首页期刊导航|The Journal of Supercomputing
期刊信息/Journal information
The Journal of Supercomputing
Kluwer Academic Publishers
The Journal of Supercomputing

Kluwer Academic Publishers

双月刊

0920-8542

The Journal of Supercomputing/Journal The Journal of SupercomputingSCI
正式出版
收录年代

    Learnable confdence‑driven asymmetric attention fusion mechanism for PET/CT tumor segmentation

    Zhaoshuo DiaoYan ZhangYang XuXi Chen...
    803.1-803.21页
    查看更多>>摘要：Positron emission tomography/computed tomography (PET/CT) imaging, combin- ing PET’s sensitivity with CT’s resolution, is pivotal in clinical tumor screening. PET/CT tumor segmentation forms the basis for treatment planning and surgical guidance. Recent advancements in deep learning-based PET/CT tumor segmenta- tion, notably via feature fusion strategies, have shown promise. However, existing fusion strategies have not accounted for the phenomenon where certain tumors are more prominent in PET images while others are more prominent in CT images, thus limiting their ability to fully exploit features from both modalities. To address this, we propose an learnable confdence-driven asymmetric attention fusion (LCAAF) mechanism. Initially, we train PET and CT segmentation branches separately and propose learnable segmentation confdence to evaluate single-modality segmenta- tion without labels. Leveraging this confdence, we design an asymmetric attention mechanism that integrates encoder features from both modalities into the decoder. The proposed mechanism enhances the contribution of the modality with higher confdence in feature fusion, thereby improving segmentation outcomes. Experi- mental results on soft tissue sarcoma (STS) and headNeck datasets demonstrate the efcacy of assigning higher weights to modalities with prominent tumor regions during feature fusion. The proposed method yields a 1.87% higher Dice value com- pared to the highest-performing CSAE-Net model on the STS dataset and a 4.88% higher Dice value compared to the highest CSAE-Net model on the HeadNeck data- set. Additionally, the proposed segmentation confdence can serve as an evaluation metric for label-free segmentation results.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    AE‑MCDM: an autoencoder‑based multi‑criteria decision‑making approach for unsupervised feature selection

    Amin HashemiMohammad Bagher DowlatshahiSiamak FarshidiParham Moradi...
    804.1-804.44页
    查看更多>>摘要：Feature selection is a fundamental technique for reducing the dimensionality of high-dimensional data by identifying the most relevant features while discarding redundant or irrelevant ones. In unsupervised settings, where labeled data are unavailable and labeling is costly, efective feature selection becomes even more challenging. This paper proposes AE-MCDM, a novel unsupervised feature selection method that integrates autoencoder-based feature extraction with multi- criteria decision-making (MCDM). The autoencoder captures high-level feature representations, and the connection weights between input features and hidden neurons refect feature importance. These weights are then processed using MCDM to rank and select the most informative features. Unlike conventional unsupervised feature selection methods, AE-MCDM leverages deep representation learning to enhance feature evaluation. To the best of our knowledge, this is the frst attempt to combine autoencoders with MCDM for feature selection. Extensive experiments on various datasets demonstrate that AE-MCDM outperforms existing methods in terms of clustering performance, measured by metrics such as accuracy, precision, recall, and normalized mutual information (NMI), while also achieving competitive computational efciency.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    CCLNet: multiclass motor imagery EEG decoding through extended common spatial patterns and CNN‑LSTM hybrid network

    Kamal SinghNitin SinghaSwati Bhalaik
    805.1-805.28页
    查看更多>>摘要：Decoding multiclass motor imagery electroencephalography (MI-EEG) data is crucial for brain–computer interface (BCI) applications. This study proposes CCLNet, a novel and high-performing approach for classifying MI-EEG data. CCLNet achieves this performance through two key innovations. First, it introduces a novel common spatial pattern (CSP)-based feature extraction method specifcally designed for multiclass MI-EEG data. This method extracts informative features by identifying spatial patterns that maximize the variance between diferent MI-EEG classes. Second, CCLNet employs a deep learning model with a convolutional neural network and long short-term memory (CNN-LSTM) architecture. The CNN component is employed to extract features to learn complex spatial patterns within the data, while the LSTM captures the temporal dependencies present in the MI-EEG data. By combining these innovations, CCLNet achieves superior classifcation accuracy compared to state-of-the-art methods. The efectiveness of CCLNet was evaluated using within-subject and cross-subject approaches on two separate datasets: BCI Competition IV-2a and HGD. On BCI Competition IV-2a, CCLNet achieved impressive accuracy, reaching 95.87% and 97.08% for within-subject and cross-subject scenarios, respectively. Furthermore, CCLNet demonstrated exceptional performance on the HGD dataset, achieving an accuracy of 98.56%. These outstanding results highlight CCLNet’s potential for real-world applications, particularly in advancing assistive technologies for individuals with motor disabilities.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    Fraud calls detection using class‑imbalanced learning on graph structures

    Nguyen Tan CamDinh Hoai Hiep
    806.1-806.40页
    查看更多>>摘要：The rapid growth of mobile networks has enhanced social interactions but has also increased fraud risks in mobile social networks, leading to fnancial and economic losses. Fraud detection systems based on Graph Neural Networks (GNNs) utilize Call Detail Record (CDR) data to analyze social behaviors, yet they struggle with data imbalance, which limits their efectiveness. In this study, we address this chal- lenge by developing an improved minority class data augmentation approach for graph-based fraud detection. Building upon existing generative models, we enhance data generation using Wasserstein GAN with Gradient Penalty (WGAN-GP) to mitigate mode collapse and Deep Denoising Difusion Models (DDPM) to generate high-quality synthetic data. These synthetic samples are then integrated with graph- based classifers to improve fraud detection performance. Experimental results dem- onstrate that our approach signifcantly improves classifcation performance, par- ticularly in terms of F1-score, recall, and generalization across multiple graph-based fraud detection models. This research contributes to advancing data augmentation techniques for imbalanced graph data, ultimately enhancing fraud detection efec- tiveness and network security in mobile telecommunications.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    Two peak‑fnding algorithms for two‑dimensional unimodal symmetric signals based on mirroring and interpolating

    Wei ChenJiao LiBin WanJinwei Hu...
    807.1-807.20页
    查看更多>>摘要：In the processing of two-dimensional single-peak symmetric signals, the peak posi- tion is a very important parameter. The current peak position estimation methods generally can be classifed into two types: one is direct methods that get the value of peak position, which always sufer the shortcomings of low precision and poor noise resistance, especially for the estimation of sparse signals, whose quality is far from satisfactory; the other type requires extra prior information about the signals and then applies the corresponding ftting methods, but these methods often lose gener- ality. In this paper, two new peak-fnding algorithms are proposed. They mainly take advantage of the symmetry of the signal to obtain a denser signal through the mir- roring and interpolating operations, which enhance the noise resistance and the esti- mation precision for the algorithm. From the experiments conducted in this paper, these two algorithms show obvious advantages compared with the centroid method, whose abilities of noisy resistance and performances in processing sparse signals are increased by approximately ffty to one hundred percent.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    A hierarchical blockchain architecture for federated learning in edge computing networks

    Shuyang RenChoonhwa Lee
    808.1-808.27页
    查看更多>>摘要：Blockchain-based federated learning (FL) has recently garnered signifcant atten- tion as a trusted decentralized learning paradigm. However, traditional FL faces critical challenges: synchronous FL sufers from stragglers that delay training, while asynchronous FL risks model instability due to inconsistent updates. Moreover, processing blockchain consensus protocols incurs substantial resource consump- tion and operational latency. To overcome these challenges, we propose a hierar- chical blockchain architecture for semi-asynchronous FL that balances efciency and security. Our approach features a two-layer design: (1) a training layer, where edge nodes asynchronously upload local models via a directed acyclic graph (DAG) to mitigate stragglers and ensure continuous progress, and (2) a blockchain layer, which periodically validates and synchronously aggregates models to maintain sta- bility and defend against malicious inputs. We further introduce novel DAG-based transaction tracking and uploading algorithms to enhance efciency, enabling rapid local updates while ensuring global model integrity through blockchain consensus. Experimental results demonstrate that our system reduces latency by 26% com- pared to typical blockchain-based FL approaches, while maintaining a stable con- vergence rate and high training accuracy. By harmonizing asynchronous fexibility with synchronous control, our work enhances the scalability and robustness of FL in resource-constrained edge environments.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    Dual‑YOLO: dual‑path UAV aerial image target detection algorithm

    Houhong LiuFei TanYuxin Jin
    809.1-809.14页
    查看更多>>摘要：In UAV aerial image target detection, the presence of small-scale objects, complex backgrounds, and weak illumination leads to difculties in feature extraction and low detection accuracy. To address these issues, this paper proposes an aerial image target detection algorithm named Dual-YOLO. First, a parallel dual-path backbone network is designed to achieve complementary feature extraction, thereby enhancing the feature extraction capability for targets. Second, a bidirectional feature pyramid network (BiFPN) structure is implemented in the neck to optimize multi- scale feature fusion, which enhances feature representation capabilities through its bidirectional cross-scale connectivity. Finally, a dynamic head framework is employed to unify the object detection head and the attention mechanism, thereby enhancing overall detection performance. Experimental results show that the Dual- YOLO algorithm achieves mean Average Precision at 50% IoU (mAP50) scores of 43.1% and 76.3% on the VisDrone2019 and HazyDet datasets, respectively, outperforming the baseline model by 9.3% and 6.4% and signifcantly enhancing detection accuracy for aerial targets.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    Improved verifcation limit for the convergence of the Collatz conjecture

    David Barina
    810.1-810.14页
    查看更多>>摘要：This article presents our project, which aims to verify the Collatz conjecture com- putationally. As a main point of the article, we introduce a new result that pushes the limit for which the conjecture is verifed up to 271. We present our baseline algo- rithm and then several sub-algorithms that enhance acceleration. The total accelera- tion from the frst algorithm we used on the CPU to our best algorithm on the GPU is 1 335×. We further distribute individual tasks to thousands of parallel workers running on several European supercomputers. Besides the convergence verifcation, our program also checks for path records during the convergence test. We found four new path records.
      原文链接:
    • NETL
    • NSTL

    Correction: YOLO‑Fast: a lightweight object detection model for edge devices

    Zijing SongXiaoyu ZhangPanlong Tan
    811.1-811.1页
    查看更多>>摘要：In this article the corresponding author’s name Xiaoyu Zhang was incorrectly written as Xiaoyv Zhang. The original article has been corrected.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature

    An algorithm for indoor positioning based on LightGBM+ExtraTrees chaotic weighted ensemble: evaluation and comparison

    Lieping ZhangYao ChenXiaolin GaoXinpeng Zheng...
    812.1-812.31页
    查看更多>>摘要：To address the issues of insufcient positioning accuracy and poor stability in exist- ing indoor positioning algorithms, this paper proposes an indoor fngerprint posi- tioning algorithm based on LightGBM and ExtraTrees chaotic weighted ensemble. First, the received signal strength indicator (RSSI) signal strengths of wireless fdel- ity (WiFi) at diferent locations are collected using a mobile device to construct a fngerprint database. Kalman fltering is then applied to preprocess the fngerprint data, removing outliers and noise to improve the data quality. The preprocessed dataset is subsequently divided into training and testing sets. Light gradient boost- ing machine (LightGBM) and extremely randomized trees (ExtraTrees) are used for modeling, and the chaos particle swarm optimization (CPSO) algorithm is employed to optimize the key parameters of both LightGBM and ExtraTrees. The optimal parameter confguration is determined based on comprehensive evaluation metrics. Finally, the optimal weight ratio of LightGBM and ExtraTrees models is determined through the CPSO algorithm for weighted fusion. Experimental results demonstrate that the proposed algorithm achieves an average positioning error of 1.1 m, which represents a reduction of 7.5–26.7% in average positioning error compared to Light- GBM, ExtraTrees, and the LightGBM+ExtraTrees algorithms. After introducing random noise, the proposed algorithm exhibits the smallest variation in average positioning error, validating its signifcant advantages in positioning accuracy and anti-interference ability. However, the algorithm in this study is primarily based on static WiFi fngerprint data and has not yet been validated in dynamic environments. Future research should further explore its applicability and robustness in more com- plex environments.
      原文链接:
    • NETL
    • NSTL
    • Springer Nature