Accurate estimation of venipuncture angle is vital for clinical success, yet existing methods often depend on manual judgment or hardware, limiting robustness and scalability. We propose InAngleNet, a non-contact stereo vision framework that integrates segmentation and geometric modeling for precise angle estimation. By reconstructing the 3D spatial relationship between the needle shaft and skin surface, our method enables reliable inference in clinical settings. To enhance perception of slender structures, we introduce the Mask-Guided Attention Module (MGAM) to inject segmentation priors into disparity estimation, and the Edge-Aware Guidance Module (EAGM) to refine structural alignment via boundary cues. We also construct a stereo venipuncture dataset covering real, phantom, and synthetic scenarios. InAngleNet achieves a mean angular error of 1.47° and a Success Rate@ 1.5° of 83 %, surpassing strong baselines and demonstrating clinical potential.

Vein puncture is a common procedure in clinical practice, and its success rate and patient experience are significantly influenced by the puncture angle and speed. This study proposes a puncture operation detection method based on the combination of You Only Look Once (YOLO) and DEcoupled Video segmentAtion (DEVA), which enables real-time quantitative evaluation of the puncture process by detecting the needle shaft and calculating the puncture angle and speed. To enhance the accuracy of context information transfer, we introduce a feature enhancement mechanism that leverages YOLO’s multi-scale features to augment the feature representation in the memory module. To tackle challenges such as speckle noise, blurred segmentation, and omissions, we improve the in-clip consensus mechanism to further enhance temporal consistency. Additionally, to further improve segmentation accuracy and better capture critical features, we introduce an enhanced attention module. Experimental results across multiple datasets demonstrate that our approach not only maintains segmentation performance but also increases processing speed by nearly threefold.

Extracting structured information and key details from medical report narratives is crucial to support healthcare data management, analysis and decision-making. However, the specialized nature of the reports, the complexity of the contents, and the high accuracy requirements of the results pose significant challenges to the structuring task. In this paper, we develop an LLM-based method to extract structure information from medical report narratives. Defining the structuring problem as mapping the narrative reports to the domain ontology, we design a framework to develop specialized LLMs that automatically learn and establish the mappings. At the core of this framework are report partitioning and interactive training data generation modules are. By separating complete reports into logically independent segments and training the LLMs on these segments independently, the trained LLMs can accurately capture the semantic relationships within each segment. Additionally, we explore different LLMs and formulate a simplistic scoring method to compare their accuracy, enabling us to select the best-performing model. Experimental evaluation on a real-world breast ultrasound report dataset demonstrates that our method achieves high accuracy with a small training dataset (400 samples). Specifically, the accuracy of structural information extraction and the attribute-value matching accuracy both exceed 96%.

Automatic tracking and segmentation of lesions in ultrasound videos could assist in early diagnosis and treatment plan development. However, this task is quite challenging due to problems such as low visual saliency of the lesions and large variation between adjacent frames. In this paper, we develop a Spatial-Temporal Fusion Network (STFNet) for unsupervised ultrasound video object segmentation. First, an Edge Blur Enhancement Module is designed to extract and preserve the edge details of the target objects in ultrasound frames for spatial feature enhancement. Then, a Dynamic Alignment Module is developed to correct the inter-frame inconsistencies by aligning target objects from adjacent frames with those in the current frame for temporal feature enhancement. To incorporate both the spatial and temporal information, we further implement a mixed training strategy. These innovations collectively refine the model’s learning process and substantially boost segmentation accuracy. Extensive evaluations on real lymphoma ultrasound video data demonstrate the competitive segmentation results of STFNet. Specifically, compared with the best results among seven competing baselines, STFNet achieves the best scores in terms of region similarity ${\mathcal{J}}$, contour accuracy ${\mathcal{F}}$ as well as their average ${\mathcal{J}}\& {\mathcal{F}}$.

Financial market volatility forecasting plays an extremely important role in the financial field. It helps investors, traders and financial institutions more accurately assess market risks and formulate effective investment strategies. However, due to the uncertainty and complexity of financial markets, the non-stationarity and volatility of financial time series make volatility forecasting more challenging. At the same time, when forecasting the volatility of multiple financial products, the distinct market behaviors and risk characteristics of different products may lead to distortion and overfitting when using the same forecasting model. Consequently, it is necessary to construct a separate volatility forecasting model for each financial product, which in turn brings about the problems of difficulty in obtaining data and insufficient samples. To address these issues, a financial market volatility forecasting method based on domain adaptation is proposed. The method first constructs a volatility forecasting model FinWaveNet for single-domain financial products. FinWaveNet introduces causal convolutional layers and inflated convolution to capture financial time series features more comprehensively. Residual networks optimize data feature capture, while Gated Recurrent Units (GRU) and highway networks ensure stable information transmission, addressing non-stationarity and volatility challenges and enhancing the accuracy in single-domain scenarios. Subsequently, we extend our approach to construct a financial volatility forecasting model using domain adaptation methods, which combines adversarial learning and maximum mean difference (MMD) method to address data feature misalignment across different domains, thereby improving the accuracy in the multi-domain scenario. Finally, through ablation and comparison experiments, the paper validates the effectiveness of the proposed method in financial field volatility forecasting, demonstrating optimal performance based on the MAE evaluation index.

Long-term time series forecasting plays a crucial role in fields such as weather, electricity, and healthcare. However, traditional methods for long-term time series forecasting still face challenges in handling long-term dependencies and nonlinear complex patterns. Pre-trained models are considered state-of-the-art deep learning technology. Recent academic research indicates that these models excel in recognizing and reasoning over complex sequential data. They show significant potential in time series tasks, yet their accuracy in long-term time series forecasting, a cross-modal task, still requires improvement. We propose a novel model named GPTLTS, which integrates GPT-2 for feature extraction from time series data and incorporates multi-scale convolution techniques to enhance feature extraction capabilities. We validate this approach through experiments on multiple time series datasets, demonstrating its effectiveness in capturing complex patterns and long-term dependencies. Experimental results show that the GPTLTS model outperforms traditional methods in time series prediction tasks, achieving higher prediction accuracy and generalization capabilities.

High-dimensional parametric models and large-scale mathematical computations limit the efficiency of IoT devices. The emergence of serverless edge computing provides a solution where users can deploy models as serverless functions and delegate provisioning and scaling to the platform. However, the resource-constrained nature of edge resources leads to their inefficiency or inability to serve large neural networks. Therefore, this paper emphasizes collaborative inference between devices and serverless edge computing for DNN models. Specifically, we consider the partitioning and offloading of DNN models and propose a DRL-based strategy to learn the joint decision of model partitioning and function memory type selection to achieve a cost-optimal service that meets the SLO requirements. In addition, we introduce a simulated annealing algorithm in deep reinforcement learning that focuses on exploration in the early stage and empirical value in the later stage. Finally, experimental results show that our algorithm can satisfy various SLOs at a low service cost and outperform three benchmarking strategies.

Currently, Coronavirus Disease 2019 (COVID-19) is still endangering world health and safety and deep learning (DL) is expected to be the most powerful method for efficient detection of COVID-19. However, patients' privacy concerns prohibit data sharing between medical institutions, leading to unexpected performance of deep neural network (DNN) models. Fortunately, federated learning (FL), as a novel paradigm, allows participating clients to collaboratively train models without exposing source data outside original location. Nevertheless, the current FL-based COVID-19 detection methods prefer optimizing secondary objectives including delay, energy consumption and privacy, while few works focus on improving the model accuracy and stability. In this paper, we propose a federated learning framework with dynamic focus for COVID-19 detection on CXR images, named FedFocus. Specifically, to improve the training efficiency and accuracy, the training loss of each model is taken as the basis for parameter aggregation weights. As training layer deepens, a constantly updated dynamic factor is designed to stabilize the aggregation process. In addition, to highly restore the real dataset, the training sets in our experiments are divided based on the population and the infection of three real cities. Extensive experiments conducted on the real-world CXR images dataset demonstrate that FedFocus outperforms the baselines in model training efficiency, accuracy and stability.

Cervical cancer, a prevalent malignancy among women, demands early screening and diagnosis for improved cure rates. Computer-aided screening systems offer accurate and rapid detection, reducing errors and enhancing efficiency. Convolutional neural networks have been extensively employed for cervical cancer cell detection in TCT images. However, these methods often overlook cervical cell morphology and intricate TCT image backgrounds, resulting in biased models. Moreover, their complexity limits practical application efficiency. Nevertheless, the causal attention mechanism can confine the model’s focus solely to the current morphology, preventing interference from other factors and effectively mitigating bias. And, lightweight strategies can enhance model efficiency and performance by reducing parameter count and computational demands. Thus, we introduce a lightweight high-performance model based on causal attention. Our cervical cancer cell detection model integrates causal attention-guided deformable convolutions in its backbone. This novel backbone improves CNN’s capacity to extract cellular morphological features, alleviating bias from target-background correlation. Additionally, a lightweight GSConv structure streamlines the neck architecture, further boosting efficiency. We also introduce a Grad-CAM-based explanation method for swift insight into model predictions. Experimental results demonstrate heightened precision, reducing parameters and computation by 20%, with a 1.9% average mAP increase. Our method reduces complexity and enhances precision in cervical cancer cell detection, showcasing its efficacy and potential in medical image object detection.

Real-time tracking and segmentation of ultrasound video sequence are prerequisite for identifying and analyzing lesions. While significant progress has been made in natural video object segmentation, developing a model for ultrasound video is still challenging due to problems such as low distinguishability and low visual saliency of the target objects, large variation between adjacent frames. These challenges are inherently complex and cannot be effectively tackled through a single process. This paper develops an end-to-end multi-stage network (EMNet) for ultrasound video object segmentation. EMNet consists of two stages. The inital mask generation stage comprises a contrast-enhanced layer to enhance visual contrast between targets and backgrounds. In this stage, a module that adopts the encoder-attention-decoder structure is designed for mask induction. After obtaining the initial segmentation mask, the mask refinement stage is followed to further improve initial segmentation. To prevent the propagation of errors, a gating mechanism is designed to control the fusion of segmentation probability maps in the initial and refinement stages. By transforming certain fixed parameters in different stage into trainable parameters and establishing an end-to-end learning process, we optimized the performance of our approach. We evaluate EMNet on real-world lymphoma ultrasound video dataset. Compared with the best results among seven competing baselines, EMNet achieves the best performance in terms of ℐ&ℱ and ℱ measures, the second-best performance with Param and FPS measures, which demonstrates the competitive performance in terms of both speed and accuracy.

Convolutional neural networks have achieved great success in the field of image recognition, however, their inherent black-box properties hinder their application in highly sensitive fields such as healthcare and autonomous driving. Therefore, interpretable artificial intelligence (XAI) has become a hot topic of academic research, which aims at exploring the working mechanism inside deep convolutional neural networks. Causality is an important means to explore the working mechanism of neural networks. For image classification, the image classification results depend on the causal features or background features of the input image. Grad-CAM is a popular interpretation method to visualize these features, and its visualization results in a deep integration of causal and context features. In this paper, we propose a background category-based feature integration method to estimate the context features, and then apply an ensemble theory approach to extract causal features from Grad-CAM. The experimental results show that our Grad-CAM-based extracted causal features not only have better causality in ResNet50, VGG16, and GoogleNet compared with Grad-CAM and Grad-CAM++, but also are similar to the visualization results of Grad-CAM ++ in terms of target object localization of visualization results. In addition, we verify that our method is applicable not only to high category fine-grained datasets but also to low fine-grained datasets through category fine-grained experiments.

Cervical cancer is one of the most common malignant tumors in women. Clinical observations have demonstrated that early screening and treatment can prevent, detect, and cure cervical cancer. TCT image detection has been widely used in cervical disease screening, yet the numerous cell counts in TCT image samples make traditional manual diagnostic screening consume a significant amount of human resources. Accurately and rapidly classifying TCT images is essential for cervical cancer screening. This paper proposed a MobileNet-Light model for cervical cancer TCT image classification based on lightweight design. The Ghost module was introduced in the lightweight model design to optimize redundant image features in the network. Moreover, the model design included an efficient channel attention mechanism based on ECA, which improved the classification accuracy while significantly reducing the parameter count. Finally, to address the black-box problem of the classification model, we used the Grad-Cam based interpretable method to provide some interpretability to the model's classification results as diagnostic basis for doctors. Experimental results demonstrated that our lightweight classification model remarkably performs in cervical cancer TCT image classification tasks.

In recent years, with the rapid development of mass spectrometry, there are more and more cases of finding biomarkers by proteomics for clinical diagnosis and treatment, and artificial intelligence is gradually replacing traditional statistical learning methods, but the process of finding biomarkers is not yet uniform and perfect, and there are gaps in validation. With this aim, this study proposes a pipeline for the analysis of biomarkers based on data independent acquisition(MS-BioAP). MS-BioAP includes three modules: mass spectrometry protein identification module, differential protein analysis module, and classification model building module. Combining statistical learning and artificial intelligence methods to complete the biomarker search, we also propose a target association scoring algorithm(TAScore), use the large amount of prior knowledge in the knowledge graph to complete the biomarker validation, and create a classification model based on the selected biomarkers. We conducted experiments using a clinical stress response mass spectrometry dataset and the final selected biomarkers achieved an accuracy of 1.0 in the classification model with both high sensitivity and specificity. The analytical pipeline is complete and comprehensive, with clinical implications. In the future, we will continue to optimize the analytical pipeline and validate this analytical pipeline with additional datasets.

Breast cancer has become one of the most common malignant tumors in women worldwide, and it seriously threatens women’s physical and mental health. In recent years, with the development of Artificial Intelligence(AI) and the accumulation of medical data, AI has begun to be deeply integrated with mammography, MRI, ultrasound, etc. to assist physicians in disease diagnosis. However, the existing breast cancer diagnosis model based on Computer Vision(CV) is greatly affected by the image quality; on the other hand, the breast cancer diagnosis model based on Natural Language Processing(NLP) cannot effectively extract the semantic information of the mammography report. The lack of model interpretability also makes the existing diagnostic models have low confidence. In this paper, we proposed Breast Cancer Causal XAI Diagnostic Model(BCCXDM). Specifically, we first structured the mammography report. Then find the causal graph based on the structured table. We combine the existing tabular learning method TabNet with causal graphs(Causal-TabNet) to enable reasoning in the graphs to preserve the correlation between features. More importantly, we use GNN and node transition probability to aggregate node information. We evaluate our model on the real-world mammography report, and compare it with other popular interpretable methods. The experimental results show that our interpretable results are closer to the diagnostic criteria of clinicians.

Pneumonia kills more than 1.4 million children per year. It is among the top diseases which caused many deaths around the world. Because the diagnosis by chest x-ray films is rather complicated, the accuracy of diagnosing pneumonia is not high even by expert radiologists. In this study, the implementation of a deep learning technique for chest x-rays is presented. For this, we have proposed a new method to simplify the process of pneumonia detection. A convolutional neural network (CNN) model “ConvNet-21” has been proposed to extract important features from the chest x-ray images. The dataset is divided into two categories: training and testing, with each category further subdivided into normal and pneumonia. Several data augmentation techniques are applied to the dataset. In this study, five different models have been discussed. There are five convolutional layers in the proposed CNN model and the remaining four are transfer learning models, VGG16, VGG19, Inception-V3, and ResNetl52V2. We have analyzed the performance of all these models by using different performance measure parameters. The accuracy obtained by our proposed CNN model is SS.94%. The accuracy obtained by VGG16, VGG19, Inception-V3 and ResNetl52V2 are 92.7S%, 91.50%, 74.19% and 77.72%, respectively.

Alzheimer’s Disease (AD) is an irreversible neurodegenerative disease that commonly occurs in the elderly. With the current accelerated aging process, the accurate diagnosis of early AD is essential for patient care and disease delay. In recent years, Magnetic Resonance Imaging (MRI) has become increasingly important in diagnosing AD due to advances in deep learning and neuroimaging technology. This paper proposes a model framework for multi-classification prediction of Alzheimer’s disease based on fusing multi-modal features. Firstly, the sMRI data are pre-processed based on ROI templates with different segmentation accuracy levels to extract morphological features including gray matter volume, surface area and cortical thickness, and then these features are combined with the corresponding Clinical Data to produce the Indicators dataset. Secondly, a 3DCNN-SE module is proposed to extract the primary features from the 3D MRI data. In order to reduce the dimensionality of the Indicators, an Indicator Selection Strategy is designed to select the most relevant features from the Indicators. Finally, a Multi-Attention-Fusion Module (MAFM) is developed to perform multi-modal data fusion on the results of feature extraction and selection, followed by a SoftMax classifier for AD disease diagnosis. We evaluated 596 patients from the Alzheimer’s Disease Neuroimaging Initiative(ADNI), including 198 patients with Alzheimer’s disease (AD), 200 patients with mild cognitive impairment (MCI), and 198 patients cognitively normal(CN). As a result, 88% accuracy is achieved on the three classifications, which is better than the related methods mentioned in literature.

Breast tumor classification with multiple medical reports such as B-ultrasound, Mammography (X-ray) and Nuclear Magnetic Resonance Imaging (MRI) is crucial to the intelligent cancer diagnosis system. Unlike the other domain texts, the medical reports have latent hierarchical syntactic structures and have hidden rich semantic information about the entities and relationships, which poses a great challenge of breast cancer classification. In this article, we proposed a Knowledge-powered Deep Breast Tumor Classification model (KDBTC), which takes the semantic information as a kind of prior knowledge and incorporated it into deep neural networks. Specially, our proposed model first uses Hierarchical Attention Bidirectional Recurrent Neural Networks (HA-BiRNNs) to encode the syntax-aware representation of medical reports in a hierarchical way. In the HA-BiRNN, a hierarchical neural network structure, consisting in two encoder layers of BiRNN (Bidirectional Recurrent Neural Networks), mirrors the hierarchical structure of medical reports, and a hierarchical attention mechanism, consisting of two levels attentions, attends to important elements within clinical report with word-level attention and sentence-level attention. Secondly, our model obtains the semantic information relevant to the medical reports from the clinical domain semantic tree, and encodes the semantic representation of medical reports by using Tree Structured Recurrent Neural Network with gated recursive units (Tree-GRUs). Finally, we classify breast tumors by combining both the syntax and semantic representations of medical reports. We evaluate our method on the real-world breast cancer medical reports, and results show that our method achieves higher performance on breast cancer classification.

Coronary heart disease is the first killer of human health. At present, the common way of coronary heart disease diagnosis is coronary angiography. This method is a kind of surgery that could cause some physical damage to the patient, and it could cause some complications and adverse reactions. Furthermore, coronary angiography is expensive. However, the heart color Doppler echocardiography report, blood biochemical indicators and other basic information can reflect the degree of heart damage in patients. Therefore, this paper proposes a combined reinforcement multitask progressive networks (CRMPN) model to predict the grade of coronary heart disease through heart color Doppler echocardiography report, blood biochemical indicators and ten basic body information items about the patient. In this model, the first step is performing deep reinforcement learning (DRL) pre-training through asynchronous advantage actor-critic (A3C). Training data is adopted to optimize the recurrent neural networks (RNN) that parameterizes the stochastic policy. In the second step, we use soft parameter sharing module, hard parameter sharing module and progressive deep network to predict coronary heart disease. The experimental results show that after DRL pre-training, the multiple tasks in the model interact with each other and learn together to achieve satisfactory results and outperform other state-of-the-art methods.

A pharmacological effect of a drug on cells, organs and systems refers to the specific biochemical interaction produced by a drug substance, which is called its mechanism of action. Drug repositioning (or drug repurposing) is a fundamental problem for the identification of new opportunities for the use of already approved or failed drugs. In this paper, we present a method based on a multi-relation unsupervised graph embedding model that learns latent representations for drugs and diseases so that the distance between these representations reveals repositioning opportunities. Once representations for drugs and diseases are obtained we learn the likelihood of new links (that is, new indications) between drugs and diseases. Known drug indications are used for learning a model that predicts potential indications. Compared with existing unsupervised graph embedding methods our method shows superior prediction performance in terms of area under the ROC curve, and we present examples of repositioning opportunities found on recent biomedical literature that were also predicted by our method.

Extracting information from the medical imaging reports with domain ontology has attracted much attention in medical natural language processing field. Based on the unstructured characteristics of medical image report text, the existing image report domain ontology is constructed by ontology learning method, including ontology learning method based on language and ontology learning method based on machine learning. However, these existing methods ignore the anatomical knowledge embedded in medical imaging reports, which is very useful for extracting entity relationships from reports. In order to solve the above problems, this paper proposes a domain ontology construction method for medical image reporting based on anatomy knowledge, which combines the prior knowledge of pathology and anatomy to obtain the basic framework of the domain ontology as the knowledge driver. In particular, our proposed approach consists of two tasks. The first task is to convert each text report into a semantic tree through an anatomic knowledge based semantic subtree generation algorithm. Semantic subtree generation algorithm mainly consists of three parts: framework positioning, relation extraction and adding relation. Firstly, the text is located based on anatomical knowledge, and the relational extraction mainly adopts the method of dependency syntactic analysis to extract the three semantic relations contained in the text. Finally, add the relationship to the corresponding branch. The second task is to obtain the domain ontology by merging semantic subtree. This step is mainly to obtain the domain ontology by merging the nodes of the XML structure semantic tree. The experimental results show that this method can be used for relational extraction and domain ontology construction, laying a good foundation for the follow-up research.

Summary Entity recognition plays an important role in building the electronic medical records (EMRs) based medical knowledge graph, which is significant for building Clinical decision support (CDS) system. Cross‐disease clinical documents are context‐related and have different interrelated semantic structures, which bring challenges for entity recognition using traditional methods. In order to solve these problems, this paper proposes a co‐training based entity recognition approach for cross‐disease clinical documents. In this model, we first build partial annotation corpus of the single disease using dependency syntax analysis and the medical statement rule unifies. Then, according to the partial annotation corpus of different diseases, the sentence level features are extracted through the Bi‐LSTM layer with memory unit and CRF methods, which optimize the whole sequence and improve the combination probability of sequence labels. Finally, the results with higher confidence are selected by cross feedback to label the corpus, which enlarges the size of corpus and improves the accuracy of the document entity recognition. The experiment result proves the availability and high efficiency of our method.

With the development of information technology, most of the medical institutions have established the web medical platform, and it will have a large number of patients' evaluating textual information. These subjective text messages contain emotional information, such as views, opinions and attitudes of patients. It takes a lot of manpower and time to analyze and evaluate positive and negative evaluations by manual methods. Therefore, this paper presents a method of emotion evaluation of medical reviews based on character-level vector convolution neural network. Aiming at the problem of text input noise caused by word segmentation and polysemy, ignoring the structure information of sentence in traditional convolution neural network model, this paper proposes a segmentation pooling convolution neural network model based on character-level vector. Using the improved skip-gram model to train the character vector, and using different convolution kernels of different sizes to extract the sentence features, this proposed model then use the method of segmentation pooling to preserve the maximum eigenvalues of each part ofthe sentence. The experiments show that the accuracy of the proposed model in the emotional analysis of medical texts is about 12% higher than that ofthe traditional convolutional neural network model. In the actual task of emotion analysis of medical texts, the accuracy of the model is as high as 88.2 %.

This paper designs and implements a NoSQL database called STNoSQL for Internet of Things on top of SensibleThings, which is a distributed platform enabling devices to communicate with each other. In the proposed method, a novel key-value design strategy is provided. Based on it, STNoSQL implements efficient range query execution without the waste of storage space, which deals with low-efficient range query in most NoSQL databases. In addition, STNoSQL provides load balance which improves the reliability and scalability of database by making a comparison between different methods during the implementation. Furthermore, in order to overcome the limitation on security in most NoSQL databases, a security mechanism including access control is proposed. Finally, a proof-of-concept application based on a specific scenario is completed and evaluations are made based on the results.

Wireless sensor networks (WSNs) have gained much attention in both industry and research communities where they are expected to bring the interaction between humans, environment, and machines to a new level. Due to the resource constraints of sensors nodes, it is infeasible to use traditional key establishment techniques that find use in fixed communication systems. In recent years a number of group key agreement protocols have been proposed for resourcelimited wireless sensor devices. However, these protocols do not satisfy some important security properties such as mutual authentication and forward secrecy. In this paper, we propose a hybrid authenticated group key agreement protocol for WSNs. This hybrid protocol reduces the high cost public-key operations at the sensor side and replaces them with efficient symmetric-key based operations. In order to provide assurance that the proposed protocol is verifiably secure and trustworthy, a formal verification is performed on the protocol's design specification.