Search

Early detection and treatment of breast cancer can significantly reduce patient mortality, and mammogram is an effective method for early screening. Computer-aided diagnosis (CAD) of mammography based on deep learning can assist radiologists in making more objective and accurate judgments. However, existing methods often depend on datasets with manual segmentation annotations. In addition, due to the large image sizes and small lesion proportions, many methods that do not use region of interest (ROI) mostly rely on multi-scale and multi-feature fusion models. These shortcomings increase the labor, money, and computational overhead of applying the model. Therefore, a deep location soft-embedding-based network with regional scoring (DLSEN-RS) is proposed. DLSEN-RS is an end-to-end mammography image classification method containing only one feature extractor and relies on positional embedding (PE) and aggregation pooling (AP) modules to locate lesion areas without bounding boxes, transfer learning, or multi-stage training. In particular, the introduced PE and AP modules exhibit versatility across various CNN models and improve the model's tumor localization and diagnostic accuracy for mammography images. Experiments are conducted on published INbreast and CBIS-DDSM datasets, and compared to previous state-of-the-art mammographic image classification methods, DLSEN-RS performed satisfactorily.

Breast cancer diagnosis is a critical step in clinical decision making, and this is achieved by making a pathological slide and gives a decision by the doctors, which is the method of final decision making for cancer diagnosis. Traditionally, the doctors usually check the pathological images by visual inspection under the microscope. Whole-slide images (WSIs) have supported the state-of-the-art diagnosis results and have been admitted as the gold standard clinically. However, this task is time-consuming and labour-intensive, and all of these limitations make low efficiency in decision making. Medical image processing protocols have been used for this task during the last decades and have obtained satisfactory results under some conditions; especially in the deep learning era, it has exhibited the advantages than those in the shallow learning period. In this paper, we proposed a novel breast cancer region mining framework based on deep pyramid architecture from multilevel and multiscale breast pathological WSIs. We incorporate the tissue- and cell-level information together and integrate these into a LSTM model for the final sequence modelling, which successfully keeps the WSIs' integration and is not mentioned by the prevalence frameworks. The experiment results demonstrated that our proposed framework greatly improved the detection accuracy than that only using tissue-level information., Competing Interests: The authors declare that they have no conflicts of interest., (Copyright © 2021 Dong Sui et al.)

Objective: To determine the diagnostic performance of a deep learning (DL) model in evaluating myometrial invasion (MI) depth on T2-weighted imaging (T2WI)-based endometrial cancer (EC) MR imaging (ECM)., Methods: We retrospectively enrolled 530 patients with pathologically proven EC at our institution between January 1, 2013, and December 31, 2017. All imaging data were reviewed on picture archiving and communication systems (PACS) server. Both sagittal and coronal T2WI-based MR images were used for lesion area determination. All MR images were divided into two groups: deep (more than 50%) and shallow (less than 50%) MI based on their pathological diagnosis. We trained a detection model based on YOLOv3 algorithm to locate the lesion area on ECM. Then, the detected regions were fed into a classification model based on DL network to identify MI depth automatically., Results: In the testing dataset, the trained model detected lesion regions with an average precision rate of 77.14% and 86.67% in both sagittal and coronal images, respectively. The classification model yielded an accuracy of 84.78%, a sensitivity of 66.67%, a specificity of 87.50%, a positive predictive value of 44.44%, and a negative predictive value of 94.59% in determining deep MI. The radiologists and trained network model together yielded an accuracy of 86.2%, a sensitivity of 77.8%, a specificity of 87.5%, a positive predictive value of 48.3%, and a negative predictive value of 96.3%., Conclusion: In this study, the DL network model derived from MR imaging provided a competitive, time-efficient diagnostic performance in MI depth identification., Key Points: • The models established with the deep learning method could help improve the diagnostic confidence and performance of MI identification based on endometrial cancer MR imaging. • The models enabled the classification of endometrial cancer MR images to the two categories with a sensitivity of 0.67, a specificity of 0.88, and an accuracy of 0.85. • Using the detected lesion region to evaluate myometrial invasion depth could remove redundant information in the image and provide more effective features.

A false data injection attack (FDIA) is the main attack method that threatens the security of smart grids. FDIAs mislead the control center to make wrong judgments by modifying the measurement data of the power grid system. Therefore, the effective and accurate detection of FDIAs is crucial for the safe operation of smart grids. However, the current deep learning-based methods do not fully exploit the short-term local characteristics and long-term dependencies of power grid data and have poor correlation with past and future time series information, resulting in a lack of credibility in the detection results. In view of this, an FDIA detection model combining a bidirectional temporal convolutional network and bidirectional gated recurrent unit with an attention mechanism (A-BiTG) was proposed. The proposed model utilizes a bidirectional time convolutional network (BiTCN) and bidirectional gated recurrent unit (BiGRU) to consider past and future temporal information in the grid. This enhances the ability of the model to capture long-term dependencies and extract features, while also solving the model's problem of exploding and vanishing gradients. In addition, an attention mechanism (AM) was added to dynamically assign weights to the extracted feature information and retain the most valuable features to improve the detection accuracy of the model. Finally, the proposed method was compared with existing methods on the IEEE 14-bus and IEEE 118-bus test systems. The results show that the proposed detection model is more robust and superior under different noise environments and FDIA signals with different intensities. [ABSTRACT FROM AUTHOR]

In recent years, the application of deep learning to remote sensing image super-resolution has achieved promising results. However, most deep learning-based methods are often trained on remote sensing datasets constructed by bicubic downsampling, and their recovery effects on real-world remote sensing images are often unsatisfactory. This is because the process of generating low-resolution (LR) images from high-resolution (HR) images to construct training data pairs (LR-HR) using simple bicubic downsampling cannot reflect the degradation process of real-world remote sensing dataset images. In this paper, we propose a multi-scale feature fusion kernel estimation with masked interpolation loss for real-world remote sensing images super-resolution (MFFILSR) to address this problem. MFFILSR is divided into two stages: degenerate kernel estimation and SR network training. In the first stage, we propose a multi-scale feature fusion kernel estimation network that can effectively fuse multi-scale information, making the estimated downsampling kernel closer to the degradation patterns of real-world remote sensing images. In the second stage, we introduce a masked interpolation loss during generator training, by masking the interpolation loss, the artefacts of generated images can be effectively reduced. Extensive experiments show that MFFILSR has satisfactory super-resolution reconstruction performance for real-world remote sensing images. [ABSTRACT FROM AUTHOR]

Deep learning has shown superiority to extract more representative features from multimedia data in recent years. Recently, the most typical graph convolutional networks (GCN) has achieved excellent performance in the semi-supervised framework-based data representation learning tasks. GCN successfully generalizes traditional convolutional neural networks to encode arbitrary graphs by exploiting the graph Laplacian-based sample structure information. However, GCN only fuses the static structure information. It is difficult to guarantee that its structure information is optimal during the training process and applicable for all practical applications. To tackle the above problem, in this paper, we propose a manifold regularized dynamic graph convolutional network (MRDGCN). The proposed MRDGCN automatically updates the structure information by manifold regularization until model fitting. In particular, we build an optimization convolution layer formulation to acquire the optimal structure information. Thus, MRDGCN can automatically learn high-level sample features to improve the performance of data representation learning. To demonstrate the effectiveness of our proposed model, we apply MRDGCN on the semi-supervised classification tasks. The extensive experiment results on human activity datasets and citation network datasets validate the performance of MRDGCN compared with GCN and other semi-supervised learning methods. [ABSTRACT FROM AUTHOR]

Many research groups in the structural health monitoring community have made efforts to utilize deep learning-based approaches for damage detection on a variety of structures. Among these approaches, structural damage detection through deep convolutional neural networks using raw structural response data has received great attention. However, structural responses are affected not only by structural properties but also by excitation characteristics. For detecting of structures' damage under seismic excitations, different seismic excitations definitely cause varied structural responses data. In practice, it is impossible to accurately predict the characteristics of future seismic excitation for pre-training the deep convolutional neural network. Therefore, it is essential to investigate the autonomous detection of structural element damage subject to unknown seismic excitation. In this article, a new approach is proposed for detecting structural damage subject to unknown seismic excitation based on a convolutional neural network with wavelet-based transmissibility of structural response data. The transmissibility functions of structural response data are used to eliminate the influence of different seismic excitations. Moreover, contrary to the traditional Fourier transform in the conventional transmissibility function, wavelet-based transmissibility function is presented using the ability in subtle information acquisition of wavelet transform. The wavelet-based transmissibility data of structural responses are used as the inputs to constructed deep convolutional neural networks. Both a numerical simulation example and an experimental test are used to validate the performance of the proposed approach based on deep convolutional neural network. [ABSTRACT FROM AUTHOR]

Deep learning-based object detection methods have achieved promising results in normal scenarios, while such methods often fail to locate objects from the disturbed images captured in inclement environments. Most existing methods utilized denoise modules to assist the detection network or exploit prior knowledge to reduce the environment interference remaining in the features, which ignores the essential role of the frequency element. From a novel perspective, we observe that the inclement environment alters the frequency content in the features, which in turn crushes the detection. To tickle this problem, we present the Magnitude Consistency Network (MCNet) to distill the irrelevant contents in the frequency domain. The MCNet is composed of the detection network and the magnitude corrector. The detection network is able to locate and classify objects. However, in inclement environments, the crucial information about the objects of interest is disrupted by the nuisance noise introduced from the inclement environments. The magnitude corrector can recover relevant information about the objects in the frequency domain by distilling the irrelevant factors and refining the affected features. Alternately optimizing the magnitude corrector and the detection network gradually makes the frequency content between the disturbed image and the clear image to be consistent. By distilling the irrelevant noise in the feature, the detection network can learn domain-invariant representations. Extensive experiments prove that the proposed method is effective and outperforms existing methods by a clear margin on four datasets. • We observe that compared to the clear images, the frequency content is altered in the inclement environment obviously, which causes the performance degradation of normal object detection approach • We propose a novel Magnitude Consistency Network (MCNet) to alleviate the distinction in the frequency components among the clear features and the disturbed features. • The experimental results demonstrate the proposed MCNet is supreme to the existing methods on the synthetic fog, synthetic rainy, natural fog and natural underwater datasets. [ABSTRACT FROM AUTHOR]

• A novel PINNFD method is proposed to solve the multi-frequency elastodynamics issue. • The proposed PINNFD allows the solution of elastodynamics in the frequency domain. • Fourier features for all frequencies can be embedded in the proposed PINNFD model. • The performance of the traditional PINN and the proposed PINNFD were compared. • The potential of the proposed PINNFD for engineering applications is validated. Despite the fact that physics-informed neural networks (PINN) have been developed rapidly in recent years, their inherent spectral bias makes it difficult to approximate multi-frequency target functions such as the solutions to elastodynamics problems. To address this challenge, this paper proposes a new PINN frequency domain (PINNFD) method on the basis of frequency domain inputs. Fourier features for all frequencies can thus be constructed and embedded in the model, which provides essential support for the PINNFD method to accurately approximate multi-frequency target functions. To validate the effectiveness of the proposed PINNFD method, the proposed method and the traditional method are applied to solve the elastodynamics problem in infinite media under various dynamic point loads, including single-frequency harmonic load, multi-frequency harmonic load, and multi-frequency random load. The results show that although the PINN with embedded Fourier features is able to achieve the simulation of elastodynamics problems under harmonic loads, it fails to solve the problems under multi-frequency random loads. Whereas the proposed PINNFD method achieves better results in all cases of loading conditions, which demonstrates the advancement of the proposed method in solving multi-frequency problems in engineering applications. [Display omitted] [ABSTRACT FROM AUTHOR]

• Extracting features by combining CNN and swin-transformer. • Lightweight method didn't degenerate the model with our 2.5D and 2-stream method. • Performance is close to previous 3D methods by only using 2D algorithms. • Only sMRI images are used as input data. Alzheimer's disease (AD) is a primary cause of dementia. Its early diagnosis is crucial to delay the progression of the disease. So far, many computer aided diagnosis (CAD) methods that combined deep learning algorithms and structural MRI have achieved encouraging results. To improve the AD diagnosis performance, more and more models are based on 3D algorithms, which make the training and deployment of these methods unaffordable. In this study, a CNN and swin-transformer based efficient model, Efficient Conv-Swin Net (ECSnet), was developed. In this model: (1) a 2.5D-subject method and two-stream structure are used to help the model to encode 3D information to 2D feature maps; (2) convolution blocks are applied in the early stages of the transformer-based backbone network to improve the generalization ability; (3) a series of lightweight approaches are applied to reduce the parameters and computational cost of the model to enable the model to train and infer efficiently. Due to the lack of multi-center data and the differences between test sets, it is difficult to make a fair comparison between the previous methods. Our model was trained on the ADNI dataset and evaluated on an independent test set from AIBL. After being lightened, our proposed method showed no performance degradation on both ADNI and AIBL compared to models such as swin-T tiny. The ECSnet achieved 92.8% balance accuracy and 91.1% sensitivity on the AIBL, which are better than those of previous works while the model is more efficient than those 3D methods. [ABSTRACT FROM AUTHOR]

Auto-Encoders, as one representative deep learning method, has demonstrated to achieve superior performance in many applications. Hence, it is drawing more and more attentions and variants of Auto-Encoders have been reported including Contractive Auto-Encoders, Denoising Auto-Encoders, Sparse Auto-Encoders and Nonnegativity Constraints Auto-Encoders. Recently, a Discriminative Auto-Encoders is reported to improve the performance by considering the within class and between class information. In this paper, we propose the Large Margin Auto-Encoders (LMAE) to further boost the discriminability by enforcing different class samples to be large marginally distributed in hidden feature space. Particularly, we stack the single-layer LMAE to construct a deep neural network to learn proper features. And finally we put these features into a softmax classifier for classification. Extensive experiments are conducted on the MNIST dataset and the CIFAR-10 dataset for classification respectively. The experimental results demonstrate that the proposed LMAE outperforms the traditional Auto-Encoders algorithm. [ABSTRACT FROM AUTHOR]

The problems of ground-borne vibration induced by running metro trains are becoming a major concern. Considering the uncertainty in train–track–tunnel–soil–building system, the vibration evaluation based on long-term monitoring is required to investigate the effect of vibration on the residents, sensitive instruments, and buildings to provide evidence for vibration mitigation design. In this study, a deep learning based approach is proposed to identify train induced vibration segments efficiently for subsequent vibration evaluations. Furthermore, an experiment is presented in which vibration monitoring on the tunnel wall, ground surface and floors in a building due to metro train passages were performed. The proposed identification approach is validated in the experiment, and then the characteristics and sources of the uncertainty for train induced ground-borne vibration are analysed to improve the quality of vibration evaluation. [ABSTRACT FROM AUTHOR]

In recent years, the rapid construction of roads and railways in cities has brought convenience to urban residents, but the noise generated by the functioning of these transport infrastructures also affects the normal life of residents in the buildings around them. Buildings adjacent to both roads and railways are exposed to multiple traffic noises at the same time, which presents difficulties for targeted noise assessment and control. To solve this problem, this paper proposes a deep learning-based traffic noise source identification method, which feeds the time-frequency spectrum of the noise or vibration monitoring signals into the trained deep learning based model, so that the sources of noise at various moments can be identified. In addition, the performance of the proposed method was presented using noise and vibration signals monitored in Beijing, and noise evaluations for two buildings were conducted as case studies to investigate the statistical characteristics of noise caused by rail and road traffic in buildings. The result demonstrates the feasibility of the proposed method, and shows that different types of noise pollution need to be attended to on different floors in buildings adjacent to underground railways, the lower floors of the building are exposed to more significant low-frequency structural noise induced by the underground railway, while the higher floors of the building are exposed to more significant airborne noise caused by road traffic. Also, extensive areas adjacent to elevated railways are affected by airborne noise, and buildings in these areas require attention to the prevention of noise. • A novel traffic noise identification method is proposed based on deep learning. • The feasibility of the method was verified using noises monitored in Beijing. • The ground floor of buildings near the metro is exposed to structural noise. • Extensive areas adjacent to elevated railways are affected by the airborne noise. [ABSTRACT FROM AUTHOR]

Deep learning has recently attracted extensive attention and developed significantly in remote sensing image super-resolution. Although remote sensing images are composed of various scenes, most existing methods consider each part equally. These methods ignore the salient objects (e.g., buildings, airplanes, and vehicles) that have more complex structures and require more attention in recovery processing. This paper proposes a saliency-guided remote sensing image super-resolution (SG-GAN) method to alleviate the above issue while maintaining the merits of GAN-based methods for the generation of perceptual-pleasant details. More specifically, we exploit the salient maps of images to guide the recovery in two aspects: On the one hand, the saliency detection network in SG-GAN learns more high-resolution saliency maps to provide additional structure priors. On the other hand, the well-designed saliency loss imposes a second-order restriction on the super-resolution process, which helps SG-GAN concentrate more on the salient objects of remote sensing images. Experimental results show that SG-GAN achieves competitive PSNR and SSIM compared with the advanced super-resolution methods. Visual results demonstrate our superiority in restoring structures while generating remote sensing super-resolution images. [ABSTRACT FROM AUTHOR]

In recent years, the application of deep learning has achieved a huge leap in the performance of remote sensing image super-resolution (SR). However, most of the existing SR methods employ bicubic downsampling of high-resolution (HR) images to obtain low-resolution (LR) images and use the obtained LR and HR images as training pairs. This supervised method that uses ideal kernel (bicubic) downsampled images to train the network will significantly degrade performance when used in realistic LR remote sensing images, usually resulting in blurry images. The main reason is that the degradation process of real remote sensing images is more complicated. The training data cannot reflect the SR problem of real remote sensing images. Inspired by the self-supervised methods, this paper proposes a cross-dimension attention guided self-supervised remote sensing single-image super-resolution method (CASSISR). It does not require pre-training on a dataset, only utilizes the internal information reproducibility of a single image, and uses the lower-resolution image downsampled from the input image to train the cross-dimension attention network (CDAN). The cross-dimension attention module (CDAM) selectively captures more useful internal duplicate information by modeling the interdependence of channel and spatial features and jointly learning their weights. The proposed CASSISR adapts well to real remote sensing image SR tasks. A large number of experiments show that CASSISR has achieved superior performance to current state-of-the-art methods. [ABSTRACT FROM AUTHOR]