Abstract: Cervical cancer is a malignant tumor that highly endangers women’s lives. Cytological screening based on image processing is the most widely used detection method for precancerous screening. Recently, with the development of machine learning theory based on deep learning, the convolutional neural network has made a revolutionary breakthrough in the field of image recognition due to its strong and effective extraction ability. In addition, it has been widely used in the field of medical image analysis such as cervical abnormal cell detection. However, due to the characteristic high resolution and large size of pathological cell images, most of its local areas do not contain cell clusters. Moreover, when the deep learning model uses the method of exhausting candidate boxes to locate and identify abnormal cells, most of the sub-images obtained do not contain cell clusters. When the number of sub-images increases gradually, a large number of images without cell clusters as input to the object detection network will make the image analysis process redundant for a long time, which drastically slows down the speed of detection of the large-scale pathological image analysis. In view of this, this paper proposed a new detection strategy for abnormal cells in cervical cancer microscopic imaging. According to the pathological cell images obtained by the membrane method, the image classification network based on deep learning was first used to determine whether there were abnormal cells in the local area. If there were abnormal cells in the local area, the single-stage object detection method was used for further pathological cell image analysis, so that the abnormal cells in the images could be quickly and accurately located and identified. Experimental results show that the proposed method can double the speed of detection of cervical cancer abnormal cells.
Abstract: Hepatocellular carcinoma (HCC) is a type of primary malignant tumor and an urgent problem to be solved, particularly in China, one of the countries with the highest prevalence of HCC. In the choice of treatment methods for patients with hepatocellular carcinoma, accurate histological grading of the lesion area undoubtedly plays a vital role that helps the management and therapy of HCC patients. However, the current pathological detection as the gold standard has defects, such as invasiveness and a large sampling error. Therefore, it is an important direction of intelligent medical treatment to provide noninvasive and accurate lesion grading using imaging technology combined with artificial intelligence technology. On the basis of the radiologists' experience in reading clinical images, this paper proposed a self-attentional guidance-based histological differentiation discrimination model combined with multi-modality fusion and an attention weight calculation scheme for dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) sequences of hepatocellular carcinoma. The model combined the spatiotemporal information contained in the enhancement sequence and learned the importance of each sequence and the slice in the sequence for the classification task. It effectively used the feature information contained in the enhancement sequence in the temporal and spatial dimensions to improve the classification performance. During the experiment, the model was trained and tested on the clinical data set of the top three hospitals in China. The experimental results show that the self-attention-guided model proposed in this paper achieves higher classification performance than several benchmark and mainstream models. Comprehensive experiments were performed on the clinical dataset with labels annotated by professional radiologists. The results show that our proposed self-attention model can achieve acceptable quantitative measuring of HCC histologic grading based on the MRI sequences. In the WHO histological classification task, the classification accuracy of the proposed model reaches 80%, the sensitivity is 82%, and the precision is 82%.
Abstract: To solve the problem that a single kernel learning support vector machine (SVM) cannot consider the learning and generalization abilities and parameter optimization of the multiple kernel function, a multiple kernel learning support vector machine (MKL-SVM) algorithm based on swarm intelligence optimization was proposed. First, the impact of five single kernel functions on the classification indexes of SVM was discussed. These kernel functions include two global kernel functions — the polynomial and sigmoid kernel functions — and three local kernel functions—the radial basis function, exponential kernel function, and Laplacian kernel function. Next, an MKL-SVM algorithm with a convex combination of a polynomial kernel having global properties and a Laplacian kernel having local properties was proposed. Then, to improve particle diversity to avoid falling into local optimal solutions during the iteration, and to reduce the model’s training time, the crossover operation in the genetic algorithm was introduced into the particle swarm optimization (PSO) algorithm. This improved swarm intelligence optimization was used to optimize the parameters of the MKL-SVM. Finally, deep learning features based on the classical model VGG16 and handcrafted features according to doctors’ suggestions were used as inputs for the recognition algorithm. In this algorithm, transfer learning was used to extract deep learning features and principal component analysis was used to reduce computational complexity through dimensionality reduction. The results show that using deep learning features is better than handcrafted features. Therefore, this paper adopts the deep learning features as input for the MKL-SVM algorithm and the hybrid swarm intelligent optimization algorithm of crossover genetic and the PSO algorithm as the optimization method. To verify the generalization ability of the proposed algorithm, the public dataset LUNA16 was selected for testing. The experimental results show that the proposed algorithm is easy to jump out of the local optimal solution, improves the learning ability and generalization ability of the algorithm, and has a better classification performance.
Abstract: Atrial septal defect (ASD) is common congenital heart disease. The detection rate of congenital heart disease has increased year by year, and ASD accounted for the largest proportion of it, reaching 37.31%. The ASD patient will suffer from shortness of breath, palpitation, weakness, etc., with symptoms worsening with advanced age. The ASD patient will not suffer from congenital heart disease if their condition is diagnosed early. Echocardiography is a powerful and cost-effective means of detecting ASD. However, the disadvantages of echocardiography, such as noise and poor imaging quality, cause misdiagnosis of ASD. Hence, research into echocardiography-based efficient and effective detection of ASD with a deep neural network is of great significance. For echocardiography is noisy and fuzzy, and the learning and feature expression ability of the traditional convolutional neural network architecture is limited, a feature view classification based atrial septal defect intelligent auxiliary diagnostic model architecture was proposed. The different views of echocardiography possess different features, demanding more precise model extraction and combined features from echocardiography. The proposed model architecture integrates the semantic characteristics of several views, significantly improving the accuracy of diagnosis. In addition, with the aim of denoising and preserving edges, a bilateral filtering algorithm was performed. Furthermore, an ASD intelligent auxiliary diagnostic system was built based on the proposed model. The results show that the accuracy of the ASD auxiliary diagnostic system reaches 97.8%, and the false-negative rate is greatly reduced compared with the traditional convolutional neural network architecture.
Abstract: For a microscopic imaging scene, an all-in-focus image of the observation object is needed. Because of the limitation of the depth of field of the camera and the typically uneven surface of the observation object, an all-in-focus image is obtained through one shot with relative difficulty. In this case, an alternative method for obtaining the all-in-focus image is usually used, which is to fuse several images focusing on different depths with the help of multi-focus image fusion technology. Multi-focus image fusion is an important branch in the field of computer vision. It aims to use image processing technology to fuse the clear regions of multiple images, focusing on different objects in the same scene, and finally to obtain an all-in-focus fusion result. With the breakthrough of machine learning theory represented by deep learning, the convolutional neural network is widely adopted in the field of multi-focus image fusion. However, most methods only focus on improving network structure and use the simple one-by-one serial fusion method, which reduces the efficiency of multiple image fusion. In addition, the defocus spread effect in the fusion process, which causes blurred artifacts in the areas near focus map boundaries, can severely affect the quality of fusion results. In the application of microscopic imaging analysis, we proposed a maximum spatial frequency in the feature map (MSFIFM) fusion strategy. By adding a post-processing module in the convolution neural network based on unsupervised learning, the redundant feature extraction process in the one-by-one serial fusion is avoided. Experiments demonstrate that this strategy can significantly improve the efficiency of multi-focus image fusion with multiple images. In addition, we presented a correction strategy that can effectively alleviate the effect of defocus spread on the fusion result under the condition of ensuring the efficiency of the algorithm fusion.
Abstract: The rapid development of artificial intelligence (AI) has injected new vitality into various industries and provided new ideas for the development of traditional Chinese medicine (TCM). The combination of AI and TCM provides more technical support for TCM auxiliary diagnosis and treatment. In the history of TCM, many methods of syndrome differentiation have been observed, among which the differentiation of Zang-fu organs is one of the important methods. The purpose of this paper is to provide support for the localization of Zang-fu in TCM through AI technology. Localization of Zang-fu organs is a method of determining the location of lesions in such organs and is an important stage in the differentiation of Zang-fu organs in TCM. In this paper, the localization model of TCM Zang-fu organs through the neural network model was established. Through the input of symptom text information, the corresponding Zang-fu label for a lesion could be output to provide support for the realization of Zang-fu syndrome differentiation in TCM-assisted diagnosis and treatment. In this paper, the localization of Zang-fu organs was abstracted as multi-label text classification in natural language processing. Using the medical record data of TCM, a Zang-fu localization model based on pretraining models a lite BERT (ALBERT) and bidirectional gated recurrent unit (Bi-GRU) was proposed. Comparison and ablation experiments finally show that the proposed method is more accurate than multilayer perceptron and the decision tree. Moreover, using an ALBERT pretraining model for text representation effectively improves the accuracy of the localization model. In terms of model parameters, the ALBERT pretraining model greatly reduces the number of model parameters compared with the BERT model and effectively reduces the model size. Finally, the F1-value of the Zang-fu localization model proposed in this paper reaches 0.8013 on the test set, which provided certain support for the TCM auxiliary diagnosis and treatment.
Abstract: Medical records, as an essential part of the health care records of residents, save all the information about the clinical treatment of patients, which are traditionally written by doctors on paper. With the development of information technologies, electronic medical records that are more easily saved and managed gradually replace the traditional ones. Intelligent auxiliary diagnosis, patients’ portrait construction, and disease prediction based on medical reports have become research hotspots in the field of intelligent medical care. To fully discover the hidden relationship between symptoms and diseases from the documents saved in electronic medical records, the development of an efficient named entity recognition algorithm is the key issue. Although several studies have been conducted on it, there is relatively little research on the information extraction of Chinese electronic medical records. To the best of our knowledge, the documents in Chinese electronic medical records contain a large number of nested named entities and short sentences. Moreover, there is weak logic among the sentences, causing a complex syntax structure. To effectively recognize the medical entities, a novel named entity recognition method based on multifeature embedding and attention mechanism was proposed. After embedding three types of features derived from characters, words, and glyphs in the input presentation layer, an attention machine was introduced to the hidden layer of the bidirectional long short-term memory network to make the model focus on the characters related to the medical entities. Finally, the optimal labels for the five types of entities in Chinese electronic medical records, including diseases, body parts, symptoms, drugs, and operations, were obtained. The experimental results for the open and self-built Chinese electronic medical records, recognition accuracy, recall rate, and F1 value of the proposed algorithm are all better than 97%, which shows that the proposed algorithm can effectively identify various entities in Chinese electronic medical records.
Abstract: In the field of medicine, in order to diagnose a patient’s condition more efficiently and conveniently, image classification has been widely leveraged. It is well established that when doctors diagnose a patient’s condition, they not only observe the patient’s image information (such as CT image) but also make final decisions incorporating the patient’s clinical diagnostic information. However, current medical image classification only puts the image into a convolution neural network to obtain the diagnostic result and does not use the clinical diagnosis information. In intelligent auxiliary diagnosis, it is necessary to combine clinical symptoms with other imaging data for comprehensive diagnosis. This paper presented a new method of assistant diagnosis for the medical field. This method combined information from patients’ imaging with numerical data (such as clinical diagnosis information) and used the combined information to automatically predict the patient’s condition. Based on this method, a medical assistant diagnosis model based on deep learning was proposed. The model takes images and numerical data as input and outputs the patient’s condition. Thus, this method is comprehensive and helps improve the accuracy of automatic diagnosis and reduce diagnostic error. Moreover, the proposed model can simultaneously process multiple types of data, thus saving diagnosis time. The effectiveness of the proposed method was verified in two groups of experiments designed in this paper. The first group of experiments shows that if the unrelated data are fused for classification, the proposed method cannot enhance the classification ability of the model, although it is able to predict multiple diseases at one time. The second group of experiments show that the proposed method could significantly improve classification results if the interrelated data are fused.
Abstract: Vital signs are important parameters for human health status assessment, and timely, accurate detection is of great significance for modern health care and intelligent medical applications. Detecting vital signs, such as heartbeat and respiration signals, provides a variety of diseases with reliable diagnosis and effective prevention. Conventional contact detection may restrict the behaviors of users, cause additional burdens, and render users uncomfortable. In recent years, noncontact detection technology has successfully achieved remote long-term detection for respiration and heartbeat signals. Compared to conventional contact-detection approaches, noncontact heartbeat and respiration detection using a millimeter-wave radar is preferable as it causes no disturbance to the subject, bringing a comfortable experience, and detects vital signs under natural conditions. However, noncontact vital signs detection is challenging owing to environmental noise. Especially, heartbeat signals are very weak and are merged with respiration harmonics and environmental noise, and their extraction and recognition are even more difficult. This paper applied a frequency-modulated continuous wave (FMCW) radar to detect vital signs. The study also presented a noncontact heartbeat and respiration signals detection approach based on wavelet analysis and autocorrelation computation (WAAC). The millimeter-wave FMCW radar first transmited the electromagnetic signal and received the reflected echo signals from the human body. Thereafter, the phase information of the intermediate frequency signals was extracted, which included respiration and heartbeat signals. The direct current offset of the phase information was corrected, and the phase was unwrapped. Finally, the wavelet packet decomposition was used to reconstruct heartbeat and respiration signals from the original signal, and an autocorrelation computation was utilized to reduce the effect of clutters on the heart rate detection. Experiments were conducted on ten subjects. Results show that the average absolute error percentage of WAAC is less than 1.65% and 1.83% for respiration and heartbeat rates, respectively.
Abstract: Continuous glucose monitoring is important in the management of diabetes. According to statistics, diabetes is the third chronic non-infectious disease that seriously endangers people's health, followed by tumor as well as cardiovascular and cerebrovascular diseases. In 2019, globally, there were a total of 460 million diabetics aged 20–79 years, which accounted for 9.1% of the total population in this cohort. Each figure is projected to increase to 592 million and by 10.1% respectively by 2035. Currently, the methods of blood glucose monitoring can be divided into invasive, minimally invasive, and noninvasive. The main methods for blood glucose monitoring include irregular sampling of fingertip blood or consecutive measurement of interstitial fluid glucose based on implantable sensors. However, these methods have some limitations, which include pain sensation, high cost, short service life, and susceptibility. Patients need to measure their blood glucose frequently. Invasive and minimally invasive monitoring will cause physical and psychological pain. Therefore, noninvasive monitoring is one of the most promising techniques for continuous monitoring of blood glucose, and it has a broad market prospect. In this study, the electrocardiogram (ECG signals) were used to achieve the noninvasive monitoring of blood glucose levels. First, 756160 ECG periodic signals of 12 volunteers for 60 d were obtained from the experiment. Second, the ECG signals were preprocessed using an infinite impulse response filter. Furthermore, a method combining convolutional neural networks and long short-term memory networks (CNN-LSTM) was proposed for blood glucose monitoring. In Addition, two modeling methods (individual modeling and group modeling) were investigated in this study. The results show that the precision of blood glucose monitoring under the condition of individual and group modeling is 80% and 88%, respectively. The F1-score of the group modeling can reach 0.95, 0.88, 0.91, 0.85, 0.92, 0.88, 0.86, 0.86, 0.87, and 0.86. Therefore, this study indicates that the proposed method based on ECG signals can provide powerful theoretical support and technical guidance for real-time and accurate blood glucose monitoring.
Abstract: Arrhythmia is a common cardiovascular disease whose occurrence is mainly related to two factors: cardiac pacing and conduction. Some severe arrhythmias can even threaten human life. An electrocardiogram (ECG) records the changes in electrical activity generated during each cardiac cycle of the heart, which can reflect the human cardiac health status and help diagnose arrhythmias. However, because of the brevity of conventional ECGs, arrhythmias, which occasionally occur in daily life, cannot be detected easily. Automatic ECG analysis-based long-term heartbeat monitoring is of great significance for the effective detection of accidental arrhythmias and then for taking indispensable measures to prevent cardiovascular diseases in time. An ensemble extreme learning machine (ELM) approach for heartbeat classification that fuses handcrafted features and deep features was proposed. The manually extracted features clearly characterize the heartbeat signal, where RR interval features reflect the time-domain characteristic, and the wavelet coefficient features reflect the time–frequency characteristic. A 1D convolutional neural network (1D CNN) was designed to automatically extract deep features for heartbeat signals. These features were fused by an ELM for heartbeat classification. Because of the instability caused by the random assignment of ELM hidden layer parameters, the bagging ensemble strategy was introduced to integrate multiple ELMs to achieve stable classification performance and good generalization ability. The proposed approach was validated on the MIT-BIH arrhythmia public dataset. The classification accuracy reaches 99.02%, and the experimental results show that the performance of the proposed approach with fused features is better than those with only deep features and only handcrafted features.
Abstract: With advancements in modern medical technology, the treatment of rhegmatogenous retinal detachment has been receiving increasing attention. Globally, vitrectomy combined with intraocular silicone oil tamponade has been widely used for rhegmatogenous retinal detachment, and the surgical equipment and technology required are increasingly advanced. In such an operation, it is crucial to understand how to achieve the best therapeutic effect with the minimum amount of silicone oil tamponade so as to reduce surgical complications. Traditional medical methods cannot effectively evaluate the effect of different silicone oil dosages on retinal hole attachment. Aiming at this concern, the current study proposed a silicone oil tamponade simulation method for retinal detachment surgery. Based on physical modeling and computer numerical discretization techniques, the intraocular force and silicone oil filling state were analyzed. Three-dimensional modeling and simulation of the silicone tamponade process were then conducted and visualized to help with medical decision-making. First, the human eyeball and surgical instruments were modeled and sampled to simulate the eyeball state during the operation. Second, based on differences in density, viscosity coefficient, and surface tension between water and silicone oil, the two-phase flow and water?silicone oil interaction were simulated. Finally, the solid?liquid interaction model was constructed to assess the movement and injection process of multiphase liquid in the eyeball. The experimental results show that this method can well present the interaction effect of multiphase fluid movement in the eyeball; understand effects such as surface tension, solid–liquid coupling, liquid stratification, and connector effect; and realize the simulation of the silicone oil injection and water drainage processes through the catheter in the intraocular cavity, which provides an effective way to predict the intraocular state after silicone oil filling and assists doctors in the field of operation process planning and effect prediction.
Abstract: Accurate identification of retinal vessels is essential for assisting doctors in screening early fundus diseases. Diabetes, hypertension, and cardiovascular disease can cause abnormalities of the retinal vascular structure. Retinal vessel segmentation maps can be quickly obtained using the automated retinal vessel segmentation technology, which saves time and cost of manually identifying retinal vessels. Aiming at the problem of incomplete and inaccurate extraction of fine retinal vessels, this paper explored the design of a multitask convolutional neural network and the topological relationship of retinal vessels. A cascaded retinal vessel segmentation network framework guided by a skeleton map was proposed. The auxiliary task of skeleton extraction was used to extract vessel centerlines, which could maximally preserve topological structure information. SAFF cascaded the two modules by remaining embedded between their feature layers. This process could effectively fuse the structural features with the vessel local features by learning pixel-wise fusion weight and thus enhancing the structural response of features in the vessel segmentation module. To obtain a complete skeleton map, the skeleton map extraction module introduced a graph-based regularization loss function for training. Compared with the latest vessel segmentation methods, the proposed approach wins the first place among the three public retinal image datasets. F1 metrics of the proposed method achieved 83.1%, 85.8%, and 82.0% on the DRIVE, STARE, and CHASEDB1 datasets, respectively. Ablation studies have shown that skeleton map-guided vessel segmentation is more effective, and graph-based regularization loss further improves accuracy of the retinal vessel segmentation compared to the vanilla network. Moreover, the framework generality is verified by replacing the skeleton map extraction and vessel segmentation modules with various convolutional networks.
Abstract: High-quality sleep is linked with physical development, cognitive function, learning, and attention in children. Since early symptoms of sleep disorders in children are not obvious and require long-term monitoring, there is an urgent need to develop a method for monitoring children’s sleep that can prevent and diagnose these disorders in advance. Polysomnography (PSG) is the basic test for sleep disorders recommended by clinical guidelines. Sleep quality can be assessed and sleep disorders can be identified by observing the changes in patterns of PSG during each sleep period. Sleep staging in children was researched and single-channel electroencephalogram (EEG) signals recorded by PSG was used in this study. On the basis of Alexnet, we use a one-dimensional convolutional neural network (1D-CNN) model instead of a two-dimensional model to propose a 1D-CNN structure composed of five convolutional layers, three pooling layers, and three fully connected layers, as well as a batch normalization layer to 1D-CNN while keeping the size of the convolutional kernel constant. Moreover, the dataset was augmented with an overlapping method to address its small size. The experimental results showed that the accuracy of this model for children’s sleep staging was 84.3%. According to the normalized confusion matrix obtained from the PSG data of Beijing Children’s Hospital, the classification performance of wake, N2, N3, and REM stages of sleep was very good. Because stage N1 sleep was misclassified as wake, N2, and REM sleep in some cases, future research should focus on improving the accuracy of stage N1 sleep. Overall, the 1D-CNN model proposed in this paper can realize automatic sleep staging for children based on single-channel EEG with sleep stage markers. In the future, more research is needed to develop a more suitable sleep staging strategy for children and to conduct experiments with a larger amount of data.
Abstract: In recent years, with the progress and development of science and technology, the research on artificial intelligence and wearable devices continues to develop, and researchers increasingly find ways to provide a kind of more comfortable and comprehensive user experience. Auditory and visual technologies have been fully developed and utilized, while as an emerging field, tactile sense is a research direction with great potential. As a type of unique human sensory channel, tactile sense has unique advantages. It can convey information about maximum joints in physiological structure of human body, such as hardness, texture, shape, size, and temperature, which cannot be transmitted by visual and auditory senses. Additionally, the tactile sense is fast and accurate, thus, it performs well in some special situations, such as supergravity scenarios, high-speed rotating scenarios, or very noisy environments. The design of vibrational tactile coding is an important way to develop tactile devices and achieve a better human computer interaction. Haptic coding has several defects. It is used in narrow application scenes and conveys unclear meaning. Compared with the mature development of vision and hearing, it is necessary to further design the vibration haptic coding patterns to overcome these defects. Research on information transmission of tactile sense is very meaningful, for example, it offers convenience for special groups such as people with visual impairment or workers engaged in their education. It provides a navigation service for visually impaired people by changing the vibration frequency and vibration intensity of a blind vest. When hearing or vision is impaired, the tactile sense is a considerate way to provide timely and accurate information assistance to the special groups. Besides, tactile sense can help express the flight attitude information in virtue of combined vibration tactile coding. However, these studies aimed at providing a set of specific coding for a specific scenario in which relatively vague pieces of information were conveyed, such as emotion and direction Based on these studies, designing a set of universal coding patterns for most usage scenarios to deliver exact and accessible information is essential. This paper discussed the resolution of tactile vibration based on the mechanism of tactile vibration perception. According to the application of directional navigation and text interaction, the vibration information coding was summarized and the experimental methods and conclusions of vibration information coding were introduced. Moreover, the prospect of the vibration information coding field was proposed.
Monthly, started in 1955 Supervising institution：Ministry of Education Sponsoring Institution：University of Science and Technology Beijing Editorial office：Editorial Department of Chinese Journal of Engineering Publisher：Science Press Chairperson：Ren-shu Yang Editor-in-Chief：Ai-xiang Wu ISSN 2095-9389CN 2095-9389
