Nowadays, teenagers and children have many problems with poor posture. In order to timely and accurately identify the health problems of teenagers' posture, this study proposes to use the bottom-up Openpose bone point marking technology based on deep learning in human posture recognition technology to take human posture images and combine them with the evaluation criteria of poor posture to quickly identify whether there are high and low shoulders, high and low hips, Genu varum Genu valgum, head position deviation and other posture problems, and give the corresponding severity level. Manual correction is made by referring to the human skeleton joint points used for posture evaluation specified in the national standard. According to the posture recognition evaluation standard, the angle value required for actual abnormal body diagnosis is measured by simulating manual measurement, and the consistency between the angle value calculated by the bone point marker model and the angle value obtained by manual measurement is analyzed, Four out of the five Cronbach's Alpha consistency coefficients obtained are all greater than 0.8, indicating high consistency. And through manual measurement, the actual severity level of various abnormal postures was obtained, which was compared with the severity level of abnormal postures calculated based on the output of bone key point coordinates from the bone point marker model, with an accuracy rate of over 93%. Therefore, bone point labeling technology can be well applied to the diagnosis of adolescent physical health problems, which is of great significance for rapid and convenient diagnosis of adolescent physical health problems and timely correction and treatment. It breaks away from the limitations of traditional manual measurement methods and improves the efficiency and accuracy of abnormal body detection.
Tongue segmentation is a key step to realize the intelligent tongue diagnosis of traditional Chinese medicine. The segmentation based on deep neural networks has achieved good results with a large amount of data with pixel-level annotation. But it is difficult to obtain a large amount of pixel-level labeled real data. In this work, we propose a simple and effective tongue segmentation method using synthetic data with rich pixel-level annotations. Firstly, we adopt the Fourier transform and its inverse transform to reduce the distribution difference between the target and source domains. Then we construct fold atrous spatial pyramid pooling (FASPP) to get multi-scale information and enhance the correlation among local features for tongue data with different sizes. Finally, the experimental results show the effectiveness of our approach.
Multimodal magnetic resonance imaging (MRI) contains complementary information in anatomical and functional images that help the accurate diagnosis and treatment evaluation of lung cancers. Accurately segmenting tumor regions in each modality can help to obtain comprehensive and precise information. Existing multimodal segmentation methods are mostly used for images with strict registration, but it is difficult to achieve it for lung MRI images. It is challenging to obtain accurate tumor segmentation in lung anatomical and functional MRI images simultaneously. In this paper, we propose a method for lung tumor segmentation in weakly paired anatomical and functional MRI images. Firstly, we use domain adaptation to narrow the gap in features across different modalities, enabling the segmentation network to adapt to different modality data simultaneously. At the same time, we explore a two-stage multimodal co-attention mechanism to help the extraction and effective fusion of multi-modal information. We evaluate the proposed method on lung tumor segmentation with a clinical dataset of 90 chest MRI scans of non-small cell lung cancer (NSCLC). The results show that this method effectively improves the segmentation accuracy of each modality, the DSC of anatomical MRI is increased to 0.81±0.19, and the DSC of functional MRI is increased to 0.77±0.23, which is significantly improved compared with several multimodal tumor segmentation methods (p <0.05).
Snakebite is a major global public health problem, with more than 5 million snakebite victims recorded each year. Timely diagnosis and appropriate treatment of snakebite are the keys to preventing serious complications and improving the patient’s prognosis. However, most medical institutions lack experience in snakebite treatment, which is prone to misdiagnosis or even missed diagnosis, and serious cases endanger the lives of patients. Artificial intelligence is widely used in computer-aided diagnosis systems (CAD) and plays a vital role in snakebite recognition. In recent years, research has tended to build models, ignoring data preparation and analysis. Therefore, there is an urgent need for a rapid and accurate diagnostic method for snakebite to assist clinical diagnostic decision-making. We propose a snakebite auxiliary diagnosis system combining edge computing and machine learning technology. The system has the advantages of simple deployment, strong flexibility, and simple operation and is not limited by the equipment conditions of some grassroots hospitals. Experimental results show that this system can quickly and accurately diagnose the types of venomous snakes, overcome the limitations of traditional snakebite diagnosis methods in various aspects, and provide reliable real-time service tools for medical professionals, especially those in remote areas, allowing rapid diagnosis and treatment of snakebite cases, ultimately reducing morbidity and mortality.
For detecting sleeping and sitting postures, a system based on unrestrained flexible pressure sensor is designed and proposed. The large-area capacitive pressure sensor, which composited by 64-row and 32-column, is designed to capture posture pressure signals. A total of 30 participants were recruited for 10 postures data acquisition, and the acquired images were normalized and bilinearly interpolated. The processed data images were randomly grouped by 80% of the training set and 20% of the test set. And the deep learning neural network (YOLOv5) was used for training. The final recognition accuracy was 99.3%. Accurate division in the posture categories of sleeping and sitting postures is achieved, which is important for realizing posture monitoring.
Surgical workflow recognition has attracted widespread attention in robot-assisted surgery since it can provide surgical context information automatically, which releases the cognitive burden of the surgeons and allows more appropriate surgical decisions. One major dilemma in this community is the limitation of clinical datasets with annotated ground truth, because it requires experienced surgeons to provide specific recognition information during the annotation progress. In this paper, we developed a clinical dataset with annotated workflow information, and we provided a potential baseline for the evaluation of this dataset by predicting different surgical steps. Specifically, our dataset was captured from the robot-assisted radical prostatectomy with lymphadenectomy performed on six patients, using the da Vinci Xi robot at European Institute of Oncology, Milan, Italy, and all annotated outputs concerning various surgical information were obtained under the supervision of an experienced surgeon. Furthermore, an advanced neural network was adopted to predict surgical steps based on this dataset by using two different training strategies (i.e., the entire dataset and the downsampled one for the balance of class), and it presented a potential baseline (0.7825 DICE and 0.7918 DICE, respectively). It is expected that this dataset could promote the development of surgical workflow recognition in the medical image community, and this dataset is now accessible at the link: https://zenodo.org/record/7644037.
Sleep is essential for human life. Automatic sleep stage classification based on polysomnography (PSG) has attracted extensive attention recently, which is fundamental for the diagnosis of sleep diseases. Single-channel sleep staging methods usually ignore the bioinformation of some modalities to reduce the cost of signal acquisition, while multi-channel methods consider more modalities with higher model complexity. To this end, we propose an effective and simple network (ESNet), to effectively learn important features from multi-channel PSG and reduce model parameters for automatic sleep staging. This model utilizes bidirectional gated recurrent unit (BiGRU) to capture epoch-level waveform information and sequence-level sleep transitions of PSG, and adopting a simple convolutional block for epoch-level channel-wise feature fusion. The sequence-level learned representation is fed into a softmax layer to train an end-to-end model. Experimental results on three public datasets show that ESNet obtains the best classification performance with satisfied model parameters compared with several baselines. And the convolutional block incorporates necessary channel-wise information for sleep staging. Our work provides an effective and simple architecture to model PSG, which is promising for deep learning-based convenient sleep monitoring in the future.
Abstract: There are many problems in traditional hospital staff sign-in systems, such as low efficiency, time required to obtain result statistics, and fraud. A system based on face recognition is a very promising strategy. In 2022, the highest accuracy of face recognition algorithms around the world has exceeded 99%, with fewer than one-thousandth of one percent false positives. Face recognition technology (FRT) has been widely used in finance, intelligent security, access control, and many other fields, and the market has seen sustained rapid growth. Baidu's face recognition algorithm has won a leading place in the most authoritative public evaluation competition in the world. In this study, a hospital staff sign-in system (H3S) has been designed and developed using the Baidu AI face recognition software development kit (SDK) along with PyQt, Python, MySQL, JSON, and other tools. Compared to traditional sign-in systems, the proposed H3S was more convenient, non-contact, highly efficient and accurate, easy for data analysis, and resistant to fraud. After its deployment, sign-in efficiency has been significantly improved, checking and analysis of sign-in data were more convenient and accurate, and fraud has been essentially eliminated.
Disease identification across multiple modalities remains a challenging task. Recent research has explored early, automatic disease prediction using single modality inputs and computer vision with deep learning techniques. In this study, we introduce the Robust Multi-modal Deep Learning-based (RMDL) framework for automatic lung disease classification using both computed tomography (CT) and X-ray scans. The RMDL framework comprises segmentation, feature extraction, and classification. Input lung images undergo pre-processing, segmentation, and region of interest (ROI) extraction. The core contribution of RMDL lies in its deep learning-based automatic feature extraction and classification. We modify the Convolutional Neural Network (CNN) model to address overfitting and exploding gradients, enabling efficient feature learning with minimal computing cost. The modified CNN model employs L1 normalization and novel convolutional layers for automatic low-dimensional feature extraction. We further utilize the Long Short-Term Memory (LSTM) classifier to mitigate the exploding gradient problem, enabling accurate and early lung disease prediction. RMDL's performance was assessed using CXRTD and CCSC datasets. CCS CONCEPTS • Keywords: Classification, CNN, automatic disease prediction, deep learning, multi-modal.
Skin is the biggest organ in the human body that protects us from hurt and supports the function of material exchange. Most skin cancers develop from skin lesions which cannot avoid in our daily life. To diagnose skin cancer, now we still rely on the traditional method, the doctor’s vision examination. Since the unstable and high misdiagnosis rate, more and more researchers are trying to use deep learning methods to assist dermatologists or even automatic diagnosis. There are a lot of works related to this topic in recent years. However, most of them are based on convolution. Convolution limits the deep learning model in awareness of spatial differences and global information. In this circumstance, we proposed a new network based on the involution and Triplet++ Attention to enhance the ability of spatial information extraction and global information extraction. Also, we find a way to improve model performance by slowing down attention compression without extra parameters. Compared to the state-of-the-art method, our model improved the performance in accuracy by 2.91 up to 99.4 in the HAM10000 dataset. In another dataset PH2, our proposed method also gets a comparable result in accuracy up to 96.67.
This paper focuses on the human lumbar vertebrae and explores how changes in the position of the center of gravity of the upper body and tissue material properties of human lumbar vertebrae impact its dynamic characteristics. The modal response frequency of human lumbar vertebrae is maximized in an upright posture. If the center of gravity of the upper body deviates excessively from its equilibrium position, both forward and backward, then it will lead to a decrease in the lumbar system response frequency and adversely affect the mobile joints within the human lumbar vertebrae. When the tissue material properties of cortical bone, cancellous bone, nucleus pulposus changes, the response frequency of human lumbar vertebrae will be affected accordingly. It has been shown that changes in elastic modulus of cancellous bone have the greatest influence on the response frequency of human lumbar vertebrae, compared to other tissues. The conclusion of this paper will help to reveal the physiological mechanisms and motion characteristics of the human lumbar vertebrae, thus guiding the practice in the fields of ergonomics, healthcare and rehabilitation therapy.
Given the intricate three-dimensional structure and variable density of lung tissue, accurate airway segmentation remains a challenging task. This study proposes a method rooted in an enhanced U2-Net t architecture that excels in identifying small peripheral bronchi in non-contrast CT scans. Replacing all ReSidual U-blocks in the U2-Net network with three-dimensional convolutions enables direct processing of CT images. Furthermore, the introduction of Instance Normalization enhances the model's capacity to discern individual features. Optimization of upsampling and downsampling mechanisms minimizes information loss during sampling and preserves crucial image details. We also employed an advanced Loss function, expediting model convergence. Experimental results indicate that these enhancements substantially elevate the performance. Our method surpasses existing strategies by extracting a more extensive array of fine branches.
Due to the unique physiological structure of human lung tissue, segmentation of the lung parenchymal regions is a very challenging task. Traditional image segmentation methods have certain limitations on practical applications. Therefore, lung tissue segmentation methods based on deep learning have become a research hotspot in recent years. Res BCDU-Net may overcome the problem of high false positives in other semi-automatic segmentation, but it still has the problem of low accuracy in edge segmentation, which is probably caused by the unclear input of a single CT image. Therefore, we propose to utilize three consecutive CT images as the input, which employs inter-slice context to restrain the false positive. When a single input CT image is not clear, the accuracy of lung segmentation results can be improved by referring to its previous and next slices. Based on the input of three continuous images, the spatial information in the process of lung segmentation is added in this paper. We build a Res BCDU-SCB network and collect and integrate spatial information through SC-BLSTM module. An SC-attention module is designed to take both finer-grained spatial information and rich semantic information into account. In this paper, we append BLSTM (Bidirectional Convolutional Long Short-term Memory) module with SC-attention module, named SC-BLSTM, to decrease performance degradation caused by ambiguous boundaries and variable lung parenchymal shapes. The results of extensive experiments on lung CT images (LIDC-IDRI database) show that it may improve the final segmentation accuracy and reduce the false positives, which confirms the effectiveness of our proposed method.
research-article Share on Mechanical Properties and Experimental Validation of Serial Pneumatic Artificial Muscle System Authors: Kai Ding School of information and engineering, Jia Xing University, China School of information and engineering, Jia Xing University, China 0009-0008-1081-1081Search about this author , Feilong Jiang School of information and engineering, Jia Xing University, China School of information and engineering, Jia Xing University, China 0000-0003-0382-1245Search about this author , Qingwei Li Disciplinary Construction, Yunnan University of Business Management, China Disciplinary Construction, Yunnan University of Business Management, China 0000-0002-2972-6169Search about this author , Hao Liu State Key Laboratory of Fluid Power and Mechatronic Systems, Zhejiang University, China State Key Laboratory of Fluid Power and Mechatronic Systems, Zhejiang University, China 0009-0004-6453-1115Search about this author , Daxia Chai AccuPath Medical (Jiaxing) Co., Ltd., China AccuPath Medical (Jiaxing) Co., Ltd., China 0000-0003-1413-9700Search about this author , Kun Liu School of information and engineering, Jia Xing University, China School of information and engineering, Jia Xing University, China 0009-0004-4427-7324Search about this author Authors Info & Claims ICBIP '23: Proceedings of the 2023 8th International Conference on Biomedical Signal and Image ProcessingJuly 2023Pages 72–76https://doi.org/10.1145/3613307.3613321Published:28 September 2023Publication History 0citation0DownloadsMetricsTotal Citations0Total Downloads0Last 12 Months0Last 6 weeks0 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access
Highly development of language-image models makes prompt-driven accurate segmentation become possible. Segment Anything Model (SAM) has recently made a breakthrough in zero-shot image segmentation, using an unprecedentedly large dataset to train a segmentation model with strong adaptability. In this paper, we investigate the capability of SAM for MRI medical images in uterus Segmentation. Experimental results that YOLOv8 with SAM implements end-to-end uterus segmentation and outperforms the traditional supervised learning U-net and U-net++ models.
At present, research mainly uses single modal and multimodal physiological signals to identify human psychological states, but these methods generally have the drawbacks of low evaluation accuracy, poor applicability, and weak effectiveness. To address the above shortcomings, this article proposes a psychological stress assessment method based on the privileged information learning paradigm framework. This method uses the SVM+ algorithm under the privileged information learning paradigm framework to predict from the WESAD dataset and compares its performance with other algorithms. According to the experimental results, the SVM+ algorithm has classification accuracy and robustness compared to other comparative algorithms. The experiment in this article can provide reference for subsequent research on convenient psychological stress assessment.
Due to the increasing number of patients with impaired immune function and the widespread use of broad-spectrum antibiotics, the rate of fungal infections is rising and invasive fungal infections (IFD) have become one of the major causes of death. Therefore, early diagnosis and treatment of IFD is the focus of medical attention. In order to automatically identify and diagnose the fungal species of infected patients, we propose a fluorescent identification method of skin fungi with a compact model and high identification accuracy. The method uses fluorescence staining to label fungi, and data enhancement processing is applied to the collected sample images to compensate for the shortcomings of less clinical data of some fungal samples. Convolutional Tokenizer and Positional Embedding are applied to provide sample image labeling in the recognition process, and Transformer is used to solve the problem of too many layers of traditional Convolutional Neural Networks and expand the image perceptual field; In the Sequence Pooling module, the attention mechanism is used to pool all the tokens. Finally, it is transformed into the final category output through a Fully Connected layer. The experimental results show that the method takes less training time, has a compact model and a smaller number of parameters, while still maintaining a high recognition accuracy, which is of great practical value in clinical diagnosis.
Ocular diseases pose significant challenges in the field of ophthalmology. The accurate classification of fundus images depicting various ocular diseases plays a vital role in early diagnosis and effective treatment planning. However, medical datasets for such conditions often suffer from limited samples and imbalanced class distributions, making classification tasks more challenging. In this study, we address the multi-classification of fundus images of ocular diseases, considering the aforementioned dataset limitations. We specifically focus on three distinct ocular diseases, along with normal fundus images. In medical data analysis, the availability of labeled data is often limited, and the distribution of classes is frequently imbalanced. To address these challenges, we propose the utilization of triple generative adversarial nets (TripleGAN), a semi-supervised generative network, for medical data classification tasks. Through rigorous experimentation and training, we demonstrate the effectiveness and robustness triple generative adversarial nets. After 500 iterations, our model achieves a stable test accuracy of approximately 70%. Furthermore, our approach exhibits resilience to loss variation, ensuring consistent performance across different stages of training. The results of our study indicate promising potential in the accurate classification of fundus images of ocular diseases, even in the presence of under-sampled and unbalanced datasets. Our findings contribute to the research in the field of ophthalmic image analysis and may aid in the development of advanced diagnostic tools for ocular diseases.
Thyroid nodule is a common chronic endocrine system disease, which may deteriorate into thyroid cancer without treatment. However, due to the distribution of many blood vessels, nerves and organs near the thyroid, the clarity of ultrasonic image is poor, and human eye recognition has great limitations.We propose a method for thyroid nodule diagnosis based on multi-scale Se-SegNet network and ResNet50. We improved and optimized the SegNet as the main backbone, and constructed a multi-scale Se-SegNet network to segment thyroid ultrasonography. Then we used ResNet50 to classify the segmented images and get the diagnosis results. The accuracy of the multi-scale Se-SegNet image segmentation model is 0.959.The accuracy of classification results was 0.967. This paper proposes a new diagnostic scheme for thyroid nodule, which can diagnose thyroid nodule quickly and accurately. The experimental results show that it has higher diagnostic accuracy than the traditional scheme..
Tumor Treating Fields (TTFields) therapy is a promising cancer treatment technique that inhibits tumor cell proliferation by applying alternating electric fields. Numerical simulations of TTFields electric fields based on realistic models have been widely used to analyze and optimize treatment implementation. However, most simulations are limited by solving tools, require laborious pre-processing of the model, and can only be performed manually. The paper presents a new implementation for simulating the electric field distribution of TTFields based on an open-source finite element solver. The reliability of the new implementation was verified by comparing with the analytical solution of a simple cylindrical model. Then, compared with the common simulation scheme on a realistic head model, the practical significance of this implementation in simulating TTFields electric field distribution was evaluated. This implementation reduces the workload of pre-processing and provides convenience for building an automated workflow for TTFields modeling.
