Lung cancer accounts for the highest number of cancer deaths globally,underscoring the urgent need for early and precise detection to enhance patient outcomes.While deep learning has made remarkable strides in analyzi...Lung cancer accounts for the highest number of cancer deaths globally,underscoring the urgent need for early and precise detection to enhance patient outcomes.While deep learning has made remarkable strides in analyzing medical images,current approaches face a fundamental challenge.They cannot adequately capture detailed local patterns and broader contextual relationships within lung Computed tomography(CT)scans.To address this limitation,we introduce AMVT-NMN(adaptive multi-scale vision transformer with neuromorphic memory networks),which combines three complementary mechanisms.The dynamic adaptive kernel networks component intelligently adjusts receptive field sizes based on input characteristics,enabling flexible feature capture across multiple scales.The neuromorphic contextual memory attention module draws inspiration from how human memory systems process information,maintaining a dynamic record of diagnostically relevant patterns to inform current predictions.The hierarchical cross-scale fusion mechanism with learnable weights synthesizes information from different resolution levels through adaptive weighting.Testing on the Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases(IQOTHNCCD)dataset demonstrates strong performance:97.9%accuracy,96.5%sensitivity,98.7%specificity,and 99.2%Area under the Curve-Receiver Operating Characteristic(AUC-ROC).These results surpass existing methods such as CNN-GD,which achieved 97.2%accuracy.Notably,the high specificity translates to fewer false alarms,potentially reducing unnecessary biopsies and follow-up imaging outcomes that matter considerably in clinical practice.Result of AMVT-NMN generalization to the Lung Image Database Consortium and Image Database Resource Initiative(LIDC-IDRI),Lung nodule analysis(LUNA16),and Non-Small Cell Lung Cancer(NSCLC)-Radiomics datasets showed AUCs of 96.5%,92.8%,and 97.2%,respectively.Ablation experiments confirm that each architectural element of AMVT-NMN contributes meaningfully to overall performance.Five-fold cross-validation yielded consistent results(97.71±0.57%),indicating reliable performance across different patient subsets.The memory-augmented design shows particular promise for handling diagnostically ambiguous cases.It is focused on pattern recognition and computational intelligence,which is useful for coping with uncertain information in intelligent diagnosis systems,meeting the growing trend for trusted artificial intelligence(AI)in decision-making.展开更多
为解决现有瓦斯体积分数预测方法存在多尺度特征提取不足、高阶非线性建模能力欠缺的问题,提出1种基于融合多尺度卷积、注意力门控与多阶核优化的瓦斯体积分数预测模型(multi-scale attention-enhanced KAN network,MSAK-Net)。该模型...为解决现有瓦斯体积分数预测方法存在多尺度特征提取不足、高阶非线性建模能力欠缺的问题,提出1种基于融合多尺度卷积、注意力门控与多阶核优化的瓦斯体积分数预测模型(multi-scale attention-enhanced KAN network,MSAK-Net)。该模型首先通过构建交互式卷积模块,以并行卷积与动态门控机制融合瓦斯序列的趋势与波动特征,然后采用注意力门控循环单元强化关键历史信息以提升模型的长程依赖建模能力,最后利用多阶核自适应网络,结合多阶核映射与多分支卷积实现多尺度整合。研究结果表明:与线性模型、Transformer模型及循环神经网络等对比模型相比,MSAK-Net在MAE(平均绝对误差)和RMSE(均方根误差)上分别降低约1.95%~20.24%、0.87%~15.49%,R2(决定系数)提升约0.13%~2.98%;MSAK-Net的PRE(峰值相对误差)降低了约5.38%~37.81%,RoCE(变化率误差)降低了约1.71%~4.17%,FAR(误报率)降低了约29.94%~66.77%。研究结果可为煤矿瓦斯体积分数预测与智能预警提供方法参考。展开更多
针对动态室内环境的变化及时变的接收信号强度(Received signal strength,RSS)对定位精度的影响,提出了一类基于核自适应滤波算法的农业无线传感器网络室内定位方法。核自适应滤波算法具体包括量化核最小均方(Quantized kernel least me...针对动态室内环境的变化及时变的接收信号强度(Received signal strength,RSS)对定位精度的影响,提出了一类基于核自适应滤波算法的农业无线传感器网络室内定位方法。核自适应滤波算法具体包括量化核最小均方(Quantized kernel least mean square,QKLMS)算法及固定预算(Fixed-budget,FB)核递推最小二乘(Kernel recursive least-squares,KRLS)算法。QKLMS算法基于一种简单在线矢量量化方法替代稀疏化,抑制核自适应滤波中径向基函数结构的增长。FB-KRLS算法是一种固定内存预算的在线学习方法,与以往的"滑窗"技术不同,每次时间更新时并不"修剪"最旧的数据,而是旨在"修剪"最无用的数据,从而抑制核矩阵的不断增长。通过构建RSS指纹信息与物理位置之间的非线性映射关系,核自适应滤波算法实现WSN的室内定位,将所提出的算法应用于仿真与物理环境下的不同实例中,在同等条件下,还与其他核学习算法、极限学习机(Extreme learning machine,ELM)等定位算法进行比较。仿真实验中2种算法在3种情形下的平均定位误差分别为0.746、0.443 m,物理实验中2种算法在2种情形下的平均定位误差分别为0.547、0.282 m。实验结果表明,所提出的核自适应滤波算法均能提高定位精度,其在线学习能力使得所提出的定位算法能自适应环境动态的变化。展开更多
基金supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number(PNURSP2026R757)Princess Nourah bint Abdulrahman University,Riyadh,Saudi Arabia and the National Natural Science Foundation of China under Grant numbers 62071153,Grant 62571163 and Grant 62571167。
摘要Lung cancer accounts for the highest number of cancer deaths globally,underscoring the urgent need for early and precise detection to enhance patient outcomes.While deep learning has made remarkable strides in analyzing medical images,current approaches face a fundamental challenge.They cannot adequately capture detailed local patterns and broader contextual relationships within lung Computed tomography(CT)scans.To address this limitation,we introduce AMVT-NMN(adaptive multi-scale vision transformer with neuromorphic memory networks),which combines three complementary mechanisms.The dynamic adaptive kernel networks component intelligently adjusts receptive field sizes based on input characteristics,enabling flexible feature capture across multiple scales.The neuromorphic contextual memory attention module draws inspiration from how human memory systems process information,maintaining a dynamic record of diagnostically relevant patterns to inform current predictions.The hierarchical cross-scale fusion mechanism with learnable weights synthesizes information from different resolution levels through adaptive weighting.Testing on the Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases(IQOTHNCCD)dataset demonstrates strong performance:97.9%accuracy,96.5%sensitivity,98.7%specificity,and 99.2%Area under the Curve-Receiver Operating Characteristic(AUC-ROC).These results surpass existing methods such as CNN-GD,which achieved 97.2%accuracy.Notably,the high specificity translates to fewer false alarms,potentially reducing unnecessary biopsies and follow-up imaging outcomes that matter considerably in clinical practice.Result of AMVT-NMN generalization to the Lung Image Database Consortium and Image Database Resource Initiative(LIDC-IDRI),Lung nodule analysis(LUNA16),and Non-Small Cell Lung Cancer(NSCLC)-Radiomics datasets showed AUCs of 96.5%,92.8%,and 97.2%,respectively.Ablation experiments confirm that each architectural element of AMVT-NMN contributes meaningfully to overall performance.Five-fold cross-validation yielded consistent results(97.71±0.57%),indicating reliable performance across different patient subsets.The memory-augmented design shows particular promise for handling diagnostically ambiguous cases.It is focused on pattern recognition and computational intelligence,which is useful for coping with uncertain information in intelligent diagnosis systems,meeting the growing trend for trusted artificial intelligence(AI)in decision-making.
摘要为解决现有瓦斯体积分数预测方法存在多尺度特征提取不足、高阶非线性建模能力欠缺的问题,提出1种基于融合多尺度卷积、注意力门控与多阶核优化的瓦斯体积分数预测模型(multi-scale attention-enhanced KAN network,MSAK-Net)。该模型首先通过构建交互式卷积模块,以并行卷积与动态门控机制融合瓦斯序列的趋势与波动特征,然后采用注意力门控循环单元强化关键历史信息以提升模型的长程依赖建模能力,最后利用多阶核自适应网络,结合多阶核映射与多分支卷积实现多尺度整合。研究结果表明:与线性模型、Transformer模型及循环神经网络等对比模型相比,MSAK-Net在MAE(平均绝对误差)和RMSE(均方根误差)上分别降低约1.95%~20.24%、0.87%~15.49%,R2(决定系数)提升约0.13%~2.98%;MSAK-Net的PRE(峰值相对误差)降低了约5.38%~37.81%,RoCE(变化率误差)降低了约1.71%~4.17%,FAR(误报率)降低了约29.94%~66.77%。研究结果可为煤矿瓦斯体积分数预测与智能预警提供方法参考。
摘要针对动态室内环境的变化及时变的接收信号强度(Received signal strength,RSS)对定位精度的影响,提出了一类基于核自适应滤波算法的农业无线传感器网络室内定位方法。核自适应滤波算法具体包括量化核最小均方(Quantized kernel least mean square,QKLMS)算法及固定预算(Fixed-budget,FB)核递推最小二乘(Kernel recursive least-squares,KRLS)算法。QKLMS算法基于一种简单在线矢量量化方法替代稀疏化,抑制核自适应滤波中径向基函数结构的增长。FB-KRLS算法是一种固定内存预算的在线学习方法,与以往的"滑窗"技术不同,每次时间更新时并不"修剪"最旧的数据,而是旨在"修剪"最无用的数据,从而抑制核矩阵的不断增长。通过构建RSS指纹信息与物理位置之间的非线性映射关系,核自适应滤波算法实现WSN的室内定位,将所提出的算法应用于仿真与物理环境下的不同实例中,在同等条件下,还与其他核学习算法、极限学习机(Extreme learning machine,ELM)等定位算法进行比较。仿真实验中2种算法在3种情形下的平均定位误差分别为0.746、0.443 m,物理实验中2种算法在2种情形下的平均定位误差分别为0.547、0.282 m。实验结果表明,所提出的核自适应滤波算法均能提高定位精度,其在线学习能力使得所提出的定位算法能自适应环境动态的变化。