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
Accurate traffic flow prediction has a profound impact on modern traffic management. Traffic flow has complex spatial-temporal correlations and periodicity, which poses difficulties for precise prediction. To address ...Accurate traffic flow prediction has a profound impact on modern traffic management. Traffic flow has complex spatial-temporal correlations and periodicity, which poses difficulties for precise prediction. To address this problem, a Multi-head Self-attention and Spatial-Temporal Graph Convolutional Network (MSSTGCN) for multiscale traffic flow prediction is proposed. Firstly, to capture the hidden traffic periodicity of traffic flow, traffic flow is divided into three kinds of periods, including hourly, daily, and weekly data. Secondly, a graph attention residual layer is constructed to learn the global spatial features across regions. Local spatial-temporal dependence is captured by using a T-GCN module. Thirdly, a transformer layer is introduced to learn the long-term dependence in time. A position embedding mechanism is introduced to label position information for all traffic sequences. Thus, this multi-head self-attention mechanism can recognize the sequence order and allocate weights for different time nodes. Experimental results on four real-world datasets show that the MSSTGCN performs better than the baseline methods and can be successfully adapted to traffic prediction tasks.展开更多
Traditional data-driven fault diagnosis methods depend on expert experience to manually extract effective fault features of signals,which has certain limitations.Conversely,deep learning techniques have gained promine...Traditional data-driven fault diagnosis methods depend on expert experience to manually extract effective fault features of signals,which has certain limitations.Conversely,deep learning techniques have gained prominence as a central focus of research in the field of fault diagnosis by strong fault feature extraction ability and end-to-end fault diagnosis efficiency.Recently,utilizing the respective advantages of convolution neural network(CNN)and Transformer in local and global feature extraction,research on cooperating the two have demonstrated promise in the field of fault diagnosis.However,the cross-channel convolution mechanism in CNN and the self-attention calculations in Transformer contribute to excessive complexity in the cooperative model.This complexity results in high computational costs and limited industrial applicability.To tackle the above challenges,this paper proposes a lightweight CNN-Transformer named as SEFormer for rotating machinery fault diagnosis.First,a separable multiscale depthwise convolution block is designed to extract and integrate multiscale feature information from different channel dimensions of vibration signals.Then,an efficient self-attention block is developed to capture critical fine-grained features of the signal from a global perspective.Finally,experimental results on the planetary gearbox dataset and themotor roller bearing dataset prove that the proposed framework can balance the advantages of robustness,generalization and lightweight compared to recent state-of-the-art fault diagnosis models based on CNN and Transformer.This study presents a feasible strategy for developing a lightweight rotating machinery fault diagnosis framework aimed at economical deployment.展开更多
Dear Editor,D2This letter presents a node feature similarity preserving graph convolutional framework P G.Graph neural networks(GNNs)have garnered significant attention for their efficacy in learning graph representat...Dear Editor,D2This letter presents a node feature similarity preserving graph convolutional framework P G.Graph neural networks(GNNs)have garnered significant attention for their efficacy in learning graph representations across diverse real-world applications.展开更多
Many spore-forming Bacillus species can cause serious human diseases,because of accidental Bacillusspore infection.Thus,developing an identification strategy with both high sensitivity and specificity is greatly in de...Many spore-forming Bacillus species can cause serious human diseases,because of accidental Bacillusspore infection.Thus,developing an identification strategy with both high sensitivity and specificity is greatly in demand.In this work,we proposed a novel approach named multi-head self-attention mechanism-guided neural network Raman platform to identify living Bacillus spores within a single-cell resolution.The multi-head self-attention mechanism-guided neural network Raman platform was created by combining single-cell Raman spectroscopy,convolutional neural network(CNN),and multi-head self-attention mechanism.To address the limited size of the original spectra dataset,Gaussian noise-based spectra augmentation was employed to increase the number of single-cell Raman spectra datasets for CNN training.Owing to the assistance of both spectra augmentation and multi-head self-attention mechanism,the obtained prediction accuracy of five Bacillus spore species was further improved from 92.29±0.82%to 99.43±0.15%.To figure out the spectra differences covered by the multi-head self-attention mechanism-guided CNN,the relative classification weight from typical Raman bands was visualized via multi-head self-attention mechanism curve.In the process of spectra augmentation from 0 to 1000,the distribution of relative classification weight varied from a discrete state to a more concentrated phase.More importantly,these highlighted four Raman bands(1017,1449,1576,and 1660 cm^(-1))were assigned large weights,showing that the spectra differences in the Raman bands produced the largest contribution to prediction accuracy.It can be foreseen that,our proposed sorting platform has great potential in accurately identifying Bacillus and its related genera species at a single-cell level.展开更多
Dear Editor,The letter proposes a tensor low-rank orthogonal compression(TLOC)model for a convolutional neural network(CNN),which facilitates its efficient and highly-accurate low-rank representation.Model compression...Dear Editor,The letter proposes a tensor low-rank orthogonal compression(TLOC)model for a convolutional neural network(CNN),which facilitates its efficient and highly-accurate low-rank representation.Model compression is crucial for deploying deep neural network(DNN)models on resource-constrained embedded devices.展开更多
Sustainable energy systems will entail a change in the carbon intensity projections,which should be carried out in a proper manner to facilitate the smooth running of the grid and reduce greenhouse emissions.The prese...Sustainable energy systems will entail a change in the carbon intensity projections,which should be carried out in a proper manner to facilitate the smooth running of the grid and reduce greenhouse emissions.The present article outlines the TransCarbonNet,a novel hybrid deep learning framework with self-attention characteristics added to the bidirectional Long Short-Term Memory(Bi-LSTM)network to forecast the carbon intensity of the grid several days.The proposed temporal fusion model not only learns the local temporal interactions but also the long-term patterns of the carbon emission data;hence,it is able to give suitable forecasts over a period of seven days.TransCarbonNet takes advantage of a multi-head self-attention element to identify significant temporal connections,which means the Bi-LSTM element calculates sequential dependencies in both directions.Massive tests on two actual data sets indicate much improved results in comparison with the existing results,with mean relative errors of 15.3 percent and 12.7 percent,respectively.The framework has given explicable weights of attention that reveal critical periods that influence carbon intensity alterations,and informed decisions on the management of carbon sustainability.The effectiveness of the proposed solution has been validated in numerous cases of operations,and TransCarbonNet is established to be an effective tool when it comes to carbon-friendly optimization of the grid.展开更多
In contemporary computer vision,convolutional neural networks(CNNs)and vision transformers(ViTs)represent the two primary architectural paradigms for image recognition.While both approaches have been widely adopted in...In contemporary computer vision,convolutional neural networks(CNNs)and vision transformers(ViTs)represent the two primary architectural paradigms for image recognition.While both approaches have been widely adopted in medical imaging applications,they operate based on fundamentally different computational principles.This report attempts to provide brief application notes on ViTs and CNNs,particularly focusing on scenarios that guide the selection of one architecture over the other in practical medical implementations.Generally,CNNs rely on convolutional kernels,localized receptive fields,and weight sharing,enabling efficient hierarchical feature extraction.These properties contribute to strong performance in detecting spatially constrained patterns such as textures,edges,and anatomical boundaries,while maintaining relatively low computational requirements.ViTs,on the other hand,decompose images into smaller segments referred to as tokens and employ self-attention mechanisms to model relationships across the entire image.This global modeling capability allows ViTs to capture long-range dependencies that may be difficult for convolution-based architectures to learn.However,ViTs typically achieve optimal performance when trained on extremely large datasets or when supported by extensive pretraining,as their reduced inductive bias requires greater data exposure to learn robust representations.This report briefly examines the architectural structure,underlying mathematical foundations,and relative performance characteristics of CNNs and ViTs,drawing upon recent findings from contemporary research.Emphasis is placed on understanding how differences in data availability,computational resources,and task requirements influence model effectiveness across medical imaging domains.Most importantly,the report serves as a concise application guide for practitioners seeking informed implementation decisions between these two influential deep learning frameworks.展开更多
Convolutional neural networks(CNNs)have shown remarkable success across numerous tasks such as image classification,yet the theoretical understanding of their convergence remains underdeveloped compared to their empir...Convolutional neural networks(CNNs)have shown remarkable success across numerous tasks such as image classification,yet the theoretical understanding of their convergence remains underdeveloped compared to their empirical achievements.In this paper,the first filter learning framework with convergence-guaranteed learning laws for end-to-end learning of deep CNNs is proposed.Novel update laws with convergence analysis are formulated based on the mathematical representation of each layer in convolutional neural networks.The proposed learning laws enable concurrent updates of weights across all layers of the deep convolutional neural network and the analysis shows that the training errors converge to certain bounds which are dependent on the approximation errors.Case studies are conducted on benchmark datasets and the results show that the proposed concurrent filter learning framework guarantees the convergence and offers more consistent and reliable results during training with a trade-off in performance compared to stochastic gradient descent methods.This framework represents a significant step towards enhancing the reliability and effectiveness of deep convolutional neural network by developing a theoretical analysis which allows practical implementation of the learning laws with automatic tuning of the learning rate to guarantee the convergence during training.展开更多
The development of brain-computer interfaces(BCI)based on motor imagery(MI)has greatly improved patients’quality of life with movement disorders.The classification of upper limb MI has been widely studied and applied...The development of brain-computer interfaces(BCI)based on motor imagery(MI)has greatly improved patients’quality of life with movement disorders.The classification of upper limb MI has been widely studied and applied in many fields,including rehabilitation.However,the physiological representations of left and right lower limb movements are too close and activated deep in the cerebral cortex,making it difficult to distinguish their features.Therefore,classifying lower limbs motor imagery is more challenging.In this study,we propose a feature extraction method based on functional connectivity,which utilizes phase-locked values to construct a functional connectivity matrix as the features of the left and right legs,which can effectively avoid the problem of physiological representations of the left and right lower limbs being too close to each other during movement.In addition,considering the topology and the temporal characteristics of the electroencephalogram(EEG),we designed a temporal-spatial convolutional network(TSGCN)to capture the spatiotemporal information for classification.Experimental results show that the accuracy of the proposed method is higher than that of existing methods,achieving an average classification accuracy of 73.58%on the internal dataset.Finally,this study explains the network mechanism of left and right foot MI from the perspective of graph theoretic features and demonstrates the feasibility of decoding lower limb MI.展开更多
With network attack technology continuing to develop,traditional anomaly traffic detection methods that rely on feature engineering are increasingly insufficient in efficiency and accuracy.Graph Neural Network(GNN),a ...With network attack technology continuing to develop,traditional anomaly traffic detection methods that rely on feature engineering are increasingly insufficient in efficiency and accuracy.Graph Neural Network(GNN),a promising Deep Learning(DL)approach,has proven to be highly effective in identifying intricate patterns in graph⁃structured data and has already found wide applications in the field of network security.In this paper,we propose a hybrid Graph Convolutional Network(GCN)⁃GraphSAGE model for Anomaly Traffic Detection,namely HGS⁃ATD,which aims to improve the accuracy of anomaly traffic detection by leveraging edge feature learning to better capture the relationships between network entities.We validate the HGS⁃ATD model on four publicly available datasets,including NF⁃UNSW⁃NB15⁃v2.The experimental results show that the enhanced hybrid model is 5.71%to 10.25%higher than the baseline model in terms of accuracy,and the F1⁃score is 5.53%to 11.63%higher than the baseline model,proving that the model can effectively distinguish normal traffic from attack traffic and accurately classify various types of attacks.展开更多
This paper introduces a fuzzy C-means-based pooling layer for convolutional neural networks that explicitly models local uncertainty and ambiguity.Conventional pooling operations,such as max and average,apply rigid ag...This paper introduces a fuzzy C-means-based pooling layer for convolutional neural networks that explicitly models local uncertainty and ambiguity.Conventional pooling operations,such as max and average,apply rigid aggregation and often discard fine-grained boundary information.In contrast,our method computes soft membershipswithin each receptive field and aggregates cluster-wise responses throughmembership-weighted pooling,thereby preserving informative structure while reducing dimensionality.Being differentiable,the proposed layer operates as standard two-dimensional pooling.We evaluate our approach across various CNN backbones and open datasets,including CIFAR-10/100,STL-10,LFW,and ImageNette,and further probe small training set restrictions on MNIST and Fashion-MNIST.In these settings,the proposed pooling consistently improves accuracy and weighted F1 over conventional baselines,with particularly strong gains when training data are scarce.Even with less than 1%of the training set,ourmethodmaintains reliable performance,indicating improved sample efficiency and robustness to noisy or ambiguous local patterns.Overall,integrating soft memberships into the pooling operator provides a practical and generalizable inductive bias that enhances robustness and generalization in modern CNN pipelines.展开更多
Graph neural networks(GNN)have shown strong performance in node classification tasks,yet most existing models rely on uniform or shared weight aggregation,lacking flexibility in modeling the varying strength of relati...Graph neural networks(GNN)have shown strong performance in node classification tasks,yet most existing models rely on uniform or shared weight aggregation,lacking flexibility in modeling the varying strength of relationships among nodes.This paper proposes a novel graph coupling convolutional model that introduces an adaptive weighting mechanism to assign distinct importance to neighboring nodes based on their similarity to the central node.Unlike traditional methods,the proposed coupling strategy enhances the interpretability of node interactions while maintaining competitive classification performance.The model operates in the spatial domain,utilizing adjacency list structures for efficient convolution and addressing the limitations of weight sharing through a coupling-based similarity computation.Extensive experiments are conducted on five graph-structured datasets,including Cora,Citeseer,PubMed,Reddit,and BlogCatalog,as well as a custom topology dataset constructed from the Open University Learning Analytics Dataset(OULAD)educational platform.Results demonstrate that the proposed model achieves good classification accuracy,while significantly reducing training time through direct second-order neighbor fusion and data preprocessing.Moreover,analysis of neighborhood order reveals that considering third-order neighbors offers limited accuracy gains but introduces considerable computational overhead,confirming the efficiency of first-and second-order convolution in practical applications.Overall,the proposed graph coupling model offers a lightweight,interpretable,and effective framework for multi-label node classification in complex networks.展开更多
This study proposes a multi-scale simplified residual convolutional neural network(MS-SRCNN)for the precise prediction of Mg-Nd binary alloy compositions from scanning electron microscope(SEM)images.A multi-scale data...This study proposes a multi-scale simplified residual convolutional neural network(MS-SRCNN)for the precise prediction of Mg-Nd binary alloy compositions from scanning electron microscope(SEM)images.A multi-scale data structure is established by spatially aligning and stacking SEM images at different magnifications.The MS-SRCNN significantly reduces computational runtime by over 90%compared to traditional architectures like ResNet50,VGG16,and VGG19,without compromising prediction accuracy.The model demonstrates more excellent predictive performance,achieving a>5%increase in R^(2) compared to single-scale models.Furthermore,the MS-SRCNN exhibits robust composition prediction capability across other Mg-based binary alloys,including Mg-La,Mg-Sn,Mg-Ce,Mg-Sm,Mg-Ag,and Mg-Y,thereby emphasizing its generalization and extrapolation potential.This research establishes a non-destructive,microstructure-informed composition analysis framework,reduces characterization time compared to traditional experiment methods and provides insights into the composition-microstructure relationship in diverse material systems.展开更多
With the rapid development of the Artificial Intelligence of Things(AIoT),convolutional neural networks(CNNs)have demonstrated potential and remarkable performance in AIoT applications due to their excellent performan...With the rapid development of the Artificial Intelligence of Things(AIoT),convolutional neural networks(CNNs)have demonstrated potential and remarkable performance in AIoT applications due to their excellent performance in various inference tasks.However,the users have concerns about privacy leakage for the use of AI and the performance and efficiency of computing on resource-constrained IoT edge devices.Therefore,this paper proposes an efficient privacy-preserving CNN framework(i.e.,EPPA)based on the Fully Homomorphic Encryption(FHE)scheme for AIoT application scenarios.In the plaintext domain,we verify schemes with different activation structures to determine the actual activation functions applicable to the corresponding ciphertext domain.Within the encryption domain,we integrate batch normalization(BN)into the convolutional layers to simplify the computation process.For nonlinear activation functions,we use composite polynomials for approximate calculation.Regarding the noise accumulation caused by homomorphic multiplication operations,we realize the refreshment of ciphertext noise through minimal“decryption-encryption”interactions,instead of adopting bootstrapping operations.Additionally,in practical implementation,we convert three-dimensional convolution into two-dimensional convolution to reduce the amount of computation in the encryption domain.Finally,we conduct extensive experiments on four IoT datasets,different CNN architectures,and two platforms with different resource configurations to evaluate the performance of EPPA in detail.展开更多
Partial least squares (PLS) model maximizes the covariance between process variables and quality variables,making it widely used in quality-related fault detection.However,traditional PLS methods focus primarily on li...Partial least squares (PLS) model maximizes the covariance between process variables and quality variables,making it widely used in quality-related fault detection.However,traditional PLS methods focus primarily on linear processes,leading to poor performance in dynamic nonlinear processes.In this paper,a novel quality-related fault detection method,named DiCAE-PLS,is developed by combining dynamic-inner convolutional autoencoder with PLS.In the proposed DiCAE-PLS method,latent features are first extracted through dynamic-inner convolutional autoencoder (DiCAE) to capture process dynamics and nonlinearity from process variables.Then,a PLS model is established to build the relationship between the extracted latent features and the final product quality.To detect quality-related faults,Hotelling's T^(2) statistic is employed.The developed quality-related fault detection is applied to the widely used industrial benchmark of the Tennessee.展开更多
With the increasing complexity of industrial automation,planetary gearboxes play a vital role in largescale equipment transmission systems,directly impacting operational efficiency and safety.Traditional maintenance s...With the increasing complexity of industrial automation,planetary gearboxes play a vital role in largescale equipment transmission systems,directly impacting operational efficiency and safety.Traditional maintenance strategies often struggle to accurately predict the degradation process of equipment,leading to excessive maintenance costs or potential failure risks.However,existing prediction methods based on statistical models are difficult to adapt to nonlinear degradation processes.To address these challenges,this study proposes a novel condition-based maintenance framework for planetary gearboxes.A comprehensive full-lifecycle degradation experiment was conducted to collect raw vibration signals,which were then processed using a temporal convolutional network autoencoder with multi-scale perception capability to extract deep temporal degradation features,enabling the collaborative extraction of longperiod meshing frequencies and short-term impact features from the vibration signals.Kernel principal component analysis was employed to fuse and normalize these features,enhancing the characterization of degradation progression.A nonlinear Wiener process was used to model the degradation trajectory,with a threshold decay function introduced to dynamically adjust maintenance strategies,and model parameters optimized through maximum likelihood estimation.Meanwhile,the maintenance strategy was optimized to minimize costs per unit time,determining the optimal maintenance timing and preventive maintenance threshold.The comprehensive indicator of degradation trends extracted by this method reaches 0.756,which is 41.2%higher than that of traditional time-domain features;the dynamic threshold strategy reduces the maintenance cost per unit time to 55.56,which is 8.9%better than that of the static threshold optimization.Experimental results demonstrate significant reductions in maintenance costs while enhancing system reliability and safety.This study realizes the organic integration of deep learning and reliability theory in the maintenance of planetary gearboxes,provides an interpretable solution for the predictive maintenance of complex mechanical systems,and promotes the development of condition-based maintenance strategies for planetary gearboxes.展开更多
The prediction of sea surface partial pressure of carbon dioxide(pCO_(2))in the South China Sea is crucial for understanding the region’s contribution to the global carbon budget and its interactions with climate cha...The prediction of sea surface partial pressure of carbon dioxide(pCO_(2))in the South China Sea is crucial for understanding the region’s contribution to the global carbon budget and its interactions with climate change.We applied the Spatiotemporal Convolutional Long Short-Term Memory(STConvLSTM)model,integrating key environmental factors including sea surface temperature(SST),sea surface salinity(SSS),and chlorophyll a(Chl a),to predict and analyze sea surface pCO_(2)in the South China Sea.The model demonstrated high accuracy in short-term predictions(1 month),with a mean absolute error(MAE)of 0.394,a root mean square error(RMSE)of 0.659,and a coefficient of determination(R^(2))of 0.998.For long-term predictions(12 months),the model maintained its predictive capability,with an MAE of 0.667,RMSE of 1.255,and R^(2)of 0.994.Feature importance analysis revealed that sea surface pCO_(2)and SST were the main drivers of the model’s predictions,whereas Chl a and SSS had relatively minor impacts.The model’s generalization ability was further validated in the northwest Pacific Ocean and tropical Pacific Ocean,where it successfully captured the spatiotemporal variation in pCO_(2)with small prediction errors.The ST-ConvLSTM model provides an efficient and accurate tool for forecasting and analyzing sea surface pCO_(2)in the South China Sea,offering new insights into global carbon cycling and climate change.This study demonstrates the potential of deep learning in marine science and provides a significant technical support for global changes and marine ecosystem research.展开更多
WiFi-based human activity recognition(HAR)provides a non-intrusive approach for ubiquitous monitoring;however,achieving both high accuracy and robustness simultaneously remains a significant challenge.This paper propo...WiFi-based human activity recognition(HAR)provides a non-intrusive approach for ubiquitous monitoring;however,achieving both high accuracy and robustness simultaneously remains a significant challenge.This paper proposes a Convolutional Neural Network with Enhanced Convolutional Block Attention Module(CNN-ECBAM)framework.The approach systematically converts raw Channel State Information(CSI)into pseudo-color images,effectively preserving essential signal characteristics for deep neural network processing.The core innovation is an Enhanced Convolutional Block Attention Module(ECBAM),tailored to CSI data characteristics,which integrates Efficient Channel Attention(ECA)and Multi-Scale Spatial Attention(MSSA).By employing learnable adaptive fusion weights,it achieves dynamic synergy between channel and spatial features,enabling the network to capture highly discriminative spatiotemporal patterns.The ECBAM module is integrated into a unified Convolutional Neural Network(CNN)to form the overall CNN-ECBAM model.Experimental results on the UT-HAR and NTU-Fi_HAR datasets demonstrate that CNN-ECBAM achieves competitive performance in recognition accuracy and outperforms mainstream baseline models.Specifically,it attains 99.20%accuracy on UT-HAR(surpassing ResNet-18 at 98.60%)and achieves 100%accuracy on NTU-Fi_HAR(exceeding GAF-CNN at 99.62%).These results validate the effectiveness of the proposed method for high-precision and reliable WiFi-based HAR.展开更多
critical for guiding treatment and improving patient outcomes.Traditional molecular subtyping via immuno-histochemistry(IHC)test is invasive,time-consuming,and may not fully represent tumor heterogeneity.This study pr...critical for guiding treatment and improving patient outcomes.Traditional molecular subtyping via immuno-histochemistry(IHC)test is invasive,time-consuming,and may not fully represent tumor heterogeneity.This study proposes a non-invasive approach using digital mammography images and deep learning algorithm for classifying breast cancer molecular subtypes.Four pretrained models,including two Convolutional Neural Networks(MobileNet_V3_Large and VGG-16)and two Vision Transformers(ViT_B_16 and ViT_Base_Patch16_Clip_224)were fine-tuned to classify images into HER2-enriched,Luminal,Normal-like,and Triple Negative subtypes.Hyperparameter tuning,including learning rate adjustment and layer freezing strategies,was applied to optimize performance.Among the evaluated models,ViT_Base_Patch16_Clip_224 achieved the highest test accuracy(94.44%),with equally high precision,recall,and F1-score of 0.94,demonstrating excellent generalization.MobileNet_V3_Large achieved the same accuracy but showed less training stability.In contrast,VGG-16 recorded the lowest performance,indicating a limitation in its generalizability for this classification task.The study also highlighted the superior performance of the Vision Transformer models over CNNs,particularly due to their ability to capture global contextual features and the benefit of CLIP-based pretraining in ViT_Base_Patch16_Clip_224.To enhance clinical applicability,a graphical user interface(GUI)named“BCMS Dx”was developed for streamlined subtype prediction.Deep learning applied to mammography has proven effective for accurate and non-invasive molecular subtyping.The proposed Vision Transformer-based model and supporting GUI offer a promising direction for augmenting diagnostic workflows,minimizing the need for invasive procedures,and advancing personalized breast cancer management.展开更多
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 Natural Science Foundation of China(Grant Nos.62472149,62376089,62202147)Hubei Provincial Science and Technology Plan Project(2023BCB04100).
文摘Accurate traffic flow prediction has a profound impact on modern traffic management. Traffic flow has complex spatial-temporal correlations and periodicity, which poses difficulties for precise prediction. To address this problem, a Multi-head Self-attention and Spatial-Temporal Graph Convolutional Network (MSSTGCN) for multiscale traffic flow prediction is proposed. Firstly, to capture the hidden traffic periodicity of traffic flow, traffic flow is divided into three kinds of periods, including hourly, daily, and weekly data. Secondly, a graph attention residual layer is constructed to learn the global spatial features across regions. Local spatial-temporal dependence is captured by using a T-GCN module. Thirdly, a transformer layer is introduced to learn the long-term dependence in time. A position embedding mechanism is introduced to label position information for all traffic sequences. Thus, this multi-head self-attention mechanism can recognize the sequence order and allocate weights for different time nodes. Experimental results on four real-world datasets show that the MSSTGCN performs better than the baseline methods and can be successfully adapted to traffic prediction tasks.
基金supported by the National Natural Science Foundation of China(No.52277055).
文摘Traditional data-driven fault diagnosis methods depend on expert experience to manually extract effective fault features of signals,which has certain limitations.Conversely,deep learning techniques have gained prominence as a central focus of research in the field of fault diagnosis by strong fault feature extraction ability and end-to-end fault diagnosis efficiency.Recently,utilizing the respective advantages of convolution neural network(CNN)and Transformer in local and global feature extraction,research on cooperating the two have demonstrated promise in the field of fault diagnosis.However,the cross-channel convolution mechanism in CNN and the self-attention calculations in Transformer contribute to excessive complexity in the cooperative model.This complexity results in high computational costs and limited industrial applicability.To tackle the above challenges,this paper proposes a lightweight CNN-Transformer named as SEFormer for rotating machinery fault diagnosis.First,a separable multiscale depthwise convolution block is designed to extract and integrate multiscale feature information from different channel dimensions of vibration signals.Then,an efficient self-attention block is developed to capture critical fine-grained features of the signal from a global perspective.Finally,experimental results on the planetary gearbox dataset and themotor roller bearing dataset prove that the proposed framework can balance the advantages of robustness,generalization and lightweight compared to recent state-of-the-art fault diagnosis models based on CNN and Transformer.This study presents a feasible strategy for developing a lightweight rotating machinery fault diagnosis framework aimed at economical deployment.
基金supported by the National Natural Science Foundation of China(62402399)the New Chongqing Youth Innovation Talent Project(CSTB2024NSCQ-QCXMX0035)。
文摘Dear Editor,D2This letter presents a node feature similarity preserving graph convolutional framework P G.Graph neural networks(GNNs)have garnered significant attention for their efficacy in learning graph representations across diverse real-world applications.
基金partially supported by the National Natural Science Foundation of China(62075137)the Guangdong Basic and Applied Basic Research Foundation(2023A1515140161)+3 种基金the Guangxi Natural Science Foundation of China(2021JJB 110003)the Dongguan Science and Technology of Social Development Program(20231800936312)the high-level talent program of Dongguan University of Technology(No.221110080)the Sanming Project of Medicine in Shenzhen(No.SZSM202103014).
文摘Many spore-forming Bacillus species can cause serious human diseases,because of accidental Bacillusspore infection.Thus,developing an identification strategy with both high sensitivity and specificity is greatly in demand.In this work,we proposed a novel approach named multi-head self-attention mechanism-guided neural network Raman platform to identify living Bacillus spores within a single-cell resolution.The multi-head self-attention mechanism-guided neural network Raman platform was created by combining single-cell Raman spectroscopy,convolutional neural network(CNN),and multi-head self-attention mechanism.To address the limited size of the original spectra dataset,Gaussian noise-based spectra augmentation was employed to increase the number of single-cell Raman spectra datasets for CNN training.Owing to the assistance of both spectra augmentation and multi-head self-attention mechanism,the obtained prediction accuracy of five Bacillus spore species was further improved from 92.29±0.82%to 99.43±0.15%.To figure out the spectra differences covered by the multi-head self-attention mechanism-guided CNN,the relative classification weight from typical Raman bands was visualized via multi-head self-attention mechanism curve.In the process of spectra augmentation from 0 to 1000,the distribution of relative classification weight varied from a discrete state to a more concentrated phase.More importantly,these highlighted four Raman bands(1017,1449,1576,and 1660 cm^(-1))were assigned large weights,showing that the spectra differences in the Raman bands produced the largest contribution to prediction accuracy.It can be foreseen that,our proposed sorting platform has great potential in accurately identifying Bacillus and its related genera species at a single-cell level.
基金supported by the Science and Technology Innovation Key R&D Program of Chongqing(CSTB2025TIAD-STX0032)National Key Research and Development Program of China(2024YFF0908200)+1 种基金the Chongqing Technology Innovation and Application Development Special Key Project(CSTB2024TIAD-KPX0018)the Southwest University Graduate Student Research Innovation(SWUB24051)。
文摘Dear Editor,The letter proposes a tensor low-rank orthogonal compression(TLOC)model for a convolutional neural network(CNN),which facilitates its efficient and highly-accurate low-rank representation.Model compression is crucial for deploying deep neural network(DNN)models on resource-constrained embedded devices.
基金funded by the Deanship of Scientific Research and Libraries at Princess Nourah bint Abdulrahman University,through the“Nafea”Program,Grant No.(NP-45-082).
文摘Sustainable energy systems will entail a change in the carbon intensity projections,which should be carried out in a proper manner to facilitate the smooth running of the grid and reduce greenhouse emissions.The present article outlines the TransCarbonNet,a novel hybrid deep learning framework with self-attention characteristics added to the bidirectional Long Short-Term Memory(Bi-LSTM)network to forecast the carbon intensity of the grid several days.The proposed temporal fusion model not only learns the local temporal interactions but also the long-term patterns of the carbon emission data;hence,it is able to give suitable forecasts over a period of seven days.TransCarbonNet takes advantage of a multi-head self-attention element to identify significant temporal connections,which means the Bi-LSTM element calculates sequential dependencies in both directions.Massive tests on two actual data sets indicate much improved results in comparison with the existing results,with mean relative errors of 15.3 percent and 12.7 percent,respectively.The framework has given explicable weights of attention that reveal critical periods that influence carbon intensity alterations,and informed decisions on the management of carbon sustainability.The effectiveness of the proposed solution has been validated in numerous cases of operations,and TransCarbonNet is established to be an effective tool when it comes to carbon-friendly optimization of the grid.
文摘In contemporary computer vision,convolutional neural networks(CNNs)and vision transformers(ViTs)represent the two primary architectural paradigms for image recognition.While both approaches have been widely adopted in medical imaging applications,they operate based on fundamentally different computational principles.This report attempts to provide brief application notes on ViTs and CNNs,particularly focusing on scenarios that guide the selection of one architecture over the other in practical medical implementations.Generally,CNNs rely on convolutional kernels,localized receptive fields,and weight sharing,enabling efficient hierarchical feature extraction.These properties contribute to strong performance in detecting spatially constrained patterns such as textures,edges,and anatomical boundaries,while maintaining relatively low computational requirements.ViTs,on the other hand,decompose images into smaller segments referred to as tokens and employ self-attention mechanisms to model relationships across the entire image.This global modeling capability allows ViTs to capture long-range dependencies that may be difficult for convolution-based architectures to learn.However,ViTs typically achieve optimal performance when trained on extremely large datasets or when supported by extensive pretraining,as their reduced inductive bias requires greater data exposure to learn robust representations.This report briefly examines the architectural structure,underlying mathematical foundations,and relative performance characteristics of CNNs and ViTs,drawing upon recent findings from contemporary research.Emphasis is placed on understanding how differences in data availability,computational resources,and task requirements influence model effectiveness across medical imaging domains.Most importantly,the report serves as a concise application guide for practitioners seeking informed implementation decisions between these two influential deep learning frameworks.
基金supported by the Ministry of Education(MOE)Singapore,Academic Research Fund(AcRF)Tier 1(RG65/22)。
文摘Convolutional neural networks(CNNs)have shown remarkable success across numerous tasks such as image classification,yet the theoretical understanding of their convergence remains underdeveloped compared to their empirical achievements.In this paper,the first filter learning framework with convergence-guaranteed learning laws for end-to-end learning of deep CNNs is proposed.Novel update laws with convergence analysis are formulated based on the mathematical representation of each layer in convolutional neural networks.The proposed learning laws enable concurrent updates of weights across all layers of the deep convolutional neural network and the analysis shows that the training errors converge to certain bounds which are dependent on the approximation errors.Case studies are conducted on benchmark datasets and the results show that the proposed concurrent filter learning framework guarantees the convergence and offers more consistent and reliable results during training with a trade-off in performance compared to stochastic gradient descent methods.This framework represents a significant step towards enhancing the reliability and effectiveness of deep convolutional neural network by developing a theoretical analysis which allows practical implementation of the learning laws with automatic tuning of the learning rate to guarantee the convergence during training.
基金supported in part by the National Natural Science Foundation of China under Grant 62172368the Natural Science Foundation of Zhejiang Province under Grant LR22F020003.
文摘The development of brain-computer interfaces(BCI)based on motor imagery(MI)has greatly improved patients’quality of life with movement disorders.The classification of upper limb MI has been widely studied and applied in many fields,including rehabilitation.However,the physiological representations of left and right lower limb movements are too close and activated deep in the cerebral cortex,making it difficult to distinguish their features.Therefore,classifying lower limbs motor imagery is more challenging.In this study,we propose a feature extraction method based on functional connectivity,which utilizes phase-locked values to construct a functional connectivity matrix as the features of the left and right legs,which can effectively avoid the problem of physiological representations of the left and right lower limbs being too close to each other during movement.In addition,considering the topology and the temporal characteristics of the electroencephalogram(EEG),we designed a temporal-spatial convolutional network(TSGCN)to capture the spatiotemporal information for classification.Experimental results show that the accuracy of the proposed method is higher than that of existing methods,achieving an average classification accuracy of 73.58%on the internal dataset.Finally,this study explains the network mechanism of left and right foot MI from the perspective of graph theoretic features and demonstrates the feasibility of decoding lower limb MI.
基金National Natural Science Foundation of China(Grant No.62103434)National Science Fund for Distinguished Young Scholars(Grant No.62176263).
文摘With network attack technology continuing to develop,traditional anomaly traffic detection methods that rely on feature engineering are increasingly insufficient in efficiency and accuracy.Graph Neural Network(GNN),a promising Deep Learning(DL)approach,has proven to be highly effective in identifying intricate patterns in graph⁃structured data and has already found wide applications in the field of network security.In this paper,we propose a hybrid Graph Convolutional Network(GCN)⁃GraphSAGE model for Anomaly Traffic Detection,namely HGS⁃ATD,which aims to improve the accuracy of anomaly traffic detection by leveraging edge feature learning to better capture the relationships between network entities.We validate the HGS⁃ATD model on four publicly available datasets,including NF⁃UNSW⁃NB15⁃v2.The experimental results show that the enhanced hybrid model is 5.71%to 10.25%higher than the baseline model in terms of accuracy,and the F1⁃score is 5.53%to 11.63%higher than the baseline model,proving that the model can effectively distinguish normal traffic from attack traffic and accurately classify various types of attacks.
文摘This paper introduces a fuzzy C-means-based pooling layer for convolutional neural networks that explicitly models local uncertainty and ambiguity.Conventional pooling operations,such as max and average,apply rigid aggregation and often discard fine-grained boundary information.In contrast,our method computes soft membershipswithin each receptive field and aggregates cluster-wise responses throughmembership-weighted pooling,thereby preserving informative structure while reducing dimensionality.Being differentiable,the proposed layer operates as standard two-dimensional pooling.We evaluate our approach across various CNN backbones and open datasets,including CIFAR-10/100,STL-10,LFW,and ImageNette,and further probe small training set restrictions on MNIST and Fashion-MNIST.In these settings,the proposed pooling consistently improves accuracy and weighted F1 over conventional baselines,with particularly strong gains when training data are scarce.Even with less than 1%of the training set,ourmethodmaintains reliable performance,indicating improved sample efficiency and robustness to noisy or ambiguous local patterns.Overall,integrating soft memberships into the pooling operator provides a practical and generalizable inductive bias that enhances robustness and generalization in modern CNN pipelines.
基金Support by Sichuan Science and Technology Program[2023YFSY0026,2023YFH0004]Guangzhou Huashang University[2024HSZD01,HS2023JYSZH01].
文摘Graph neural networks(GNN)have shown strong performance in node classification tasks,yet most existing models rely on uniform or shared weight aggregation,lacking flexibility in modeling the varying strength of relationships among nodes.This paper proposes a novel graph coupling convolutional model that introduces an adaptive weighting mechanism to assign distinct importance to neighboring nodes based on their similarity to the central node.Unlike traditional methods,the proposed coupling strategy enhances the interpretability of node interactions while maintaining competitive classification performance.The model operates in the spatial domain,utilizing adjacency list structures for efficient convolution and addressing the limitations of weight sharing through a coupling-based similarity computation.Extensive experiments are conducted on five graph-structured datasets,including Cora,Citeseer,PubMed,Reddit,and BlogCatalog,as well as a custom topology dataset constructed from the Open University Learning Analytics Dataset(OULAD)educational platform.Results demonstrate that the proposed model achieves good classification accuracy,while significantly reducing training time through direct second-order neighbor fusion and data preprocessing.Moreover,analysis of neighborhood order reveals that considering third-order neighbors offers limited accuracy gains but introduces considerable computational overhead,confirming the efficiency of first-and second-order convolution in practical applications.Overall,the proposed graph coupling model offers a lightweight,interpretable,and effective framework for multi-label node classification in complex networks.
基金funded by the National Natural Science Foundation of China(No.52204407)the Natural Science Foundation of Jiangsu Province(No.BK20220595)the China Postdoctoral Science Foundation(No.2022M723689).
文摘This study proposes a multi-scale simplified residual convolutional neural network(MS-SRCNN)for the precise prediction of Mg-Nd binary alloy compositions from scanning electron microscope(SEM)images.A multi-scale data structure is established by spatially aligning and stacking SEM images at different magnifications.The MS-SRCNN significantly reduces computational runtime by over 90%compared to traditional architectures like ResNet50,VGG16,and VGG19,without compromising prediction accuracy.The model demonstrates more excellent predictive performance,achieving a>5%increase in R^(2) compared to single-scale models.Furthermore,the MS-SRCNN exhibits robust composition prediction capability across other Mg-based binary alloys,including Mg-La,Mg-Sn,Mg-Ce,Mg-Sm,Mg-Ag,and Mg-Y,thereby emphasizing its generalization and extrapolation potential.This research establishes a non-destructive,microstructure-informed composition analysis framework,reduces characterization time compared to traditional experiment methods and provides insights into the composition-microstructure relationship in diverse material systems.
基金supported by the Natural Science Foundation of China No.62362008the Major Scientific and Technological Special Project of Guizhou Province([2024]014).
文摘With the rapid development of the Artificial Intelligence of Things(AIoT),convolutional neural networks(CNNs)have demonstrated potential and remarkable performance in AIoT applications due to their excellent performance in various inference tasks.However,the users have concerns about privacy leakage for the use of AI and the performance and efficiency of computing on resource-constrained IoT edge devices.Therefore,this paper proposes an efficient privacy-preserving CNN framework(i.e.,EPPA)based on the Fully Homomorphic Encryption(FHE)scheme for AIoT application scenarios.In the plaintext domain,we verify schemes with different activation structures to determine the actual activation functions applicable to the corresponding ciphertext domain.Within the encryption domain,we integrate batch normalization(BN)into the convolutional layers to simplify the computation process.For nonlinear activation functions,we use composite polynomials for approximate calculation.Regarding the noise accumulation caused by homomorphic multiplication operations,we realize the refreshment of ciphertext noise through minimal“decryption-encryption”interactions,instead of adopting bootstrapping operations.Additionally,in practical implementation,we convert three-dimensional convolution into two-dimensional convolution to reduce the amount of computation in the encryption domain.Finally,we conduct extensive experiments on four IoT datasets,different CNN architectures,and two platforms with different resource configurations to evaluate the performance of EPPA in detail.
基金supported in part by the National Natural Science Foundation of China(62573387)the Natural Science Foundation of Zhejiang province,China(LY24F030004)the Fundamental Research Funds of Zhejiang Sci-Tech University(25222139-Y).
文摘Partial least squares (PLS) model maximizes the covariance between process variables and quality variables,making it widely used in quality-related fault detection.However,traditional PLS methods focus primarily on linear processes,leading to poor performance in dynamic nonlinear processes.In this paper,a novel quality-related fault detection method,named DiCAE-PLS,is developed by combining dynamic-inner convolutional autoencoder with PLS.In the proposed DiCAE-PLS method,latent features are first extracted through dynamic-inner convolutional autoencoder (DiCAE) to capture process dynamics and nonlinearity from process variables.Then,a PLS model is established to build the relationship between the extracted latent features and the final product quality.To detect quality-related faults,Hotelling's T^(2) statistic is employed.The developed quality-related fault detection is applied to the widely used industrial benchmark of the Tennessee.
基金funded by scientific research projects under Grant JY2024B011.
文摘With the increasing complexity of industrial automation,planetary gearboxes play a vital role in largescale equipment transmission systems,directly impacting operational efficiency and safety.Traditional maintenance strategies often struggle to accurately predict the degradation process of equipment,leading to excessive maintenance costs or potential failure risks.However,existing prediction methods based on statistical models are difficult to adapt to nonlinear degradation processes.To address these challenges,this study proposes a novel condition-based maintenance framework for planetary gearboxes.A comprehensive full-lifecycle degradation experiment was conducted to collect raw vibration signals,which were then processed using a temporal convolutional network autoencoder with multi-scale perception capability to extract deep temporal degradation features,enabling the collaborative extraction of longperiod meshing frequencies and short-term impact features from the vibration signals.Kernel principal component analysis was employed to fuse and normalize these features,enhancing the characterization of degradation progression.A nonlinear Wiener process was used to model the degradation trajectory,with a threshold decay function introduced to dynamically adjust maintenance strategies,and model parameters optimized through maximum likelihood estimation.Meanwhile,the maintenance strategy was optimized to minimize costs per unit time,determining the optimal maintenance timing and preventive maintenance threshold.The comprehensive indicator of degradation trends extracted by this method reaches 0.756,which is 41.2%higher than that of traditional time-domain features;the dynamic threshold strategy reduces the maintenance cost per unit time to 55.56,which is 8.9%better than that of the static threshold optimization.Experimental results demonstrate significant reductions in maintenance costs while enhancing system reliability and safety.This study realizes the organic integration of deep learning and reliability theory in the maintenance of planetary gearboxes,provides an interpretable solution for the predictive maintenance of complex mechanical systems,and promotes the development of condition-based maintenance strategies for planetary gearboxes.
基金Supported by the National Key Research and Development Program of China(No.2023YFC3008202)the National Natural Science Foundation of China(No.42406019)the Scientific Research Fund of Zhejiang Provincial Education Department(No.Y202353066)。
文摘The prediction of sea surface partial pressure of carbon dioxide(pCO_(2))in the South China Sea is crucial for understanding the region’s contribution to the global carbon budget and its interactions with climate change.We applied the Spatiotemporal Convolutional Long Short-Term Memory(STConvLSTM)model,integrating key environmental factors including sea surface temperature(SST),sea surface salinity(SSS),and chlorophyll a(Chl a),to predict and analyze sea surface pCO_(2)in the South China Sea.The model demonstrated high accuracy in short-term predictions(1 month),with a mean absolute error(MAE)of 0.394,a root mean square error(RMSE)of 0.659,and a coefficient of determination(R^(2))of 0.998.For long-term predictions(12 months),the model maintained its predictive capability,with an MAE of 0.667,RMSE of 1.255,and R^(2)of 0.994.Feature importance analysis revealed that sea surface pCO_(2)and SST were the main drivers of the model’s predictions,whereas Chl a and SSS had relatively minor impacts.The model’s generalization ability was further validated in the northwest Pacific Ocean and tropical Pacific Ocean,where it successfully captured the spatiotemporal variation in pCO_(2)with small prediction errors.The ST-ConvLSTM model provides an efficient and accurate tool for forecasting and analyzing sea surface pCO_(2)in the South China Sea,offering new insights into global carbon cycling and climate change.This study demonstrates the potential of deep learning in marine science and provides a significant technical support for global changes and marine ecosystem research.
基金Supported by Anhui Provincial Engineering Research Center for Sports and Health Information Monitoring Technology(KF2023012)。
文摘WiFi-based human activity recognition(HAR)provides a non-intrusive approach for ubiquitous monitoring;however,achieving both high accuracy and robustness simultaneously remains a significant challenge.This paper proposes a Convolutional Neural Network with Enhanced Convolutional Block Attention Module(CNN-ECBAM)framework.The approach systematically converts raw Channel State Information(CSI)into pseudo-color images,effectively preserving essential signal characteristics for deep neural network processing.The core innovation is an Enhanced Convolutional Block Attention Module(ECBAM),tailored to CSI data characteristics,which integrates Efficient Channel Attention(ECA)and Multi-Scale Spatial Attention(MSSA).By employing learnable adaptive fusion weights,it achieves dynamic synergy between channel and spatial features,enabling the network to capture highly discriminative spatiotemporal patterns.The ECBAM module is integrated into a unified Convolutional Neural Network(CNN)to form the overall CNN-ECBAM model.Experimental results on the UT-HAR and NTU-Fi_HAR datasets demonstrate that CNN-ECBAM achieves competitive performance in recognition accuracy and outperforms mainstream baseline models.Specifically,it attains 99.20%accuracy on UT-HAR(surpassing ResNet-18 at 98.60%)and achieves 100%accuracy on NTU-Fi_HAR(exceeding GAF-CNN at 99.62%).These results validate the effectiveness of the proposed method for high-precision and reliable WiFi-based HAR.
基金funded by the Ministry of Higher Education(MoHE)Malaysia through the Fundamental Research Grant Scheme—Early Career Researcher(FRGS-EC),grant number FRGSEC/1/2024/ICT02/UNIMAP/02/8.
文摘critical for guiding treatment and improving patient outcomes.Traditional molecular subtyping via immuno-histochemistry(IHC)test is invasive,time-consuming,and may not fully represent tumor heterogeneity.This study proposes a non-invasive approach using digital mammography images and deep learning algorithm for classifying breast cancer molecular subtypes.Four pretrained models,including two Convolutional Neural Networks(MobileNet_V3_Large and VGG-16)and two Vision Transformers(ViT_B_16 and ViT_Base_Patch16_Clip_224)were fine-tuned to classify images into HER2-enriched,Luminal,Normal-like,and Triple Negative subtypes.Hyperparameter tuning,including learning rate adjustment and layer freezing strategies,was applied to optimize performance.Among the evaluated models,ViT_Base_Patch16_Clip_224 achieved the highest test accuracy(94.44%),with equally high precision,recall,and F1-score of 0.94,demonstrating excellent generalization.MobileNet_V3_Large achieved the same accuracy but showed less training stability.In contrast,VGG-16 recorded the lowest performance,indicating a limitation in its generalizability for this classification task.The study also highlighted the superior performance of the Vision Transformer models over CNNs,particularly due to their ability to capture global contextual features and the benefit of CLIP-based pretraining in ViT_Base_Patch16_Clip_224.To enhance clinical applicability,a graphical user interface(GUI)named“BCMS Dx”was developed for streamlined subtype prediction.Deep learning applied to mammography has proven effective for accurate and non-invasive molecular subtyping.The proposed Vision Transformer-based model and supporting GUI offer a promising direction for augmenting diagnostic workflows,minimizing the need for invasive procedures,and advancing personalized breast cancer management.
