Search

• An advanced detection neural network, the improved Faster R-CNN, is presented to realize the localization and classification of axillary lymph nodes (ALNs) simultaneously, which not only classifies lymph node metastasis, but also provides doctors with the position of ALNs in the image for reference. • As ALN usually appears on the left and right sides of CECT images, an attention model, which has the ability of extracting spatial information, was added into the network, aiming to further improve the performance of the algorithm in positioning ALN. • A novel model constructed by channel fusion and bottle neck architecture is proposed and placed before the flatten operation to better fuse and transmit spatial information with less computational cost. Background and objective: Preoperative imaging diagnosis of axillary lymph node (ALN) metastasis is particularly important for breast cancer patients. This paper focuses on developing non-invasive and automatic schemes for accurate localization and classification (metastasis prediction) of ALN via contrast-enhanced computed tomography (CECT) image and deep learning models. Methods: Based on a two-stage strategy, a novel detection neural network is proposed, where the convolutional block attention module is utilized to extract spacial information and the bottleneck feature fusion module is designed for feature fusion in different scales. Results: Owing to the two embedded modules, the proposed convolutional neural network (CNN) model outperforms Faster R-CNN, YOLOv3, and EfficientDet in the sense that the achieved mAP is 0.454, higher than 0.247, 0.335, and 0.329, respectively. In particular, considering the function of classification only, the proposed model reaches the best performance on most indices (accuracy of 0.968, positive predictive value of 0.972, negative predictive value of 0.966, specificity of 0.983), compared with the methods that have been frequently adopted to predict ALN. In addition, the proposed CNN model has the function of locating ALN, which is lacking in existing models. Conclusions: In this paper, a supervised deep learning method is proposed to detect ALN in CECT images. The positive effect of new added modules are verified, and the benefits of spatial information in ALN detection are confirmed. Further, the two subtasks called localization and classification are evaluated separately, where the proposed model achieves the best performance on most indices. The source code mentioned in this article will be released later. [ABSTRACT FROM AUTHOR]

• Autonomous driving, virtual reality and all kinds of robots integrated into our life rely on facial expression recognition technology. • Facial expression recognition and computer vision is based on deep learning technology and convolutional neural network. • Whether it is two-stage target detection or single-stage target detection, performance of algorithm is measured by detection speed and accuracy. • Large variety of training data with accurate expression tags can fundamentally improve expression recognition rate. Facial expression recognition technology will play an increasingly important role in our daily life. Autonomous driving, virtual reality and all kinds of robots integrated into our life depend on the development of facial expression recognition technology. Many tasks in the field of computer vision are based on deep learning technology and convolutional neural network. The paper proposes an occluded expression recognition model based on the generated countermeasure network. The model is divided into two modules, namely, occluded face image restoration and face recognition. Firstly, this paper summarizes the research status of deep facial expression recognition methods in recent ten years and the development of related facial expression database. Then, the current facial expression recognition methods based on deep learning are divided into two categories: Static facial expression recognition and dynamic facial expression recognition. The two methodswill be introduced and summarized respectively. Aiming at the advanced deep expression recognition algorithms in the field, the performance of these algorithms on common expression databases is compared, and the strengths and weaknesses of these algorithms are analyzed in detail. As the task of facial expression recognition is gradually transferred from the controlled laboratory environment to the challenging real-world environment, with the rapid development of deep learning technology, deep neural network can learn discriminative features, and is gradually applied to automatic facial expression recognition task. The current deep facial expression recognition system is committed to solve the following two problems: (1) Overfitting due to lack of sufficient training data; (2) In the real world environment, other variables that have nothing to do with expression bring interference problems. From the perspective of algorithm, combining other expression models, such as facial action unit model and pleasure arousal dimension model, as well as other multimodal models, such as audio mode, 3D face depth information and human physiological information, can make expression recognition more practical. [ABSTRACT FROM AUTHOR]

• An improved deep residual convolutional neural network is proposed in this paper for arrhythmia classification. It uses raw ECG data to perform an efficient arrhythmia classification because it does not require extensive data preprocessing, such as R-peak localization and feature extraction. The experiment in this paper also shows that the network model trained by the training set obtained by using the overlapping segmentation method has better classification performance. In addition, the classification performance of the network model trained by focal loss as the loss function is further improved. The effectiveness of the proposed method is illustrated by high classification performance under inter-patient paradigms, even though single-lead raw ECG data is used only. Background and Objective: Early detection of arrhythmias has become critical due to the increased mortality from cardiovascular disease, and ECG is an effective tool for diagnosing cardiovascular disease and detecting arrhythmias. Classification based on ECG signal segments is more suitable for clinical application. Methods: An improved deep residual convolutional neural network is proposed to classify arrhythmias automatically. Firstly, the overlapping segmentation method is used to segment the ECG signals in the MIT-BIH database into segments of 5 seconds in length to overcome the imbalance between classes, and these segments of the ECG signals are re-labeled. Then the discrete wavelet transform (DWT) is used to denoise these segments and the improved deep residual convolutional neural network is used for arrhythmia classification. In addition, the focal loss function is used to overcome the imbalanced classification difficulty between classes. Results: The proposed method gives 94.54% sensitivity, 93.33% positive predictivity, and 80.80% specificity for normal segments. For the supraventricular ectopic segment, the proposed method gives 35.22% sensitivity, 65.88% positive predictivity, and 98.83% specificity. For the ventricular ectopic segment, the proposed method gives 88.35% sensitivity, 79.86% positive predictivity, and 94.92% specificity. Conclusion: The results of this study indicate that the proposed improved deep residual convolutional neural network model trained by the training set obtained by using the overlapping segmentation method is comparable to a classical method and three state-of-art methods. In addition, the classification performance of the network model trained by focal loss as the loss function is further improved. [ABSTRACT FROM AUTHOR]

Epilepsy is a brain disorder consisting of abnormal electrical discharges of neurons resulting in epileptic seizures. The nature and spatial distribution of these electrical signals make epilepsy a field for the analysis of brain connectivity using artificial intelligence and network analysis techniques since their study requires large amounts of data over large spatial and temporal scales. For example, to discriminate states that would otherwise be indistinguishable from the human eye. This paper aims to identify the different brain states that appear concerning the intriguing seizure type of epileptic spasms. Once these states have been differentiated, an attempt is made to understand their corresponding brain activity. The representation of brain connectivity can be done by graphing the topology and intensity of brain activations. Graph images from different instants within and outside the actual seizure are used as input to a deep learning model for classification purposes. This work uses convolutional neural networks to discriminate the different states of the epileptic brain based on the appearance of these graphs at different times. Next, we apply several graph metrics as an aid to interpret what happens in the brain regions during and around the seizure. Results show that the model consistently finds distinctive brain states in children with epilepsy with focal onset epileptic spasms that are indistinguishable under the expert visual inspection of EEG traces. Furthermore, differences are found in brain connectivity and network measures in each of the different states. Computer-assisted discrimination using this model can detect subtle differences in the various brain states of children with epileptic spasms. The research reveals previously undisclosed information regarding brain connectivity and networks, allowing for a better understanding of the pathophysiology and evolving characteristics of this particular seizure type. From our data, we speculate that the prefrontal, premotor, and motor cortices could be more involved in a hypersynchronized state occurring in the few seconds immediately preceding the visually evident EEG and clinical ictal features of the first spasm in a cluster. On the other hand, a disconnection in centro-parietal areas seems a relevant feature in the predisposition and repetitive generation of epileptic spasms within clusters. [ABSTRACT FROM AUTHOR]

• P300 ERP detection can be used to interpret EEG signals. • Deep learning is utilized for the detection of the P300 component. • A new CNN is designed as TGT-MHOG-CNN using new preprocessing techniques. • TGT-MHOG-CNN is proven to be superior in accuracy, timing, and complexity. Incorporating the time-frequency localization properties of Gabor transform (GT), the complexity understandings of convolutional neural network (CNN), and histogram of oriented gradients (HOG) efficacy in distinguishing positive peaks can exhibit their characteristics to reveal an effective solution in the detection of P300 evoked related potential (ERP). By applying a drastic number of convolutional layers, the majority of deep networks elicit sufficient properties for the output determination, leading to gigantic and time-consuming structures. In this paper, we propose a novel deep learning framework by the combination of tuned GT, and modified HOG with the CNN as "TGT-MHOG-CNN" for detection of P300 ERP in EEG signal. In the proposed method, GT is tuned based on triangular function for EEG signals, and then spectrograms including time-frequency information are captured. The function's parameters are justified to differentiate the signals with the P300 component. Furthermore, HOG is modified (MHOG) for the 2-D EEG signal, and consequently, gradients patterns are extracted for the target potentials. MHOG is potent in distinguishing the positive peak in the general waveform; however, GT unravels time-frequency information, which is ignored in the gradient histogram. These outputs of GT and MHOG do not share the same nature in the images nor overlap. Therefore, more extensive information is reached without redundancy or excessive information by fusing them. Combining GT and MHOG provides different patterns which benefit CNN for more precise detection. Consequently, TGT-MHOG-CNN ends in a more straightforward structure than other networks, and therefore, the whole performance is acceptable with faster rates and very high accuracy. BCI Competition II and III datasets are used to evaluate the performance of the proposed method. These datasets include a complete record for P300 ERP with BCI2000 using a paradigm, and it has numerous noises, including power and muscle-based noises. The objective is to predict the correct character in each provided character selection epochs. Compared to state-of-the-art methods, simulation results indicate striking abilities of the proposed framework for P300 ERP detection. Our best record reached the P300 ERP classification rates of over 98.7% accuracy and 98.7% precision for BCI Competition II and 99% accuracy and 100% precision for BCI Competition III datasets, with superiority in execution time for the mentioned datasets. [ABSTRACT FROM AUTHOR]

• The collected tongue images are sorted and marked to form a relevant data set. • The convolutional neural network model based on deep learning improves the ability to extract tongue features. • The ResNet34 algorithm architecture in the model improves the accuracy, sensitivity, and specificity of the model when processing the dataset. • The health detector for tongue image recognition designed by using the research model can effectively improve the accuracy of judging the user's health status. Tongue diagnosis is one of the characteristics of traditional Chinese medicine (TCM), but traditional tongue diagnosis is affected by many factors, and its differential diagnosis results are not widely recognized. The appearance of tongue diagnosis instruments is the product of the modernization of tongue diagnosis, and it has standard and objective advantages in clinical practice. In this study, based on standard tongue images, a tongue image dataset and detection model were constructed. And based on the deep learning convolutional neural network (CNN) algorithm and visual question answering technology, a human-computer interaction intelligent health detector for tongue image recognition is constructed. In this research, 1420 tongue images were collected. After screening, experts judged them, and annotated the tongue images to form tongue image datasets. Then the artificial intelligence network framework based on deep learning convolutional neural network (CNN), that is, ResNet34, is applied to this dataset to automatically extract image features and realize tongue images classification. Finally, the VGG16 network framework is applied to the dataset to compare the classification model and compare with the classification effect. In this paper, relevant datasets were formed by collating the tongue images collected by annotation, which verified that the ResNet34 architecture could better perform the task of tooth mark and tongue feature recognition. Compared with similar learning tasks in existing studies, the accuracy of the teeth-printed tongue recognition model proposed in this study is more than 10% higher, which indicates that the CNN algorithm can distinguish teeth-printed tongue more accurately and effectively. At the same time, using datasets and models combined with visual question and answer technology, an AI health detector for TCM tongue image identification is designed, which can make health assessments and give suggestions to users. This study adopts a convolutional neural network model based on deep learning, which can reduce the extraction of tongue features more quickly and conveniently. At the same time, the model architecture has excellent performance and strong generalization ability and is more accurate in judging users' health status. [ABSTRACT FROM AUTHOR]

• An improved convolutional module MCB is proposed. • A hybrid down-sampling module HDSB is proposed to reduce the feature loss. • The Skip Connection+ structure is proposed to capture more context information. Background and objective: Medical image segmentation plays an important role in clinic. Recently, with the development of deep learning, many convolutional neural network (CNN)-based medical image segmentation algorithms have been proposed. Among them, U-Net is one of the most famous networks. However, the standard convolutional layers used by U-Net limit its capability to capture abundant features. Additionally, the consecutive maximum pooling operations in U-Net cause certain features to be lost. This paper aims to improve the feature extraction capability of U-Net and reduce the feature loss during the segmentation process. Meanwhile, the paper also focuses on improving the versatility of the proposed segmentation model. Methods: Firstly, in order to enable the model to capture richer features, we have proposed a novel multiscale convolutional block (MCB). MCB adopts a wider and deeper structure, which can be applied to different types of segmentation tasks. Secondly, a hybrid down-sampling block (HDSB) has been proposed to reduce the feature loss via replacing the maximum pooling layer. Thirdly, we have proposed a context module (CIF) based on atrous convolution and SKNet to extract sufficient context information. Finally, we combined the CIF module with Skip Connection of U-Net, and further proposed the Skip Connection+ structure. Results: We name the proposed network MHSU-Net. MHSU-Net has been evaluated on three different datasets, including lung, cell contour, and pancreas. Experimental results demonstrate that MHSU-Net outperforms U-Net and other state-of-the-art models under various evaluation metrics, and owns greater potential in clinical applications. Conclusions: The proposed modules can greatly improve the feature extraction capability of the segmentation model and effectively reduce the feature loss during the segmentation process. MHSU-Net can also be applied to different types of medical image segmentation tasks. [ABSTRACT FROM AUTHOR]

• The diagnostic effect of 64-slice spiral CT and MRI high-resolution images based on deep CNN in lung cancer is studied. • 74 patients with highly suspected lung cancer were selected as the research objects. • The enhanced 64-slice spiral CT and MRI were used to detect and diagnose respectively, and the images and accuracy of CT diagnosis and MRI diagnosis were retrospectively analyzed. • The accuracy of CT diagnosis is 94.6% (70/74), and the accuracy of MRI diagnosis is 89.2% (66/74). • Enhanced CT and MRI examinations based on CNN to improve image clarity have high application value in the diagnosis of lung cancer. To study the diagnostic effect of 64-slice spiral CT and MRI high-resolution images based on deep convolutional neural networks(CNN) in lung cancer. In this paper, we Select 74 patients with highly suspected lung cancer who were treated in our hospital from January 2017 to January 2021 as the research objects. The enhanced 64-slice spiral CT and MRI were used to detect and diagnose respectively, and the images and accuracy of CT diagnosis and MRI diagnosis were retrospectively analyzed. The accuracy of CT diagnosis is 94.6% (70/74), and the accuracy of MRI diagnosis is 89.2% (66/74). CT examination has the advantages of non-invasive, convenient operation and fast examination. MRI is showing there are advantages in the relationship between the chest wall and the mediastinum, and the relationship between the lesion and the large blood vessels. Enhanced CT and MRI examinations based on convolutional neural networks(CNN) to improve image clarity have high application value in the diagnosis of lung cancer patients, but the focus of performance is different. [ABSTRACT FROM AUTHOR]

• Convolutional Neural Networks can accurately detect trabeculae in the left ventricle. • The Deep Learning solution outperforms the traditional trabecular segmentation tool. • Produced segmentations enable the automatic assessment of the level of trabeculation. • Automatic and accurate diagnosis based on the output of the network is possible. Background and Objective: Left ventricular non-compaction (LVNC) is an uncommon cardiomyopathy characterised by a thick and spongy left ventricle wall caused by the high presence of trabeculae (hyper-trabeculation). Recently, the percentage of the trabecular volume to the total volume of the external wall of the left ventricle (V T %) has been proposed to diagnose this illness. Methods: This paper presents the use of a deep learning-based method to measure the (V T %) value and diagnose this rare cardiomyopathy. The population used in this research was composed of 277 patients suffering from hypertrophic cardiomyopathy. 134 patients only suffered hypertrophic cardiomyopathy, and 143 also suffered left ventricular non-compaction. Our deep learning solution is based on a 2D U-Net. This artificial neural network (ANN) was trained on short-axis magnetic resonance imaging to segment the left ventricle's internal cavity, external wall, and trabecular tissue. 5-fold cross-validation was performed to ensure the robustness of the results. The Dice coefficient of the three classes was computed as a measure of the precision of the segmentation. Based on this segmentation, the percentage of the trabecular volume (VT%) was computed. Two specialist cardiologists rated the segmentation produced by the neural network for 25 patients to evaluate the clinical validity of the outputs. The computed VT% was used to automatically diagnose the 277 patients depending on whether or not a given threshold was exceeded. A receiver operating characteristic analysis was also performed. Results: According to the cross-validation results, the average and standard deviation of the Dice coefficient for the internal cavity, external wall, and trabeculae were 0.96 ± 0.00 , 0.89 ± 0.00 , and 0.84 ± 0.00 , respectively. The cardiologists rated 99.5 % of the evaluated segmentations as clinically valid for diagnosis, outperforming existing automatic traditional tools. The area under the ROC curve was 0.94 (95 % confidence interval, 0.91–0.96). The accuracy, sensitivity, and specificity values of diagnosis using a threshold of 25 % were 0.87, 0.93, and 0.80, respectively. Conclusions: The U-Net neural network can achieve excellent results in the delineation of different cardiac structures of short-axis cardiac MRI. The high-quality segmentation allows for the correct measurement of left ventricular hyper-trabeculation and a definitive diagnosis of LVNC illness. Using this kind of solution could lead to more objective and faster analysis, reducing human error and time spent by cardiologists. [ABSTRACT FROM AUTHOR]

• CNN is used to segment multiple bone joints to assist medical diagnosis. • Image segmentation using local perception and features. • A loss function is added to adjust the position of the candidate area. • Fine segmentation of the edges of the bone areais is performed to prevent confusion with fat. Magnetic resonance imaging (MRI) is gradually replacing computed tomography (CT) in the examination of bones and joints. The accurate and automatic segmentation of the bone structure in the MRI of the shoulder joint is essential for the measurement and diagnosis of bone injuries and diseases. The existing bone segmentation algorithms cannot achieve automatic segmentation without any prior knowledge, and their versatility and accuracy are relatively low. For this reason, an automatic segmentation algorithm based on the combination of image blocks and convolutional neural networks is proposed. First, we establish 4 segmentation models, including 3 U-Net-based bone segmentation models (humeral segmentation model, joint bone segmentation model, humeral head and articular bone segmentation model as a whole) and a block-based Alex Net segmentation model; Then we use 4 segmentation models to obtain the candidate bone area, and accurately detect the location area of the humerus and joint bone by voting. Finally, the Alex Net segmentation model is further used in the detected bone area to segment the bone edge with the accuracy of the pixel level. The experimental data is obtained from 8 groups of patients in the orthopedics department of our hospital. Each group of scan sequence includes about 100 images, which have been segmented and labeled. Five groups of patients were used for training and five-fold cross-validation, and three groups of patients were used to test the actual segmentation effect. The average accuracy of Dice Coefficient, Positive Predicted Value (PPV) and Sensitivity reached 0.91 ± 0.02, respectively. 0.95 ± 0.03 and 0.95 ± 0.02. The method in this paper is for a small sample of patient data sets, and only through deep learning on 2D medical images, very accurate shoulder joint segmentation results can be obtained, provide clinical diagnostic guidance to orthopedics. At the same time, the proposed algorithm framework has a certain versatility and is suitable for the precise segmentation of specific organs and tissues in MRI based on a small sample data. [ABSTRACT FROM AUTHOR]

• S-transform spectrogram is introduced to boost the ECG quality assessment performance for the first time. • A double-input CNN is designed to integrate the hand-crafted statistical features and deep-learned S-Transform spectrogram features, improving both the sensitivity and specificity in evaluating ECG quality. • Enhance the model generalization ability by applying a novel online augmentation scheme during the training stage. • A log-odds analysis scheme combining with a gradient-based method is proposed to localize the abnormal ECG signal in time, frequency, and spatial domains. Background and Objective Electrocardiogram (ECG) quality assessment is significant for automatic diagnosis of cardiovascular disease and reducing the massive workload of reviewing continuous ECGs. Hence, how to design an appropriate algorithm for objectively evaluating the multi-lead ECG recordings is particularly important. Despite the deep learning methods performing well in many fields, as a data-driven method, it may not be entirely suitable for ECG analysis due to the difficulty in obtaining sufficient data and the low signal-to-noise ratio of ECG recordings. In this study, with the aim of providing an accurate and automatic ECG quality assessment scheme, we propose an innovative ECG quality assessment algorithm based on hand-crafted statistical features and deep-learned spectral features. Methods In this paper, a novel approach, combining the deep-learned Stockwell transform (S-Transform) spectrogram features and hand-crafted statistical features, is proposed for ECG quality assessment. Firstly, a double-input convolutional neural network (CNN) is established. Then, the S-Transform with a novel online augmentation scheme is performed on the multi-lead raw ECG signal received from one input layer to obtain proper time-frequency representation. After that, the CNN with three convolutional layers is employed to extract robust deep-learned features automatically. Simultaneously, the hand-crafted statistical features, including lead-fall, baseline drift, and R peak features, are calculated and fed into another input layer for feature fusion training. Finally, the deep-learned and hand-crafted features are concatenated and further fused by a fully connected layer for quality classification. Furthermore, a log-odds analysis scheme combining with a gradient-based method can localize the abnormal zone in time, frequency, and spatial domains. Results and Conclusion Our proposed method is evaluated on a publicly available database with 10-fold cross-validation. The experimental results demonstrate that the proposed assessment algorithm reached a mean accuracy of 93.09%, a mean F1-score of 0.8472, and a sensitivity of 0.9767. Moreover, comprehensive experiments indicate that the fusion of CNN features and statistical features has complementary advantages and ideal interpretability, achieving end-to-end multi-lead ECG assessment with satisfying performance. [ABSTRACT FROM AUTHOR]

• Here proposed a refined 1D-CNN model for six types of Cardiovascular diseases (CVDs) prediction, in which standard 12-lead ECG signals are the inputs. • The model achieved maximum accuracy of 99.84%, 99.80%, and 99.59% with three-seconds, two-seconds and one-second ECG segments, respectively. • As an alternative to any complex preprocessing, durations of raw ECG signals have been considered as input with simple min-max normalization. Background and Objective : Automatic screening tools can be applied to detect cardiovascular diseases (CVDs), which are the leading cause of death worldwide. As an effective and non-invasive method, electrocardiogram (ECG) based approaches are widely used to identify CVDs. Hence, this paper proposes a deep convolutional neural network (CNN) to classify five CVDs using standard 12-lead ECG signals. Methods : The Physiobank (PTB) ECG database is used in this study. Firstly, ECG signals are segmented into different intervals (one-second, two-seconds and three-seconds), without any wave detection, and three datasets are obtained. Secondly, as an alternative to any complex preprocessing, durations of raw ECG signals have been considered as input with simple min-max normalization. Lastly, a ten-fold cross-validation method is employed for one-second ECG signals and also tested on other two datasets (two-seconds and three-seconds). Results : Comparing to the competing approaches, the proposed CNN acquires the highest performance, having an accuracy, sensitivity, and specificity of 99.59%, 99.04%, and 99.87%, respectively, with one-second ECG signals. The overall accuracy, sensitivity, and specificity obtained are 99.80%, 99.48%, and 99.93%, respectively, using two-seconds of signals with pre-trained proposed models. The accuracy, sensitivity, and specificity of segmented ECG tested by three-seconds signals are 99.84%, 99.52%, and 99.95%, respectively. Conclusion : The results of this study indicate that the proposed system accomplishes high performance and keeps the characterizations in brief with flexibility at the same time, which means that it has the potential for implementation in a practical, real-time medical environment. [ABSTRACT FROM AUTHOR]

• Continuous ECG monitoring requires robust R-wave detection. • Convolutional neural networks (CNN) outperform other techniques. • Adaptive cross data and cross lead training further increases performance. • Evaluation is based on 4 ECG datasets with 4.5 million annotated beats. Automated R-wave detection plays a vital role in electrocardiography (ECG) and ECG-based computer-aided diagnosis. Recently, a multi-level one-dimensional (1D) deep learning approach was presented that shows good performance as compared to traditional methods. In this paper, we present several improvements of the multi-level 1D convolutional neural network (CNN)-based deep learning approach using: (i) adaptive deep learning, (ii) cross-database training, and (iii) cross-lead training. For this, we consider ECG signals from four publicly available databases: MIT-BIH, INCART, TELE, and SDDB, having 109,404, 175,660, 6,708, and 1,684,447 annotated beats, respectively. Except for TELE, all databases provide at least two-lead recordings. To evaluate the improvements, experiments are performed with adaptive k-times cross-trained databases validation scheme (k = 5). The hypothesis tested are: (i) the improvements outperform the state-of-the-art, (ii) cross-database training and adaptive deep learning contribute, and (iii) additional databases or cross-lead training further improves the results. Our proposed approach outperforms the state-of-the-art. In terms of F-measure, F = 99.75% and F = 95.25% is obtained for the MIT-BIH and TELE databases, respectively. Further, cross-database training (F = 98.02%) is found to be more effective than training on individual databases (F = 97.33%). The performance of our approach further improves when additional databases and different leads are used for training. Existing state-of-the-art methods perform low on noisy and pathological signals. Adaptive cross-data training identifies the optimal model. Using multiple datasets and leads allows analyzing noisy, pathological and mobile-recorded long-term ECG signals without ground truths. These conclusions are based on the comprehensive evaluation of four different databases, and in total, about 4.5 million annotated beats. [ABSTRACT FROM AUTHOR]

• MRI is important for the measurement and diagnosis of bone injury and disease. • Automatic segmentation method combining PCNN and FCN is proposed. • Precise shoulder segmentation results can be obtained by performing deep learning. • AlexNet can segment orthopedic MRI with higher accuracy. Magnetic resonance imaging (MRI) has been known to replace computed tomography (CT) for bone and skeletal joint examination. The accurate automatic segmentation of bone structure in shoulder MRI is important for the measurement and diagnosis of bone injury and disease. Existing bone segmentation algorithms cannot achieve automatic segmentation without any prior knowledge, and their versatility and accuracy are relatively low. Therefore, an automatic segmentation combining pulse coupled neural network (PCNN) and full convolutional neural networks (FCN) is proposed. By constructing the block-based AlexNet segmentation model and U-Net-based bone segmentation module, we implemented the humeral segmentation model, articular bone segmentation model, humeral head and articular bone segmentation model synthesis model. We use this four kinds of segmentation models to obtain candidate bone regions, and accurately detect the positions of humerus and articular bone by voting. Finally, we perform an AlexNet segmentation model in the detected bone area in one step to segment accuracy at the pixel level. The experimental data came from 8 groups of patients in Shengjing Hospital affiliated to China Medical University. The scanning volume of each group is approximately 100 images. Five fold cross-validations and for training were recorded, and five sets of data were carefully separated. After using our technique in the three groups of patients tested, the positive predictive value of dice coefficient (PPV) and the average accuracy of sensitivity were very significant, which reached 0.96±0.02, 0.97±0.02 and 0.94±0.03, respectively. The method used in the experiment in this paper is based on a small amount of patient sample data. The deep learning required for the experiment needs to be performed through 2D medical images. The shoulder segmentation data obtained in this way can be very accurate. [ABSTRACT FROM AUTHOR]

• A two-stage vessel extraction framework is designed to learn well-defined vessel features from the global features (learned by the pre-trained VGG-16 base network) using the Vessel Specific Convolutional (VSC) blocks, Skip chain Convolutional (SC) layers, and feature map summations • VSC blocks and SC layers help to selectively learn the blood vessel features by performing better feature learning and feature propagation • The feature map summation plays a significant role in extracting the appropriate blood vessels apart from the background • Supervision at the output of every summation strengthens gradient propagation Deep learning techniques are instrumental in developing network models that aid in the early diagnosis of life-threatening diseases. To screen and diagnose the retinal fundus and coronary blood vessel disorders, the most important step is the proper segmentation of the blood vessels. This paper aims to segment the blood vessels from both the coronary angiogram and the retinal fundus images using a single VSSC Net after performing the image-specific preprocessing. The VSSC Net uses two-vessel extraction layers with added supervision on top of the base VGG-16 network. The vessel extraction layers comprise of the vessel-specific convolutional blocks to localize the blood vessels, skip chain convolutional layers to enable rich feature propagation, and a unique feature map summation. Supervision is associated with the two-vessel extraction layers using separate loss/sigmoid function. Finally, the weighted fusion of the individual loss/sigmoid function produces the desired blood vessel probability map. It is then binary segmented and validated for performance. The VSSC Net shows improved accuracy values on the standard retinal and coronary angiogram datasets respectively. The computational time required to segment the blood vessels is 0.2 seconds using GPU. Moreover, the vessel extraction layer uses a lesser parameter count of 0.4 million parameters to accurately segment the blood vessels. The proposed VSSC Net that segments blood vessels from both the retinal fundus images and coronary angiogram can be used for the early diagnosis of vessel disorders. Moreover, it could aid the physician to analyze the blood vessel structure of images obtained from multiple imaging sources. [ABSTRACT FROM AUTHOR]

• Clinicians predict bone age by manually reading X-ray hand bone images. • SVM automatic bone age assessment method has bone age prediction capability. • ST-ResNet network model is based on ResNet and spatial transformer. • ST-ResNet network model demonstrates better bone age prediction accuracy. • Deep learning can facilitate bone age detection more effectively. Bone age prediction can be performed by medical experts manually assessment of X-ray images of the hand bone. In practice, the workload is huge, resource consumption is large, measurement takes a long time, and it is easily influenced by human factors. As such, manual estimation of bone age takes a long time and the results fluctuate greatly depending on the proficiency of the radiologist. The left-hand X-ray image data was identified and pre-processed. X-ray image analysis method using on deep neural network was used to automatically extract the key features of the left-hand joint bone age, and evaluation performance of the model was implemented. In this paper, the deep learning method can be used to obtain the X-ray bone image features, and the convolutional neural network is used to automatically assess the age of bone. The feature region extraction method based on deep learning can extract feature information with superior performance compared to the traditional image analysis technique. Based on the residual network (ResNet) model in the deep learning algorithm, the average absolute error of the age of bones detected by the bone age assessment model is 0.455 better than traditional methods and only end-to-end deep learning methods. When the learning rate is greater than 0.0005, the MAE of Inception Resnet v2 model is higher than most models. Accuracy of bone age prediction is as high as 97.6%. In comparison with the traditional machine learning feature extraction technique, the convolutional neural network based on feature extraction has better performance in the bone age regression model, and further improves the accuracy of image-based age of bone assessment. [ABSTRACT FROM AUTHOR]

• Data augmentation performed to create a new sample image by performing transformation like flipping, rotation etc., to solve the problem of overfitting created due to the model complexity of a larger number of parameters. • Extra convolutional layer added to extract more important features and increase the overall accuracy in the detection of the left ventricle. • Drop out layer introduced during the feature extraction process to enable the model to learn only the important features and drop the irrelevant features to fix the problem overfitting and increases generalizability. deep learning algorithms have not been successfully used for the left ventricle (LV) detection in echocardiographic images due to overfitting and vanishing gradient descent problem. This research aims to increase accuracy and improves the processing time of the left ventricle detection process by reducing the overfitting and vanishing gradient problem. the proposed system consists of an enhanced deep convolutional neural network with an extra convolutional layer, and dropout layer to solve the problem of overfitting and vanishing gradient. Data augmentation was used for increasing the accuracy of feature extraction for left ventricle detection. four pathological groups of datasets were used for training and evaluation of the model: heart failure without infarction, heart failure with infarction, and hypertrophy, and healthy. The proposed model provided an accuracy of 94% in left ventricle detection for all the groups compared to the other current systems. The results showed that the processing time was reduced from 0.45 s to 0.34 s in an average. the proposed system enhances accuracy and decreases processing time in the left ventricle detection. This paper solves the issues of overfitting of the data. [ABSTRACT FROM AUTHOR]

• We present a novel method of reusing ground-truth information to improve the performance of detecting cerebral microbleeds by using the SSD algorithm. • The ground-truth information is using to stabilize the feature in forwarding propagation of the convolution network. • Compared with baseline results, using the feature enhancement achieve more effective results. • This method is flexible to be integrated into other object detecting methods. Cerebral microbleeds (CMBs) are cerebral small vascular diseases and are often used to diagnose symptoms such as stroke and dementia. Manual detection of cerebral microbleeds is a time-consuming and error-prone task, so the application of microbleed detection algorithms based on deep learning is of great significance. This study presents the feature enhancement technology applying to improve the performances of detecting CMBs. The primary purpose of the feature enhancement is emphasizing the meaningful features, leading deep learning network easier and correctly to optimize. In this study, we applied feature enhancement in detecting CMBs from brain MRI images. Feature enhancement enhanced specific intervals and suppressed the useless intervals of the feature map. This method was applied in SSD-512 and SSD-300 algorithm, using VGG architecture pre-trained in the ImageNet dataset. The proposed method was applied in SSD-512. Moreover, the model was trained and tested on the sequence of SWAN images of brain MRI images. The results of the experiment demonstrate that our method effectively improves the detection performance of the SSD network in detecting CMBs. We train SSD-512 120000 iterations and test results on the test datasets, by applying the feature enhancement layer, improving the precision with 3.3% and the mAP of 2.3%. In the same way, we trained SSD-300, improving the mAP of 2.0%. 2.8% and 7.4% precision are improved by applying feature enhancement layer In ResNet-34 and MobileNet. The proposed method achieved more effective performance, demonstrated that feature enhancement can be a helpful algorithm to enhance the deep learning model. [ABSTRACT FROM AUTHOR]