Various deep learning models have been proposed for the accurate assisted diagnosis of early-stage Alzheimer’s disease(AD).Most studies predominantly employ Convolutional Neural Networks(CNNs),which focus solely on l...Various deep learning models have been proposed for the accurate assisted diagnosis of early-stage Alzheimer’s disease(AD).Most studies predominantly employ Convolutional Neural Networks(CNNs),which focus solely on local features,thus encountering difficulties in handling global features.In contrast to natural images,Structural Magnetic Resonance Imaging(sMRI)images exhibit a higher number of channel dimensions.However,during the Position Embedding stage ofMulti Head Self Attention(MHSA),the coded information related to the channel dimension is disregarded.To tackle these issues,we propose theRepBoTNet-CESA network,an advanced AD-aided diagnostic model that is capable of learning local and global features simultaneously.It combines the advantages of CNN networks in capturing local information and Transformer networks in integrating global information,reducing computational costs while achieving excellent classification performance.Moreover,it uses the Cubic Embedding Self Attention(CESA)proposed in this paper to incorporate the channel code information,enhancing the classification performance within the Transformer structure.Finally,the RepBoTNet-CESA performs well in various AD-aided diagnosis tasks,with an accuracy of 96.58%,precision of 97.26%,and recall of 96.23%in the AD/NC task;an accuracy of 92.75%,precision of 92.84%,and recall of 93.18%in the EMCI/NC task;and an accuracy of 80.97%,precision of 83.86%,and recall of 80.91%in the AD/EMCI/LMCI/NC task.This demonstrates that RepBoTNet-CESA delivers outstanding outcomes in various AD-aided diagnostic tasks.Furthermore,our study has shown that MHSA exhibits superior performance compared to conventional attention mechanisms in enhancing ResNet performance.Besides,the Deeper RepBoTNet-CESA network fails to make further progress in AD-aided diagnostic tasks.展开更多
This paper proposes a novel method for the automatic diagnosis of keratitis using feature vector quantization and self-attention mechanisms(ADK_FVQSAM).First,high-level features are extracted using the DenseNet121 bac...This paper proposes a novel method for the automatic diagnosis of keratitis using feature vector quantization and self-attention mechanisms(ADK_FVQSAM).First,high-level features are extracted using the DenseNet121 backbone network,followed by adaptive average pooling to scale the features to a fixed length.Subsequently,product quantization with residuals(PQR)is applied to convert continuous feature vectors into discrete features representations,preserving essential information insensitive to image quality variations.The quantized and original features are concatenated and fed into a self-attention mechanism to capture keratitis-related features.Finally,these enhanced features are classified through a fully connected layer.Experiments on clinical low-quality(LQ)images show that ADK_FVQSAM achieves accuracies of 87.7%,81.9%,and 89.3% for keratitis,other corneal abnormalities,and normal corneas,respectively.Compared to DenseNet121,Swin transformer,and InceptionResNet,ADK_FVQSAM improves average accuracy by 3.1%,11.3%,and 15.3%,respectively.These results demonstrate that ADK_FVQSAM significantly enhances the recognition performance of keratitis based on LQ slit-lamp images,offering a practical approach for clinical application.展开更多
Convolutional neural network(CNN)with the encoder-decoder structure is popular in medical image segmentation due to its excellent local feature extraction ability but it faces limitations in capturing the global featu...Convolutional neural network(CNN)with the encoder-decoder structure is popular in medical image segmentation due to its excellent local feature extraction ability but it faces limitations in capturing the global feature.The transformer can extract the global information well but adapting it to small medical datasets is challenging and its computational complexity can be heavy.In this work,a serial and parallel network is proposed for the accurate 3D medical image segmentation by combining CNN and transformer and promoting feature interactions across various semantic levels.The core components of the proposed method include the cross window self-attention based transformer(CWST)and multi-scale local enhanced(MLE)modules.The CWST module enhances the global context understanding by partitioning 3D images into non-overlapping windows and calculating sparse global attention between windows.The MLE module selectively fuses features by computing the voxel attention between different branch features,and uses convolution to strengthen the dense local information.The experiments on the prostate,atrium,and pancreas MR/CT image datasets consistently demonstrate the advantage of the proposed method over six popular segmentation models in both qualitative evaluation and quantitative indexes such as dice similarity coefficient,Intersection over Union,95%Hausdorff distance and average symmetric surface distance.展开更多
Tumour segmentation in medical images(especially 3D tumour segmentation)is highly challenging due to the possible similarity between tumours and adjacent tissues,occurrence of multiple tumours and variable tumour shap...Tumour segmentation in medical images(especially 3D tumour segmentation)is highly challenging due to the possible similarity between tumours and adjacent tissues,occurrence of multiple tumours and variable tumour shapes and sizes.The popular deep learning‐based segmentation algorithms generally rely on the convolutional neural network(CNN)and Transformer.The former cannot extract the global image features effectively while the latter lacks the inductive bias and involves the complicated computation for 3D volume data.The existing hybrid CNN‐Transformer network can only provide the limited performance improvement or even poorer segmentation performance than the pure CNN.To address these issues,a short‐term and long‐term memory self‐attention network is proposed.Firstly,a distinctive self‐attention block uses the Transformer to explore the correlation among the region features at different levels extracted by the CNN.Then,the memory structure filters and combines the above information to exclude the similar regions and detect the multiple tumours.Finally,the multi‐layer reconstruction blocks will predict the tumour boundaries.Experimental results demonstrate that our method outperforms other methods in terms of subjective visual and quantitative evaluation.Compared with the most competitive method,the proposed method provides Dice(82.4%vs.76.6%)and Hausdorff distance 95%(HD95)(10.66 vs.11.54 mm)on the KiTS19 as well as Dice(80.2%vs.78.4%)and HD95(9.632 vs.12.17 mm)on the LiTS.展开更多
With the continuous growth of online news articles,there arises the necessity for an efficient abstractive summarization technique for the problem of information overloading.Abstractive summarization is highly complex...With the continuous growth of online news articles,there arises the necessity for an efficient abstractive summarization technique for the problem of information overloading.Abstractive summarization is highly complex and requires a deeper understanding and proper reasoning to come up with its own summary outline.Abstractive summarization task is framed as seq2seq modeling.Existing seq2seq methods perform better on short sequences;however,for long sequences,the performance degrades due to high computation and hence a two-phase self-normalized deep neural document summarization model consisting of improvised extractive cosine normalization and seq2seq abstractive phases has been proposed in this paper.The novelty is to parallelize the sequence computation training by incorporating feed-forward,the self-normalized neural network in the Extractive phase using Intra Cosine Attention Similarity(Ext-ICAS)with sentence dependency position.Also,it does not require any normalization technique explicitly.Our proposed abstractive Bidirectional Long Short Term Memory(Bi-LSTM)encoder sequence model performs better than the Bidirectional Gated Recurrent Unit(Bi-GRU)encoder with minimum training loss and with fast convergence.The proposed model was evaluated on the Cable News Network(CNN)/Daily Mail dataset and an average rouge score of 0.435 was achieved also computational training in the extractive phase was reduced by 59%with an average number of similarity computations.展开更多
An improved cycle-consistent generative adversarial network(CycleGAN) method for defect data augmentation based on feature fusion and self attention residual module is proposed to address the insufficiency of defect s...An improved cycle-consistent generative adversarial network(CycleGAN) method for defect data augmentation based on feature fusion and self attention residual module is proposed to address the insufficiency of defect sample data for light guide plate(LGP) in production,as well as the problem of minor defects.Two optimizations are made to the generator of CycleGAN:fusion of low resolution features obtained from partial up-sampling and down-sampling with high-resolution features,combination of self attention mechanism with residual network structure to replace the original residual module.Qualitative and quantitative experiments were conducted to compare different data augmentation methods,and the results show that the defect images of the LGP generated by the improved network were more realistic,and the accuracy of the you only look once version 5(YOLOv5) detection network for the LGP was improved by 5.6%,proving the effectiveness and accuracy of the proposed method.展开更多
Devices in Industrial Internet of Things are vulnerable to voice adversarial attacks.Studying adversarial speech samples is crucial for enhancing the security of automatic speech recognition systems in Industrial Inte...Devices in Industrial Internet of Things are vulnerable to voice adversarial attacks.Studying adversarial speech samples is crucial for enhancing the security of automatic speech recognition systems in Industrial Internet of Things devices.Current black-box attack methods often face challenges such as complex search processes and excessive perturbation generation.To address these issues,this paper proposes a black-box voice adversarial attack method based on enhanced neural predictors.This method searches for minimal perturbations in the perturbation space,employing an optimization process guided by a self-attention neural predictor to identify the optimal perturbation direction.This direction is then applied to the original sample to generate adversarial samples.To improve search efficiency,a pruning strategy is designed to discard samples below a threshold in the early search stages,reducing the number of searches.Additionally,a dynamic factor based on feedback from querying the automatic speech recognition system is introduced to adaptively adjust the search step size,further accelerating the search process.To validate the performance of the proposed method,experiments are conducted on the LibriSpeech dataset.Compared with the mainstream methods,the proposed method improves the signal-to-noise ratio by 0.8 dB,increases sample similarity by 0.43%,and reduces the average number of queries by 7%.Experimental results demonstrate that the proposed method offers better attack effectiveness and stealthiness.展开更多
Spike transformers cannot be pretrained due to objective factors such as lack of datasets and memory constraints,which results in a significant performance gap compared to pretrained artificial neural networks(ANNs),t...Spike transformers cannot be pretrained due to objective factors such as lack of datasets and memory constraints,which results in a significant performance gap compared to pretrained artificial neural networks(ANNs),thereby hindering their practical applicability.To address this issue,we propose a hybrid attention spike transformer that utilises self-attention with compound tokens and channel attention-based token processing to better capture the inductive biases of the data.We also add convolution in patch splitting and feedforward networks,which not only provides local information but also leverages the translation invariance and locality of convolutions to help the model converge.Experiments on static datasets and neuromorphic datasets demonstrate that our method achieves state-of-the-art performance in the spiking neural networks(SNNs)field.Notably,we achieve a top-1 accuracy of 80.59%on CIFAR-100 with only 4 time steps.As far as we know,it is the first exploration of the spike transformer with multiattention fusion,achieving outstanding effectiveness.展开更多
基金the Key Project of Zhejiang Provincial Natural Science Foundation under Grants LD21F020001,Z20F020022the National Natural Science Foundation of China under Grants 62072340,62076185the Major Project of Wenzhou Natural Science Foundation under Grants 2021HZSY0071,ZS2022001.
文摘Various deep learning models have been proposed for the accurate assisted diagnosis of early-stage Alzheimer’s disease(AD).Most studies predominantly employ Convolutional Neural Networks(CNNs),which focus solely on local features,thus encountering difficulties in handling global features.In contrast to natural images,Structural Magnetic Resonance Imaging(sMRI)images exhibit a higher number of channel dimensions.However,during the Position Embedding stage ofMulti Head Self Attention(MHSA),the coded information related to the channel dimension is disregarded.To tackle these issues,we propose theRepBoTNet-CESA network,an advanced AD-aided diagnostic model that is capable of learning local and global features simultaneously.It combines the advantages of CNN networks in capturing local information and Transformer networks in integrating global information,reducing computational costs while achieving excellent classification performance.Moreover,it uses the Cubic Embedding Self Attention(CESA)proposed in this paper to incorporate the channel code information,enhancing the classification performance within the Transformer structure.Finally,the RepBoTNet-CESA performs well in various AD-aided diagnosis tasks,with an accuracy of 96.58%,precision of 97.26%,and recall of 96.23%in the AD/NC task;an accuracy of 92.75%,precision of 92.84%,and recall of 93.18%in the EMCI/NC task;and an accuracy of 80.97%,precision of 83.86%,and recall of 80.91%in the AD/EMCI/LMCI/NC task.This demonstrates that RepBoTNet-CESA delivers outstanding outcomes in various AD-aided diagnostic tasks.Furthermore,our study has shown that MHSA exhibits superior performance compared to conventional attention mechanisms in enhancing ResNet performance.Besides,the Deeper RepBoTNet-CESA network fails to make further progress in AD-aided diagnostic tasks.
基金supported by the National Natural Science Foundation of China(Nos.62276210,82201148 and 62376215)the Key Research and Development Project of Shaanxi Province(No.2025CY-YBXM-044)+3 种基金the Natural Science Foundation of Zhejiang Province(No.LQ22H120002)the Medical Health Science and Technology Project of Zhejiang Province(Nos.2022RC069 and 2023KY1140)the Natural Science Foundation of Ningbo(No.2023J390)the Ningbo Top Medical and Health Research Program(No.2023030716).
文摘This paper proposes a novel method for the automatic diagnosis of keratitis using feature vector quantization and self-attention mechanisms(ADK_FVQSAM).First,high-level features are extracted using the DenseNet121 backbone network,followed by adaptive average pooling to scale the features to a fixed length.Subsequently,product quantization with residuals(PQR)is applied to convert continuous feature vectors into discrete features representations,preserving essential information insensitive to image quality variations.The quantized and original features are concatenated and fed into a self-attention mechanism to capture keratitis-related features.Finally,these enhanced features are classified through a fully connected layer.Experiments on clinical low-quality(LQ)images show that ADK_FVQSAM achieves accuracies of 87.7%,81.9%,and 89.3% for keratitis,other corneal abnormalities,and normal corneas,respectively.Compared to DenseNet121,Swin transformer,and InceptionResNet,ADK_FVQSAM improves average accuracy by 3.1%,11.3%,and 15.3%,respectively.These results demonstrate that ADK_FVQSAM significantly enhances the recognition performance of keratitis based on LQ slit-lamp images,offering a practical approach for clinical application.
基金National Key Research and Development Program of China,Grant/Award Number:2018YFE0206900China Postdoctoral Science Foundation,Grant/Award Number:2023M731204+2 种基金The Open Project of Key Laboratory for Quality Evaluation of Ultrasound Surgical Equipment of National Medical Products Administration,Grant/Award Number:SMDTKL-2023-1-01The Hubei Province Key Research and Development Project,Grant/Award Number:2023BCB007CAAI-Huawei MindSpore Open Fund。
文摘Convolutional neural network(CNN)with the encoder-decoder structure is popular in medical image segmentation due to its excellent local feature extraction ability but it faces limitations in capturing the global feature.The transformer can extract the global information well but adapting it to small medical datasets is challenging and its computational complexity can be heavy.In this work,a serial and parallel network is proposed for the accurate 3D medical image segmentation by combining CNN and transformer and promoting feature interactions across various semantic levels.The core components of the proposed method include the cross window self-attention based transformer(CWST)and multi-scale local enhanced(MLE)modules.The CWST module enhances the global context understanding by partitioning 3D images into non-overlapping windows and calculating sparse global attention between windows.The MLE module selectively fuses features by computing the voxel attention between different branch features,and uses convolution to strengthen the dense local information.The experiments on the prostate,atrium,and pancreas MR/CT image datasets consistently demonstrate the advantage of the proposed method over six popular segmentation models in both qualitative evaluation and quantitative indexes such as dice similarity coefficient,Intersection over Union,95%Hausdorff distance and average symmetric surface distance.
基金supported by the National Key Research and Development Program of China under Grant No.2018YFE0206900the National Natural Science Foundation of China under Grant No.61871440 and CAAI‐Huawei Mind-Spore Open Fund.
文摘Tumour segmentation in medical images(especially 3D tumour segmentation)is highly challenging due to the possible similarity between tumours and adjacent tissues,occurrence of multiple tumours and variable tumour shapes and sizes.The popular deep learning‐based segmentation algorithms generally rely on the convolutional neural network(CNN)and Transformer.The former cannot extract the global image features effectively while the latter lacks the inductive bias and involves the complicated computation for 3D volume data.The existing hybrid CNN‐Transformer network can only provide the limited performance improvement or even poorer segmentation performance than the pure CNN.To address these issues,a short‐term and long‐term memory self‐attention network is proposed.Firstly,a distinctive self‐attention block uses the Transformer to explore the correlation among the region features at different levels extracted by the CNN.Then,the memory structure filters and combines the above information to exclude the similar regions and detect the multiple tumours.Finally,the multi‐layer reconstruction blocks will predict the tumour boundaries.Experimental results demonstrate that our method outperforms other methods in terms of subjective visual and quantitative evaluation.Compared with the most competitive method,the proposed method provides Dice(82.4%vs.76.6%)and Hausdorff distance 95%(HD95)(10.66 vs.11.54 mm)on the KiTS19 as well as Dice(80.2%vs.78.4%)and HD95(9.632 vs.12.17 mm)on the LiTS.
文摘With the continuous growth of online news articles,there arises the necessity for an efficient abstractive summarization technique for the problem of information overloading.Abstractive summarization is highly complex and requires a deeper understanding and proper reasoning to come up with its own summary outline.Abstractive summarization task is framed as seq2seq modeling.Existing seq2seq methods perform better on short sequences;however,for long sequences,the performance degrades due to high computation and hence a two-phase self-normalized deep neural document summarization model consisting of improvised extractive cosine normalization and seq2seq abstractive phases has been proposed in this paper.The novelty is to parallelize the sequence computation training by incorporating feed-forward,the self-normalized neural network in the Extractive phase using Intra Cosine Attention Similarity(Ext-ICAS)with sentence dependency position.Also,it does not require any normalization technique explicitly.Our proposed abstractive Bidirectional Long Short Term Memory(Bi-LSTM)encoder sequence model performs better than the Bidirectional Gated Recurrent Unit(Bi-GRU)encoder with minimum training loss and with fast convergence.The proposed model was evaluated on the Cable News Network(CNN)/Daily Mail dataset and an average rouge score of 0.435 was achieved also computational training in the extractive phase was reduced by 59%with an average number of similarity computations.
基金supported by the Jiangsu Province IUR Cooperation Project (No.BY2021258)the Wuxi Science and Technology Development Fund Project (No.G20212028)。
文摘An improved cycle-consistent generative adversarial network(CycleGAN) method for defect data augmentation based on feature fusion and self attention residual module is proposed to address the insufficiency of defect sample data for light guide plate(LGP) in production,as well as the problem of minor defects.Two optimizations are made to the generator of CycleGAN:fusion of low resolution features obtained from partial up-sampling and down-sampling with high-resolution features,combination of self attention mechanism with residual network structure to replace the original residual module.Qualitative and quantitative experiments were conducted to compare different data augmentation methods,and the results show that the defect images of the LGP generated by the improved network were more realistic,and the accuracy of the you only look once version 5(YOLOv5) detection network for the LGP was improved by 5.6%,proving the effectiveness and accuracy of the proposed method.
基金supported in part by the Natural Science Foundation of China under Grant 62273272,Grant 62303375,and Grant 61873277in part by the Key Research and Development Program of Shaanxi Province under Grant 2024CY2-GJHX-49 and Grant 2024CY2-GJHX-43+1 种基金in part by the Youth Innovation Team of Shaanxi Universitiesand in part by the Key Scientific Research Programof Education Department of Shaanxi Province under Grant 24JR111.
文摘Devices in Industrial Internet of Things are vulnerable to voice adversarial attacks.Studying adversarial speech samples is crucial for enhancing the security of automatic speech recognition systems in Industrial Internet of Things devices.Current black-box attack methods often face challenges such as complex search processes and excessive perturbation generation.To address these issues,this paper proposes a black-box voice adversarial attack method based on enhanced neural predictors.This method searches for minimal perturbations in the perturbation space,employing an optimization process guided by a self-attention neural predictor to identify the optimal perturbation direction.This direction is then applied to the original sample to generate adversarial samples.To improve search efficiency,a pruning strategy is designed to discard samples below a threshold in the early search stages,reducing the number of searches.Additionally,a dynamic factor based on feedback from querying the automatic speech recognition system is introduced to adaptively adjust the search step size,further accelerating the search process.To validate the performance of the proposed method,experiments are conducted on the LibriSpeech dataset.Compared with the mainstream methods,the proposed method improves the signal-to-noise ratio by 0.8 dB,increases sample similarity by 0.43%,and reduces the average number of queries by 7%.Experimental results demonstrate that the proposed method offers better attack effectiveness and stealthiness.
基金supported by STI 2030-Major Projects No.2022ZD0208800,in part by NSFC No.62088101 Autonomous Intelligent Unmanned Systemsin part by the Baima Lake Laboratory Joint Funds of the Zhejiang Provincial Natural Science Foundation of China under Grant No.LBMHD25F030001.
文摘Spike transformers cannot be pretrained due to objective factors such as lack of datasets and memory constraints,which results in a significant performance gap compared to pretrained artificial neural networks(ANNs),thereby hindering their practical applicability.To address this issue,we propose a hybrid attention spike transformer that utilises self-attention with compound tokens and channel attention-based token processing to better capture the inductive biases of the data.We also add convolution in patch splitting and feedforward networks,which not only provides local information but also leverages the translation invariance and locality of convolutions to help the model converge.Experiments on static datasets and neuromorphic datasets demonstrate that our method achieves state-of-the-art performance in the spiking neural networks(SNNs)field.Notably,we achieve a top-1 accuracy of 80.59%on CIFAR-100 with only 4 time steps.As far as we know,it is the first exploration of the spike transformer with multiattention fusion,achieving outstanding effectiveness.
