. External validation on a public multi-domain compound fault dataset achieved 80.99% accuracy, with a Mean Squared Error (MSE) of 0.0799 and a Mean Absolute Error (MAE) of 0.1452. These results indicate that the proposed framework can identify bearing fault components under real compound machine fault interference and provide a practical solution for HSR bearing health monitoring, early fault warning, and maintenance decision support when labeled field data are limited.
Human trajectory prediction has significant practical applications in various scenarios, such as autonomous driving, social robots and so on. Recently, it has been widely studied by diffusion models in order to model the inherent multi-modality of human motions. However, existing diffusion-based approaches only focus on modeling the social interactions via a single encoder and neglect the scene interactions, which results in producing unreasonable trajectories across obstacles or road boundaries. To address this issue, we propose the Interaction-Aware Diffusion Model (IADM), a novel diffusion-based framework considering both human motions and surrounding scene layout by treating the social and scene interactions as conditions in the parameterized reverse Markov chain. To implement IADM, we design two encoders, i.e., social encoder and scene encoder, where the social encoder models the social interactions via attention mechanism, and the scene encoder preserves spatial information of the scene when learning the scene interactions. Furthermore, we devise the dual-guidance decoder consisting of the motion-guided temporal module and the scene-guided spatial module to intensify the collaboratively guidance of the social and scene interactions. Extensive experiments on the ETH/UCY dataset, Stanford Drone Dataset and Intersection Drone Dataset validate the superiority of our method, achieving state-of-the-art results.
Cross-border e-commerce has expanded rapidly in recent years, increasing the complexity of sustainable supply chain management (SSCM) and amplifying operational risks that are difficult to assess using traditional approaches. This article investigates sustainable supply chain (SSC) risk prediction in cross-border e-commerce and proposes a hybrid, data-driven assessment framework that integrates expert knowledge with machine learning. First, qualitative and multi-source risk indicators are structured and quantified using the Analytic Hierarchy Process (AHP) to obtain comparable numerical representations. Second, a support vector machine (SVM) model is constructed for risk prediction, and its key hyperparameters are optimized using meta-heuristic search to enhance predictive performance. Experimental results based on the enterprise dataset indicate that the proposed genetic algorithm-particle swarm optimization-support vector machine (GA-PSO-SVM) approach outperforms all benchmark methods, achieving an approximately 20% reduction in root mean square error (RMSE) compared with the non-optimized baseline. Overall, the proposed method offers an interpretable and practically applicable framework for intelligent SSCM risk assessment in cross-border e-commerce, thereby providing methodological support for risk monitoring and decision-making in sustainable supply chains.
Integrating multi-omics data to understand biological processes in human diseases is a complex bioinformatic task. Machine learning (ML), particularly deep learning (DL) models, offers a promising approach to multi-omics data integration and analysis. However, existing DL models generally integrate multi-omics data by concatenating the input data space or learned feature space, which is a sub-optimal approach. In addition, single classifiers are commonly used in DL-based methods, which can compromise the performance. Furthermore, the gradient descent optimization technique in DL suffers from a high computational cost and local sub-optimal solutions. To address these challenges, this article presents a novel cancer subtype classification framework using multi-omics integration and an ensemble-based parallel DL/ML architecture. Specifically, a multimodal autoencoder is used for effective feature learning across omics types, overcoming the limitations of naïve concatenation. A hybrid ensemble model comprising DL and ML learners with a meta-learner enhances classification robustness beyond single models. To improve optimization and computation, we incorporate a hybrid Back-Propagation and Particle Swarm Optimization (PSO) strategy and execute the entire framework on a parallel processing platform, reducing computation time while enhancing global search capability. The proposed framework is evaluated empirically with two benchmark data sets from The Cancer Genome Atlas (TCGA), namely the TCGA Pan-cancer and TCGA Breast Invasive Carcinoma (BRCA) data sets. The results indicate a high performance with accuracy rates of 89.51% and 90.9% for TCGA Pan-cancer and TCGA BRCA, respectively. The parallel implementation of the proposed framework reduces the computation time, resulting in a speed-up of 3 times and 2.5 times for TCGA Pan-cancer and TCGA BRCA, respectively. The findings ascertain the efficacy of the proposed framework for the classification of cancer subtypes, offering a promising solution for implementation in real-world environments.
Investments are the foundation of financial markets, directing resources toward activities that generate future returns. Stock trading plays a central role in wealth creation, yet forecasting stock prices remains difficult due to the nonlinear and volatile nature of financial data. This study presents an ensemble deep learning framework for short-term stock price forecasting in the Saudi market. It addresses the limited use of advanced ensemble methods in emerging economies. The framework uses multivariate data from ten companies listed on the Tadawul All Share Index and compares two ensemble strategies: averaging and stacking with standalone models (Gated Recurrent Unit, Long Short-Term Memory, and Bidirectional Recurrent Neural Network). A total of 34 engineered technical indicators and a rolling-window setup were applied for five-day-ahead predictions. The results show that ensemble models consistently outperform individual architectures across all evaluation metrics. Stacking ensembles, especially those combining Gated Recurrent Unit and Long Short-Term Memory, achieved the highest accuracy. Averaging ensembles provided stable and efficient alternatives. Overall, the findings confirm the robustness and adaptability of ensemble methods for financial forecasting and offer practical insights for investors and policymakers seeking reliable, data-driven prediction tools.
Automated silicon wafer defect inspection is critical for sustaining yield and reliability in semiconductor manufacturing. Traditional image processing and feature engineering approaches struggle with complex and evolving defect patterns, while deep learning methods, although effective, increase computational cost and risk of overfitting. Many existing models address class imbalance through synthetic data, which can introduce artifacts and misrepresent the actual defect distributions. Additionally, conventional convolutional neural networks remain limited in capturing global context, leading to the misclassification of visually similar defects. This study addresses these issues through a simple yet effective two-stage approach. The proposed ensemble attention convolutional neural network (EACNN) first performs binary defect detection to distinguish between defective and non-defective samples, followed by a dedicated defect-type classification for fine-grained categorization. This staged design decomposes the complex multi-class problem into manageable subtasks and mitigates class dominance without synthetic data. The model further incorporates spatial and channel attention mechanisms to strengthen discriminative feature learning and improve the representation of salient defect regions. Focal loss is employed to mitigate class imbalance by down-weighting easily classified examples and placing greater emphasis on hard, misclassified samples. Optimization strategies, including batch normalization, dropout, and global average pooling, are adopted to improve generalization, reduce parameter complexity, and accelerate convergence. Experiments on the WM-811K dataset through 10-fold cross-validation demonstrated the effectiveness of EACNN and achieved 95.57% accuracy, 93.67% precision, 95.41% recall, 98.76% specificity, 94.47% F-measure, and 94.51% G-Mean. Comparisons with classical, attention-based, and imbalance-handling methods confirmed the superior performance of EACNN. The results underscore EACNN’s ability to learn highly discriminative features while preserving the integrity of real defect distributions.
This article presents CAD2TechSpec, a novel framework for automating design processes within computer-aided design (CAD) systems by leveraging multimodal large language models (LLMs). The framework enables the analysis and generation of detailed design specifications, including the automated creation of machining process plans. Our system architecture combines 3D model rendering, dimensionality reduction techniques, and the capabilities of multimodal LLMs to produce structured JavaScript Object Notation (JSON) representations of manufacturing workflows. Experiments conducted on a big CAD model (ABC) dataset demonstrate that CAD2TechSpec significantly reduces design time while enhancing the accuracy and completeness of technical specifications. The proposed approach holds considerable promise for high-tech industries such as precision manufacturing and mechanical engineering, where efficiency and precision in design processes are critical.
Background Character recognition in ancient Tamil inscriptions is a difficult process because of the complex nature, overlapping, and varied sizes of the characters. Traditional Optical Character Recognition (OCR) methods cannot perform optimally in such situations, making it important to develop an efficient character recognition system. Methodology The proposed system is named CUGRA, which represents the integration of five models, namely, Canny, U-Net, Graph Neural Networks (GNN), Recurrent Neural Networks (RNN), and attention models. Its application for Tamil inscription recognition is referred to as CUGRA-TIR. The process starts with the application of the Canny model for enhancement, followed by the application of the U-Net model for segmentation. A Convolutional Neural Network (CNN) is used for feature extraction, and the features extracted are converted into graph representations by the application of the GNN model, followed by the application of the RNN model with an attention mechanism. The proposed approach mainly focuses on the recognition of characters after preprocessing and segmentation and does not address reconstruction or restoration of the characters. Results On an expanded dataset of ancient Tamil inscriptions, the model was tested against conventional OCR systems and existing deep learning techniques. In five separate trials for model training, a mean accuracy of 96.12% was recorded for model validation, with a median accuracy of 98.98% and an interquartile range of 9.18%. In a separate test set, a total accuracy of 84% was recorded for classification, with a weighted F1-score of 0.83. Conclusion From the results, it is clear that the combined spatial, structural, temporal recognition model capable of reading inscriptions is now much more accurate. The data and code used to make the model are available on GitHub, enabling others to scrutinize them and utilize them to further enhance this work.
Data publishers typically apply anonymization techniques to safeguard privacy when sharing data for secondary analysis. However, the current models for estimating reidentification risk fall short of accurately assessing the potential privacy risks associated with disclosed data. For instance, in this article, we show that existing risk estimation models fail to precisely estimate population uniqueness risks as they overlook sensitive and other attributes. To address such problems, we present a generative, copula-based model that can precisely estimate the probability of reidentifying a specific individual based on population uniqueness, taking into account quasi-identifying and sensitive attributes. Further, we prove that risk estimation models should include the percentage of high-risk records in the dataset to provide additional insights into the reidentification risks. We applied the model to 12 real-world datasets and our results show over 91% true positive recognition and less than 19% error rate for high-risk records. This error rate is notably lower than the current best rate of 40%. Further, we found that 43% of Brazilians can be uniquely identified using the five attributes combination and 79% of the United States population can be uniquely identified using the nine attributes considered. Our results show that the proposed model shows more than 81% accuracy in estimating the reidentification risk of high-risk records. Similarly, we emphasize the importance of standardizing the inclusion of the maximum risk of the disclosing dataset in the estimation of reidentification risk.
Predicting medical insurance costs from increasingly complex healthcare data requires models that can capture nonlinear relationships among demographic, behavioural, and health-related factors. Existing attention-based approaches for tabular data typically model either additive or multiplicative feature interactions independently, limiting their ability to capture complementary patterns. We constructed a computational model to predict medical insurance costs based on Arithmetic Multi-attention Transformer (AMFormer), a deep learning architecture that introduces a dual-stream arithmetic attention mechanism jointly modeling additive and multiplicative interactions within a unified framework. This design enables the model to capture both linear feature contributions and their complex multiplicative dependencies through logarithmic-space transformations, distinguishing it from prior approaches that treat these interactions separately. We evaluate our model on a benchmark medical insurance dataset, comparing it against representative traditional machine learning methods and state-of-the-art deep learning baselines. Our model achieves better performance compared to other methods. Ablation analyses demonstrate that the synergistic combination of additive and multiplicative attention streams with logarithmic transformation is critical for these performance gains.
Background Client-side web analytics is often interpreted as if recorded events directly reflect user behavior, although actual measurement depends on multiple observable stages. Because Google Analytics 4 (GA4) and Google Tag Manager (GTM) function as black boxes from the analyst side, practical improvement requires focusing on controllable site-side conditions. This study examined whether disagreement across observable layers becomes visible under controlled conditions and whether that profile varies by dwell time, click type, operating system, and browser. Methods A Three-Layer Monitoring System (3LMS) was developed to record dataLayer events, GA4-related payload requests, and event counts displayed in the GA4 interface. Using Python/Playwright, 900 automated trials were conducted on a minimal test site across three click types and three dwell-time conditions under Windows 10 + Chrome, Windows 11 + Chrome, and Windows 11 + Firefox. Pairwise layer comparisons were evaluated statistically where applicable. Supplementary analyses added pseudo-tag logging, URL-change monitoring, and ultra-short dwell-time conditions in Windows 11 + Chrome. Results Disagreement was most evident under short dwell times. In both Chrome environments, the 1 s condition produced the largest discrepancies, whereas agreement improved at 2 s and was largely complete at 3 s. Under Windows 10 + Chrome, retention at 1 s declined progressively from the dataLayer to the payload layer and then to the GA4 layer. Under Windows 11 + Chrome, the 1 s discrepancy was concentrated mainly at the final GA4 layer despite complete upstream retention. Under Windows 11 + Firefox, upstream monitoring appeared incomplete despite complete downstream GA4 retention, indicating that cross-browser comparability of observability cannot be assumed automatically. Conclusions This study developed a 3LMS as a quality assurance (QA) framework for examining disagreement across publicly observable stages of the GA4/GTM measurement pathway under controlled site-side conditions. The findings support its use as a bounded comparative framework within the tested environment. The 3LMS may serve as a bounded observational QA framework for evaluating site-side conditions associated with layer-wise disagreement in GA4/GTM measurement.
Liver tumors impose a significant global health burden, underscoring the urgent need for efficient and accurate diagnostic methods. Computer-assisted techniques, particularly deep learning-based segmentation models, have shown considerable promise in this domain. However, they continue to face persistent challenges in liver tumor segmentation, including severe background noise interference, indistinct lesion boundaries, and the difficulty of simultaneously improving segmentation accuracy while maintaining a balanced trade-off between Recall and Precision. To address these issues, this article proposes a novel residual “Enhancement-Calibration” U-Net architecture, termed REC-UNet. The model consists of two task-specific modules: a Residual Enhancement Module (REM) and a Calibration Module (CM). REM leverages residual connections and cross-dimensional attention to enhance tumor feature representation for accurate segmentation, thereby establishing a foundation for balancing Recall and Precision. CM further mitigates noise propagation from shallow to deep feature layers, refining segmentation precision while sustaining high levels of both Recall and Precision. Experiments on the LiTS2017 and MSD_Task08 liver tumor datasets demonstrate that REC-UNet achieves superior performance over mainstream models, with a 4.34% improvement in Dice and a 4.24% improvement in Intersection over Union (IoU) over the second-best model (VM-UNet) on LiTS2017. We further validate the model on an in-house clinical liver tumor Magnetic Resonance Imaging (MRI) dataset, where it attains a Dice score of 87.88% and an IoU of 86.91%, while maintaining a well-balanced trade-off between Recall and Precision. Importantly, REC-UNet achieves high overall segmentation accuracy across diverse lesion sizes and contrast conditions without relying on explicit size-stratified optimization. These results confirm the robust generalizability of REC-UNet and highlight its significant clinical value for computer-assisted liver tumor diagnosis.
Accurate traffic flow prediction at specific road segments is essential for optimizing signal control, mitigating congestion, and improving the efficiency of urban transportation systems. Mobile Crowdsensing (MCS) enables large-scale monitoring by collecting geotagged data from participating vehicles and aggregating them at a centralized server. However, most existing solutions for privacy preservation introduce noise or coarse spatial aggregation of location data, which distorts spatiotemporal patterns and degrades the utility of prediction models, while users remain vulnerable to deanonymization and trajectory re-identification attacks. Approaches based on Differential Privacy (DP) offer formal guarantees by injecting calibrated noise into trajectories or model updates, but this perturbation is particularly harmful for short-horizon traffic flow prediction, where detailed local patterns are crucial. We instead shift from perturbation to encryption-based computation and aim to preserve the utility of the prediction model while still enforcing strong location privacy. We propose a location privacy-preserving traffic flow prediction framework that moves all location-sensitive operations into an encrypted recruitment protocol. A Paillier additively homomorphic encryption scheme supports geometric range queries, in particular, point in rectangle tests for task regions, directly over encrypted coordinates. Service requesters encode task areas as encrypted rectangles, vehicles encrypt their current positions, and edge nodes assist in homomorphic operations. The crowdsensing server can decide whether a vehicle lies inside a task area or satisfies distance constraints without observing raw locations and without modifying the traffic measurements used for learning. On top of this privacy-preserving data acquisition pipeline, we build a Gated Recurrent Unit (GRU) based traffic flow prediction model and evaluate it on real-world data. Because the privacy layer leaves traffic flow values intact, the GRU operates on high-fidelity time series under strict location privacy constraints. Experiments on the PeMS dataset and the NYC Green Taxi Trip Record dataset include recurrent, feed-forward, and representative mainstream spatiotemporal baselines. Models that explicitly capture road-network dependencies achieve lower forecasting errors, while under the evaluated settings, the ablation results show only marginal changes in prediction metrics, because the privacy mechanism is confined to the recruitment stage and does not perturb downstream traffic-flow values. Within this framework, GRU provides a competitive trade-off between predictive accuracy, computational efficiency, and implementation complexity for short-horizon traffic flow forecasting.
Background Accurate brain tumor segmentation from multi-modal Magnetic Resonance Imaging (MRI) is critical for clinical prognosis, yet existing 3D architectures face a persistent dilemma. Traditional Convolutional Neural Network (CNN) suffer from restricted receptive fields, while emerging Vision Transformers incur prohibitive quadratic computational complexity and excessive parameterization, severely hindering their deployment on resource-constrained clinical devices. Methods To address this fundamental trade-off between global contextual modeling and computational efficiency, this study proposes DCLA-UNet, a highly parameter-efficient 3D segmentation framework driven by two core architectural innovations. First, we design a lightweight encoder-decoder structure integrating a Slim Large-Kernel Module (SLKM) that utilizes depthwise axial convolutions to enlarge the effective receptive field, coupled with a Multi-Scale Fusion Module (MSFM) employing parallel dilated branches for robust semantic reconstruction. Second, we introduce a Dynamic Cross-Layer Compressive Attention (DCLA) mechanism that leverages extreme single-channel compression and multi-kernel spatial alignment to adaptively modulate bottleneck features, effectively bridging the semantic gap and suppressing background noise. Results Extensive evaluations across the BraTS (2019, 2020, 2021) and MSD BrainTumour bench- marks systematically validate the efficacy and generalization capability of DCLA-UNet. The proposed architecture requires only 0.53M trainable parameters—a 94.76% reduction compared to the 3D U-Net baseline—and consumes just 39.23 Giga Floating Point Operations (GFLOPs). Despite this extreme compression, DCLA-UNet achieves a competitive mean Dice Similarity Coefficient (DSC) of 87.5 ± 0.6% on BraTS2021, outperforming emerging lightweight architectures such as SegFormer3D and Mamba3D. While it matches the DSC of heavier models like MogaNet using only 9.2% of its parameters, a trade-off is observed in the 95th percentile Hausdorff Distance (HD_95) (7.6 ± 1.8), objectively reflecting the inherent limits of boundary precision under strict parameter constraints. Conclusion DCLA-UNet successfully breaks the conventional bottleneck of 3D global context modeling by achieving a competitive balance between parameter efficiency and segmentation accuracy. It provides a robust, parameter-efficient solution for 3D medical image segmentation in storage-constrained clinical scenarios. (The source code is available at https://github.com/Helium-327/DCLA-UNet-3D).
