Search

Background: This study intends to establish a combined prediction model that integrates the clinical symptoms,the lung lesion volume, and the radiomics features of patients with COVID-19, resulting in a new model to predict the severity of COVID-19. Methods: The clinical data of 386 patients with COVID-19 at several hospitals, as well as images of certain patients during their hospitalization, were collected retrospectively to create a database of patients with COVID-19 pneumonia. The contour of lungs and lesion locations may be retrieved from CT scans using a CT-image-based quantitative discrimination and trend analysis method for COVID-19 and the Mask R-CNN deep neural network model to create 3D data of lung lesions. The quantitative COVID-19 factors were then determined, on which the diagnosis of the development of the patients' symptoms could be established. Then, using an artificial neural network, a prediction model of the severity of COVID-19 was constructed by combining characteristic imaging features on CT slices with clinical factors. ANN neural network was used for training, and tenfold cross-validation was used to verify the prediction model. The diagnostic performance of this model is verified by the receiver operating characteristic (ROC) curve. Results: CT radiomics features extraction and analysis based on a deep neural network can detect COVID-19 patients with an 86% sensitivity and an 85% specificity. According to the ROC curve, the constructed severity prediction model indicates that the AUC of patients with severe COVID-19 is 0.761, with sensitivity and specificity of 79.1% and 73.1%, respectively. Conclusions: The combined prediction model for severe COVID-19 pneumonia, which is based on deep learning and integrates clinical aspects, pulmonary lesion volume, and radiomics features of patients, has a remarkable differential ability for predicting the course of disease in COVID-19 patients. This may assist in the early prevention of severe COVID-19 symptoms. [ABSTRACT FROM AUTHOR]

Graph neural networks (GNNs) have drawn increasing attention in recent years and achieved remarkable performance in many graph-based tasks, especially in semi-supervised learning on graphs. However, most existing GNNs are based on the message-passing paradigm to iteratively aggregate neighborhood information in a single topology space. Despite their success, the expressive power of GNNs is limited by some drawbacks, such as inflexibility of message source expansion, negligence of node-level message output discrepancy, and restriction of single message space. To address these drawbacks, we present a novel message-passing paradigm, based on the properties of multi-step message source, node-specific message output, and multi-space message interaction. To verify its validity, we instantiate the new message-passing paradigm as a Dual-Perception Graph Neural Network (DPGNN), which applies a node-to-step attention mechanism to aggregate node-specific multi-step neighborhood information adaptively. Our proposed DPGNN can capture the structural neighborhood information and the feature-related information simultaneously for graph representation learning. Experimental results on six benchmark datasets with different topological structures demonstrate that our method outperforms the latest state-of-the-art models, which proves the superiority and versatility of our method. To our knowledge, we are the first to consider node-specific message passing in the GNNs. [ABSTRACT FROM AUTHOR]

Artificial neural network (ANN) was proposed as an effective method to help multipath ultrasonic flowmeter (UFM) reduce its measurement error when determining the flowrate of complex flow field. However, the effectiveness of the ANN method heavily depends on the network architecture specified by the designer, and also the provided initial weights and layer biases. This hinders the ANN to be widely used in actual UFMs. This paper proposes a genetic algorithm (GA) optimized ANN method (GANN) to be used in UFMs. The GA is utilized to determine ANN architecture based on its efficient parallel advantage in global search to replace traditional trial and error approach or empirical method. In addition, the initial weights and biases are optimized by GA, which is capable to avoid the network getting stuck with local minimum and qualify it for good generalization ability. Tests are implemented on numerical models simulating a six-path UFM installed downstream single or double elbows with complex flow field passing through it. Comparisons between the results of classical Gaussian quadrature, existing ANN, and proposed GANN are given for verification. The flowrate determination error of GANN is only 40% $\sim 62$ % of existing ANN, which is extremely smaller than the Gaussian quadrature method. In addition, the output of GANN is much more stable than existing ANN, as proper initial weights and layer biases will be generated by GA for iteration process even random data are given for training. [ABSTRACT FROM PUBLISHER]

• A new deep learning framework is proposed in this paper, which can leverage the object label inter-dependent for weakly supervised learning. • We creatively introduce a novel initialization method of label embeddings for inter-relationships learning. • Ablation studies of our model are provided in detail. • Our models can be applied to object classification and weakly supervised pointwise object localization. In computer vision, the research community has been looking to how to benefit from weakly supervised learning that utilizes easily obtained image-level labels to train neural network models. The existing deep convolutional neural networks for weakly supervised learning, however, generally do not fully exploit the label dependencies in an image. To make full use of this information, in this paper, we propose a new framework for weakly supervised learning of deep convolutional neural networks, introducing graph convolutional networks to capture the semantic label co-occurrence in an image. Moreover, we propose a novel initialization method for label embedding in graph convolutional networks, which enables a smoother optimization for interrelationships learning. Extensive experiments and comparisons on four public benchmark datasets (PASCAL VOC 2007, PASCAL VOC 2012, Microsoft COCO, and NUS-WIDE) show the superior performance of our approach in both image classification and weakly supervised pointwise object localization. These results lead us to conclude that the label dependencies in the input image can provide valuable evidence for learning strongly localized features. [ABSTRACT FROM AUTHOR]

• A neural network method to distinguish different lasers in dual-laser systems by a superimposed spot image is proposed. • This technique can obtain the pointing errors of dual lasers by one measurement of a superimposed without any beam-splitting devices, thus suppling simplified structure design with better reliability and synchronization. • This technique has the potential to separate more lasers in multi-beam systems by methods such as increasing the tilting angle of a detector and reducing the focal length of the lens. In various laser applications with multiple beams, measuring the pointing errors of different beams is a prerequisite. Normally, this is done by multiple measurements of corresponding laser spots after separating each beam. In this study, a neural network method is utilized to predict the position and angular errors of dual lasers from a single superimposed spot image on a tilted detector, and thus separates the two laser spots. We find that the key point of decoupling the superimposed image is to enlarge the size difference of the two spots on the detector. Also, optimized intensities of the dual spots to balance their prediction abilities, can help obtain a better prediction performance. This technique provides a novel strategy for advancing pointing measurement in multi-beam optical systems with simplified structure design, better reliability and synchronization. [ABSTRACT FROM AUTHOR]

Leveraging on the underlying low-dimensional structure of data, low-rank and sparse modeling approaches have achieved great success in a wide range of applications. However, in many applications the data can display structures beyond simply being low-rank or sparse. Fully extracting and exploiting hidden structure information in the data is always desirable and favorable. To reveal more underlying effective manifold structure, in this paper, we explicitly model the data relation. Specifically, we propose a structure learning framework that retains the pairwise similarities between the data points. Rather than just trying to reconstruct the original data based on self-expression, we also manage to reconstruct the kernel matrix, which functions as similarity preserving. Consequently, this technique is particularly suitable for the class of learning problems that are sensitive to sample similarity, e.g., clustering and semisupervised classification. To take advantage of representation power of deep neural network, a deep auto-encoder architecture is further designed to implement our model. Extensive experiments on benchmark data sets demonstrate that our proposed framework can consistently and significantly improve performance on both evaluation tasks. We conclude that the quality of structure learning can be enhanced if similarity information is incorporated. • We preserve the pairwise relations during low-dimensional structure learning process. • Combining with auto-encoder, a more discriminative latent representation is achieved. • Evaluations on shallow and deep models with clustering and classification tasks. • Our framework can be readily applied to other self-expression methods and tasks. [ABSTRACT FROM AUTHOR]

To reduce the large integration errors brought by traditional methods in gas flow measurement under complex flow field, artificial neural network, support vector machine and other intelligent algorithms have been put into use. But these intelligent methods consume much time for training and need intensive user intervention for network design. This paper proposes to apply extreme learning machine to multipath ultrasonic flowmeters, which can analytically determine the output weights of networks instead of error backpropagation algorithm and iterative tuning of the parameters, and therefore provide high metering accuracy at extremely fast learning speed as well as require least human intervention. To test its effectiveness under different flow field and sensitivity to complex flow profiles, extreme learning machine is applied to determine the flow rate under two piping configurations, which can produce mild and severe flow disturbances. The determination errors are compared with a traditional integration method on the position of 5D and 10D as well as with the path orientation of 0° and 90°. Then 7 installation angles and 9 installation positions are respectively configured to study the performance and sensitivity of UFMs to installation effects. Finally, a comparison between extreme learning machine and other two intelligent algorithms is made in training and test time, the mean squared error and the maximal metering error under severe flow disturbance. It is found that extreme learning machine has rather high determination accuracy for flow rate at extremely fast learning speed and it is insensitive to the installation effects of ultrasonic gas flowmeter. [ABSTRACT FROM AUTHOR]