Search

Convolutional Neural Networks (CNNs) have become essential in deep learning applications, especially in computer vision, yet their complex internal mechanisms pose significant challenges to interpretability, crucial for ethical applications. Addressing this, our paper explores CNNs by examining their topological changes throughout the learning process, specifically employing persistent homology, a core method within Topological Data Analysis (TDA), to observe the dynamic evolution of their structure. This approach allows us to identify consistent patterns in the topological features of CNN kernels, particularly through shifts in Betti curves, which is a key concept in TDA. Our analysis of these Betti curves, initially focusing on the zeroth and first Betti numbers (respectively referred to as Betti-0 and Betti-1, which denote the number of connected components and loops), reveals insights into the learning dynamics of CNNs and potentially indicates the effectiveness of the learning process. We also discover notable differences in topological structures when CNNs are trained on grayscale versus color datasets, indicating the need for more extensive parameter space adjustments in color image processing. This study not only enhances the understanding of the intricate workings of CNNs but also contributes to bridging the gap between their complex operations and practical, interpretable applications. [ABSTRACT FROM AUTHOR]

Background: The number of patients undergoing proton therapy has increased in recent years. Current treatment planning systems (TPS) calculate dose maps using three-dimensional (3D) maps of relative stopping power (RSP) and mass density. The patient-specific maps of RSP and mass density were obtained by translating the CT number (HU) acquired using single-energy computed tomography (SECT) with appropriate conversions and coefficients. The proton dose calculation uncertainty of this approach is 2.5%-3.5% plus 1 mm margin. SECT is the major clinical modality for proton therapy treatment planning. It would be intriguing to enhance proton dose calculation accuracy using a deep learning (DL) approach centered on SECT. Objectives: The purpose of this work is to develop a deep learning method to generate mass density and relative stopping power (RSP) maps based on clinical single-energy CT (SECT) data for proton dose calculation in proton therapy treatment. Methods: Artificial neural networks (ANN), fully convolutional neural networks (FCNN), and residual neural networks (ResNet) were used to learn the correlation between voxel-specific mass density, RSP, and SECT CT number (HU). A stoichiometric calibration method based on SECT data and an empirical model based on dual-energy CT (DECT) images were chosen as reference models to evaluate the performance of deep learning neural networks. SECT images of a CIRS 062M electron density phantom were used as the training dataset for deep learning models. CIRS anthropomorphic M701 and M702 phantoms were used to test the performance of deep learning models. Results: For M701, the mean absolute percentage errors (MAPE) of the mass density map by FCNN are 0.39%, 0.92%, 0.68%, 0.92%, and 1.57% on the brain, spinal cord, soft tissue, bone, and lung, respectively, whereas with the SECT stoichiometric method, they are 0.99%, 2.34%, 1.87%, 2.90%, and 12.96%. For RSPmaps, the MAPE of FCNN on M701 are 0.85%, 2.32%, 0.75%, 1.22%, and 1.25%, whereas with the SECT reference model, they are 0.95%, 2.61%, 2.08%, 7.74%, and 8.62%. Conclusion: The results show that deep learning neural networks have the potential to generate accurate voxel-specific material property information, which can be used to improve the accuracy of proton dose calculation. Advances in knowledge: Deep learning-based frameworks are proposed to estimate material mass density and RSP from SECT with improved accuracy compared with conventional methods. [ABSTRACT FROM AUTHOR]

Background: Breast cancer is one of the most common malignant tumors in women. A noninvasive ultrasound examination can identify mammary-gland-related diseases and is well tolerated by dense breast, making it a preferred method for breast cancer screening and of significant clinical value. However, the diagnosis of breast nodules or masses via ultrasound is performed by a doctor in real time, which is time-consuming and subjective. Junior doctors are prone to missed diagnoses, especially in remote areas or grass-roots hospitals, due to limited medical resources and other factors, which bring great risks to a patient's health. Therefore, there is an urgent need to develop fast and accurate ultrasound image analysis algorithms to assist diagnoses. Methods: We propose a breast ultrasound image-based assisted-diagnosis method based on convolutional neural networks, which can effectively improve the diagnostic speed and the early screening rate of breast cancer. Our method consists of two stages: tumor recognition and tumor classification. (1) Attention-based semantic segmentation is used to identify the location and size of the tumor; (2) the identified nodules are cropped to construct a training dataset. Then, a convolutional neural network for the diagnosis of benign and malignant breast nodules is trained on this dataset. We collected 2057 images from 1131 patients as the training and validation dataset, and 100 images of the patients with accurate pathological criteria were used as the test dataset. Results: The experimental results based on this dataset show that the MIoU of tumor location recognition is 0.89 and the average accuracy of benign and malignant diagnoses is 97%. The diagnosis performance of the developed diagnostic system is basically consistent with that of senior doctors and is superior to that of junior doctors. In addition, we can provide the doctor with a preliminary diagnosis so that it can be diagnosed quickly. Conclusion: Our proposed method can effectively improve diagnostic speed and the early screening rate of breast cancer. The system provides a valuable aid for the ultrasonic diagnosis of breast cancer. [ABSTRACT FROM AUTHOR]

Objective: For several years, computer technology has been utilized to diagnose lung nodules. When compared to traditional machine learning methods for image processing, deep-learning methods can improve the accuracy of lung nodule diagnosis by avoiding the laborious pre-processing step of the picture (extraction of fake features, etc.). Our goal is to investigate how well deep-learning approaches classify lung nodule malignancy. Method: We evaluated the performance of deep-learning methods on lung nodule malignancy classification via a systematic literature search. We conducted searches for appropriate articles in the PubMed and ISI Web of Science databases and chose those that employed deep learning to classify or predict lung nodule malignancy for our investigation. The figures were plotted, and the data were extracted using SAS version 9.4 and Microsoft Excel 2010, respectively. Results: Sixteen studies that met the criteria were included in this study. The articles classified or predicted pulmonary nodule malignancy using classification and summarization, using convolutional neural network (CNN), autoencoder (AE), and deep belief network (DBN). The AUC of deep-learning models is typically greater than 90% in articles. It demonstrated that deep learning performed well in the diagnosis and forecasting of lung nodules. Conclusion: It is a thorough analysis of the most recent advancements in lung nodule deep-learning technologies. The advancement of image processing techniques, traditional machine learning techniques, deep-learning techniques, and other techniques have all been applied to the technology for pulmonary nodule diagnosis. Although the deep-learning model has demonstrated distinct advantages in the detection of pulmonary nodules, it also carries significant drawbacks that warrant additional research. [ABSTRACT FROM AUTHOR]

Woven fabric composite structures are applied in a wide range of industrial applications. Composite structures are vulnerable to damage from working in complex conditions and environments, which threatens the safety of the in-service structure. Damage detection based on Lamb waves is one of the most promising structural health monitoring (SHM) techniques for composite materials. In this paper, based on guided Lamb waves, a lightweight deep-learning approach is proposed for identifying damaged regions in woven fabric composite structures. The designed deep neural networks are built using group convolution and depthwise separated convolution, which can reduce the parameters considerably. The input of this model is a multi-channel matrix transformed by a one-dimensional guided wave signal. In addition, channel shuffling is introduced to increase the interaction between features, and a multi-head self-attention module is designed to increase the model's global modeling capabilities. The relevant experimental results show that the proposed SHM approach can achieve a recognition accuracy of 100% after only eight epochs of training, and the proposed LCANet has only 4.10% of the parameters of contrastive SHM methods, which further validates the effectiveness and reliability of the proposed method. [ABSTRACT FROM AUTHOR]

With the development of deep learning techniques, convolutional neural networks (CNN) are increasingly being used in image recognition for marine surveys and underwater object classification. Automatic recognition of targets on side-scan sonar (SSS) images using CNN can improve recognition accuracy and efficiency. However, the vast selection of CNN models makes it challenging to select models for target recognition in SSS images. Therefore, this paper aims to compare different CNN models' prediction accuracy and computational performance comprehensively. First, four traditional CNN models were applied to train and predict the same submarine SSS dataset using both the original model and models with transfer learning methods. Then, we examined and studied the prediction accuracy and computation performance of four CNN models. Results showed that transfer learning enhances the accuracy of all CNN models, with lesser improvements for AlexNet and VGG-16 and greater improvements for GoogleNet and ResNet101. GoogleNet has the highest prediction of accuracy (100% in the train dataset and 94.27% in the test dataset) and good computational difficulty. The findings of this work are useful for future model selection in target recognition in SSS images. [ABSTRACT FROM AUTHOR]

Solar radio observation is a method used to study the Sun. It is very important for space weather early warning and solar physics research to automatically classify solar radio spectrums in real time and judge whether there is a solar radio burst. As the number of solar radio burst spectrums is small and uneven, this paper proposes a classification method for solar radio spectrums based on the Swin transformer. First, the method transfers the parameters of the pretrained model to the Swin transformer model. Then, the hidden layer weights of the Swin transformer are frozen, and the fully connected layer of the Swin transformer is trained on the target dataset. Finally, parameter tuning is performed. The experimental results show that the method can achieve a true positive rate of 100%, which is more accurate than previous methods. Moreover, the number of our model parameters is only 20 million, which is 80% lower than that of the traditional VGG16 convolutional neural network with more than 130 million parameters. [ABSTRACT FROM AUTHOR]

Seismic full wave inversion (FWI) is a widely used non-linear seismic imaging method used to reconstruct subsurface velocity images, however it is time consuming, has high computational cost and depend heavily on human interaction. Recently, deep learning has accelerated it's use in several data-driven techniques, however most deep learning techniques suffer from overfitting and stability issues. In this work, we propose an edge computing-based data-driven inversion technique based on supervised deep convolutional neural network to accurately reconstruct the subsurface velocities. Deep learning based data-driven technique depends mostly on bulk data training. In this work, we train our deep convolutional neural network (DCN) (UNet and InversionNet) on the raw seismic data and their corresponding velocity models during the training phase to learn the non-linear mapping between the seismic data and velocity models. The trained network is then used to estimate the velocity models from new input seismic data during the prediction phase. The prediction phase is performed on a resource-constrained edge device such as Raspberry Pi. Raspberry Pi provides real-time and on-device computational power to execute the inference process. In addition, we demonstrate robustness of our models to perform inversion in the presence on noise by performing both noise-aware and no-noise training and feeding the resulting trained models with noise at different signal-to-noise (SNR) ratio values. We make great efforts to achieve very feasible inference times on the Raspberry Pi for both models. Specifically, the inference times per prediction for UNet and InversionNet models on Raspberry Pi were 22 and 4 s respectively whilst inference times for both models on the GPU were 2 and 18 s which are very comparable. Finally, we have designed a user-friendly interactive graphical user interface (GUI) to automate the model execution and inversion process on the Raspberry Pi. [ABSTRACT FROM AUTHOR]

The emergence of multimodal medical imaging technology greatly increases the accuracy of clinical diagnosis and etiological analysis. Nevertheless, each medical imaging modal unavoidably has its own limitations, so the fusion of multimodal medical images may become an effective solution. In this paper, a novel fusion method on the multimodal medical images exploiting convolutional neural network (CNN) and extreme learning machine (ELM) is proposed. As a typical representative in deep learning, CNN has been gaining more and more popularity in the field of image processing. However, CNN often suffers from several drawbacks, such as high computational costs and intensive human interventions. To this end, the model of convolutional extreme learning machine (CELM) is constructed by incorporating ELM into the traditional CNN model. CELM serves as an important tool to extract and capture the features of the source images from a variety of different angles. The final fused image can be obtained by integrating the significant features together. Experimental results indicate that, the proposed method is not only helpful to enhance the accuracy of the lesion detection and localization, but also superior to the current state-of-the-art ones in terms of both subjective visual performance and objective criteria. [ABSTRACT FROM AUTHOR]

With the development of convolutional neural networks, the effect of pedestrian detection has been greatly improved by deep learning models. However, the presence of pseudo pedestrians will lead to accuracy reduction in pedestrian detection. To solve the problem that the existing pedestrian detection algorithms cannot distinguish pseudo pedestrians from real pedestrians, a real and pseudo pedestrian detection method with CA-YOLOv5s based on stereo image fusion is proposed in this paper. Firstly, the two-view images of the pedestrian are captured by a binocular stereo camera. Then, a proposed CA-YOLOv5s pedestrian detection algorithm is used for the left-view and right-view images, respectively, to detect the respective pedestrian regions. Afterwards, the detected left-view and right-view pedestrian regions are matched to obtain the feature point set, and the 3D spatial coordinates of the feature point set are calculated with Zhengyou Zhang's calibration method. Finally, the RANSAC plane-fitting algorithm is adopted to extract the 3D features of the feature point set, and the real and pseudo pedestrian detection is achieved by the trained SVM. The proposed real and pseudo pedestrian detection method with CA-YOLOv5s based on stereo image fusion effectively solves the pseudo pedestrian detection problem and efficiently improves the accuracy. Experimental results also show that for the dataset with real and pseudo pedestrians, the proposed method significantly outperforms other existing pedestrian detection algorithms in terms of accuracy and precision. [ABSTRACT FROM AUTHOR]

Emotion is the human brain reacting to objective things. In real life, human emotions are complex and changeable, so research into emotion recognition is of great significance in real life applications. Recently, many deep learning and machine learning methods have been widely applied in emotion recognition based on EEG signals. However, the traditional machine learning method has a major disadvantage in that the feature extraction process is usually cumbersome, which relies heavily on human experts. Then, end-to-end deep learning methods emerged as an effective method to address this disadvantage with the help of raw signal features and time-frequency spectrums. Here, we investigated the application of several deep learning models to the research field of EEG-based emotion recognition, including deep neural networks (DNN), convolutional neural networks (CNN), long short-term memory (LSTM), and a hybrid model of CNN and LSTM (CNN-LSTM). The experiments were carried on the well-known DEAP dataset. Experimental results show that the CNN and CNN-LSTM models had high classification performance in EEG-based emotion recognition, and their accurate extraction rate of RAW data reached 90.12 and 94.17%, respectively. The performance of the DNN model was not as accurate as other models, but the training speed was fast. The LSTM model was not as stable as the CNN and CNN-LSTM models. Moreover, with the same number of parameters, the training speed of the LSTM was much slower and it was difficult to achieve convergence. Additional parameter comparison experiments with other models, including epoch, learning rate, and dropout probability, were also conducted in the paper. Comparison results prove that the DNN model converged to optimal with fewer epochs and a higher learning rate. In contrast, the CNN model needed more epochs to learn. As for dropout probability, reducing the parameters by ~50% each time was appropriate. [ABSTRACT FROM AUTHOR]

Accurate localization of surgical instruments is of utmost importance for precise robot-assisted surgeries. With the development of artificial intelligence, deep convolutional neural networks (DCNNs) have been widely employed for automatic image segmentation, owing to their strong ability to generate contextual features, especially in the encoder–decoder framework. However, existing segmentation networks lack the feature capturing capability on micro objects and have shortcomings in processing local semantic features. These limitations can affect the precise segmentation of surgical instruments. In response to these issues, this paper proposes a multi-scale attention fusion network called MAF-Net, which comprises residual dense module, a multi-scale atrous convolution (MSAC) module, and an attention fusion module (AFM). To improve the processing ability of local contextual features, we propose replacing skip connections with residual dense module to acquire stronger contexts. Furthermore, a MSAC module is proposed for local feature enhancement, thereby enhancing attention on multi-scale features. In addition, an AFM module is introduced to integrate multi-scale information by cross-scale feature fusion. Experimental results, using two public datasets, Endovis2017 and kvasir-instrument, demonstrate that the proposed network has the ability to achieve precise surgical instrument segmentation and outperforms related advanced methods. • A novel multi-scale attention fusion network is proposed for accurate and automatic segmentation of surgical instruments. • A residual dense module is proposed to improve the process ability of local contextual features. • An attention fusion module is proposed to integrate multi-scale information by cross-scale feature fusion. [ABSTRACT FROM AUTHOR]

Defect detection is a very important link for much manufacturing and processing applications which could be used for quality control and precise maintenance decision. However, faced with the weak-texture and low-contrast industrial environment, high-precision defect detection still faces a certain challenge due to diverse and complex of defects. Meanwhile, due to a minimal portion image pixels of defects, the pixel-level defect detection task is always against class-unbalance issue which also will affect the detection performance. Recently, with the strong automatic feature representation ability, deep learning has shown an excellent detection performance on defect identification and location. Nevertheless, it still has some demerits, such as insufficient processing of feature maps, lack of temporal modeling information, etc. To address these issues, on the basis of the encoder–decoder architecture, a pixel-level deep segmentation network is proposed for automatic defect detection to construct an end-to-end defect segmentation model. To realize effective feature representation, a residual attention network is proposed to construct the backbone network, which could also make the segmentation network better emphasize target regions. Meanwhile, to improve the network propagation ability of subtle context features, a bidirectional convolutional long short-term memory (ConvLSTM) block is introduced to optimize the skip connections to learn long-range spatial contexts. Besides, a weighted loss function is proposed for model training to address the class-unbalance issue. Combined with multiple public data sets, through qualitative and quantitative analysis, experimental results demonstrate that the proposed defect segmentation network achieves a better performance compared to other state-of-the-art segmentation methods. [ABSTRACT FROM AUTHOR]

Retinal fundus vessels contain rich geometric features, including both thick and thin vessels, which is particularly important for accurate clinical diagnosis of cardiovascular diseases. Nowadays, deep convolution neural networks (DCNNs) have shown remarkable performance on fundus vessel segmentation combined with effective contextual feature expression ability, especially U-Net network and its variants. However, due to multiple convolution and pooling operations, existing methods using DCNNs will cause the information loss of details or micro objects, which impacts the accuracy of thin vascular detection and cause poor final segmentation performance. To address this problem, a multi-task symmetric network, called GDF-Net, is proposed for accurate retinal fundus image segmentation, which is composed by two symmetrical segmentation networks (global segmentation network branch and detail enhancement network branch) and a fusion network branch. To address the information loss issue, two symmetrical segmentation networks are proposed to extract global contextual features and detail features respectively. To strength both advantages of two symmetric segmentation networks, this paper presented a fusion network to perform feature integration to improve the segmentation accuracy for retinal fundus vessels. To better validate the usefulness and excellence of the GDF-Net, experiments demonstrate that the GDF-Net has achieved a competitive performance on retinal vessel segmentation. • A multi-task symmetric network is proposed for fundus vessels segmentation. • Two symmetric networks are proposed to extract global and detail features respectively. • An attention fusion network is proposed for merging multi-source information. [ABSTRACT FROM AUTHOR]

For the clinical diagnosis, it is essential to obtain accurate morphology data of retinal blood vessels from patients, and the morphology of retinal blood vessels can well help doctors to judge the patient's condition and give targeted therapeutic measures. Conventional manual retinal blood vessel segmentation by the doctors from the fundus images is time-consuming and laborious, while it also requires the rich doctor's expertise. With the strong context feature expression ability of deep convolutional neural networks (DCNN), it has shown a promising performance on retinal blood vessel segmentation, specially U-shape network (U-Net) and its variant. However, due to the information loss issue caused by multiple pooling operations and insufficient process issue of local context features by skip connections, most of segmentation methods still exist a certain shortcoming on accurate fine vessel detection. To address this issue, based on the encoder–decoder framework, a novel retinal vessel segmentation network, called ResDO-UNet, is proposed to provide an automatic and end-to-end detection scheme from fundus images. To enhance feature extraction capabilities, combined with depth-wise over-parameterized convolutional layer (DO-conv), a residual DO-conv (ResDO-conv) network is proposed to act as the backbone network to acquire strong context features. In addition, to reduce the effect of information loss caused by multiple pooling operations, taking advantages of max pooling and average pooling layers, a pooling fusion block (PFB) is proposed to realize nonlinear fusion pooling. Meanwhile, faced with insufficient process of local context features by skip connections, an attention fusion block (AFB) is proposed to realize effective multi-scale feature expression. Combined with the three public available data sets on retinal vessel segmentation, including DRIVE, STARE and CHASE_DB1, the proposed segmentation network could reach a state-of-the-art detection performance compared to other related advanced work. • A novel residual DO-conv network is proposed for automatic and accurate retinal vessel segmentation. • To address the information loss issue, a pooling fusion block is proposed to realize nonlinear fusion pooling. • An attention fusion block is proposed to realize effective multi-scale feature expression. • Proposed model achieves a competitive performance on multiple public benchmark image sets. [ABSTRACT FROM AUTHOR]