Cardiovascular diseases and psychological disorders represent two major threats to human physical and mental health. Research on electrocardiogram (ECG) signals offers valuable opportunities to address these issues. However, existing methods are constrained by limitations in understanding ECG features and transferring knowledge across tasks. To address these challenges, this study developed a multi-resolution feature encoding network based on residual networks, which effectively extracted local morphological features and global rhythm features of ECG signals, thereby enhancing feature representation. Furthermore, a model compression-based continual learning method was proposed, enabling the structured transfer of knowledge from simpler tasks to more complex ones, resulting in improved performance in downstream tasks. The multi-resolution learning model demonstrated superior or comparable performance to state-of-the-art algorithms across five datasets, including tasks such as ECG QRS complex detection, arrhythmia classification, and emotion classification. The continual learning method achieved significant improvements over conventional training approaches in cross-domain, cross-task, and incremental data scenarios. These results highlight the potential of the proposed method for effective cross-task knowledge transfer in ECG analysis and offer a new perspective for multi-task learning using ECG signals.
Humans ; Electrocardiography/methods* ; Mental Health ; Algorithms ; Signal Processing, Computer-Assisted ; Machine Learning ; Arrhythmias, Cardiac/diagnosis* ; Cardiovascular Diseases ; Neural Networks, Computer ; Mental Disorders

Deep learning method can be used to automatically analyze electrocardiogram (ECG) data and rapidly implement arrhythmia classification, which provides significant clinical value for the early screening of arrhythmias. How to select arrhythmia features effectively under limited abnormal sample supervision is an urgent issue to address. This paper proposed an arrhythmia classification algorithm based on an adaptive multi-feature fusion network. The algorithm extracted RR interval features from ECG signals, employed one-dimensional convolutional neural network (1D-CNN) to extract time-domain deep features, employed Mel frequency cepstral coefficients (MFCC) and two-dimensional convolutional neural network (2D-CNN) to extract frequency-domain deep features. The features were fused using adaptive weighting strategy for arrhythmia classification. The paper used the arrhythmia database jointly developed by the Massachusetts Institute of Technology and Beth Israel Hospital (MIT-BIH) and evaluated the algorithm under the inter-patient paradigm. Experimental results demonstrated that the proposed algorithm achieved an average precision of 75.2%, an average recall of 70.1% and an average F ₁-score of 71.3%, demonstrating high classification accuracy and being able to provide algorithmic support for arrhythmia classification in wearable devices.
Humans ; Arrhythmias, Cardiac/diagnosis* ; Algorithms ; Electrocardiography/methods* ; Neural Networks, Computer ; Signal Processing, Computer-Assisted ; Deep Learning ; Classification Algorithms

Fatigue driving is one of the leading causes of traffic accidents, posing a significant threat to drivers and road safety. Most existing methods focus on studying whole-brain multi-channel electroencephalogram (EEG) signals, which involve a large number of channels, complex data processing, and cumbersome wearable devices. To address this issue, this paper proposes a fatigue detection method based on frontal EEG signals and constructs a fatigue driving detection model using an asymptotic hierarchical fusion network. The model employed a hierarchical fusion strategy, integrating an attention mechanism module into the multi-level convolutional module. By utilizing both cross-attention and self-attention mechanisms, it effectively fused the hierarchical semantic features of power spectral density (PSD) and differential entropy (DE), enhancing the learning of feature dependencies and interactions. Experimental validation was conducted on the public SEED-VIG dataset. The proposed model achieved an accuracy of 89.80% using only four frontal EEG channels. Comparative experiments with existing methods demonstrate that the proposed model achieves high accuracy and superior practicality, providing valuable technical support for fatigue driving monitoring and prevention.
Humans ; Electroencephalography/methods* ; Automobile Driving ; Fatigue/diagnosis* ; Accidents, Traffic/prevention & control* ; Signal Processing, Computer-Assisted ; Neural Networks, Computer ; Algorithms ; Prefrontal Cortex/physiology*

Medical visual question answering (MVQA) plays a crucial role in the fields of computer-aided diagnosis and telemedicine. Due to the limited size and uneven annotation quality of the MVQA datasets, most existing methods rely on additional datasets for pre-training and use discriminant formulas to predict answers from a predefined set of labels. This approach makes the model prone to overfitting in low resource domains. To cope with the above problems, we propose an image-aware generative MVQA method based on image caption prompts. Firstly, we combine a dual visual feature extractor with a progressive bilinear attention interaction module to extract multi-level image features. Secondly, we propose an image caption prompt method to guide the model to better understand the image information. Finally, the image-aware generative model is used to generate answers. Experimental results show that our proposed method outperforms existing models on the MVQA task, realizing efficient visual feature extraction, as well as flexible and accurate answer outputs with small computational costs in low-resource domains. It is of great significance for achieving personalized precision medicine, reducing medical burden, and improving medical diagnosis efficiency.
Humans ; Image Processing, Computer-Assisted/methods* ; Diagnosis, Computer-Assisted/methods* ; Algorithms ; Telemedicine

To address the current issues of data imbalance and scarcity in photoplethysmography (PPG) data for type 2 diabetes mellitus (T2DM) prediction, this study proposes an improved conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP). The algorithm integrated gated recurrent unit (GRU) networks and self-attention mechanisms to construct a generator, aiming to produce high-quality PPG signals. Various data augmentation methods, including the improved CWGAN-GP, were employed to expand the PPG dataset, and multiple classifiers were applied for T2DM prediction analysis. Experimental results showed that the model trained on data generated by the improved CWGAN-GP achieved the optimal prediction performance. The highest accuracy reached 0.895 0, and compared with other data enhancement methods, this approach exhibited significant advantages in terms of precision and F1-score. The generated data notably enhances the accuracy and generalization ability of T2DM prediction models, providing a more reliable technical basis for non-invasive early T2DM screening based on PPG signals.
Photoplethysmography/methods* ; Diabetes Mellitus, Type 2/diagnosis* ; Humans ; Algorithms ; Neural Networks, Computer ; Signal Processing, Computer-Assisted ; Prediction Algorithms

Lung cancer is a leading cause of cancer-related deaths worldwide, with its high mortality rate primarily attributed to delayed diagnosis. Radiomics, by extracting abundant quantitative features from medical images, offers novel possibilities for early diagnosis and precise treatment of lung cancer. This article reviewed the latest advancements in radiomics for lung cancer management, particularly its integration with artificial intelligence (AI) to optimize diagnostic processes and personalize treatment strategies. Despite existing challenges, such as non-standardized image acquisition parameters and limitations in model reproducibility, the incorporation of AI significantly enhanced the precision and efficiency of image analysis, thereby improving the prediction of disease progression and the formulation of treatment plans. We emphasized the critical importance of standardizing image acquisition parameters and discussed the role of AI in advancing the clinical application of radiomics, alongside future research directions.
Humans ; Lung Neoplasms/diagnosis* ; Artificial Intelligence ; Early Detection of Cancer/methods* ; Precision Medicine ; Image Processing, Computer-Assisted/methods* ; Tomography, X-Ray Computed ; Radiomics

Thyroid nodules are a common endocrine disorder, and their early detection and accurate diagnosis are crucial for the prevention of thyroid cancer. However, the highly heterogeneous morphology and boundaries of thyroid nodules pose significant challenges to their precise identification and classification. Traditional diagnostic approaches rely heavily on physicians' experience, which increases the risk of misdiagnosis and missed diagnoses. With the rapid advancement of computer-aided diagnosis (CAD) technologies, applying deep learning algorithms to the analysis of thyroid nodule ultrasound images has shown great potential. This paper reviews the latest research progress on deep learning-based CAD methods for thyroid nodules, with a focus on their applications in image preprocessing, segmentation and classification. The advantages and limitations of current techniques are analyzed, and potential future directions are discussed. This review aims to highlight the potential of deep learning in thyroid nodule diagnosis and to provide a foundation for selecting feasible pathways for future clinical applications.
Humans ; Thyroid Nodule/diagnostic imaging* ; Deep Learning ; Ultrasonography/methods* ; Diagnosis, Computer-Assisted/methods* ; Algorithms ; Thyroid Neoplasms/diagnostic imaging* ; Image Processing, Computer-Assisted/methods*

OBJECTIVES: We propose a new melanoma segmentation model, SG-UNet, to enhance the precision of melanoma segmentation in dermascopy images to facilitate early melanoma detection. METHODS: We utilized a U-shaped convolutional neural network, UNet, and made improvements to its backbone, skip connections, and downsampling pooling sections. In the backbone, with reference to the structure of VGG, we increased the number of convolutions from 10 to 13 in the downsampling part of UNet to achieve a deepened network hierarchy that allowed capture of more refined feature representations. To further enhance feature extraction and detail recognition, we replaced the traditional convolution the backbone section with self-calibrated convolution to enhance the model's ability to capture both spatial and channel dimensional features. In the pooling part, the original pooling layer was replaced by Haar wavelet downsampling to achieve more effective multi-scale feature fusion and reduce the spatial resolution of the feature map. The global attention mechanism was then incorporated into the skip connections at each layer to enhance the understanding of contextual information of the image. RESULTS: The experimental results showed that the SG-UNet model achieved significantly improved segmentation accuracy on ISIC 2017 and ISIC 2018 datasets as compared with other current state-of-the-art segmentation models, with Dice reached 92.41% and 86.62% and IoU reaching 92.31% and 86.48% on the two datasets, respectively. CONCLUSIONS The proposed model is capable of effective and accurate segmentation of melanoma from dermoscopy images.
Melanoma/diagnosis* ; Humans ; Neural Networks, Computer ; Dermoscopy/methods* ; Skin Neoplasms ; Image Processing, Computer-Assisted/methods* ; Calibration ; Algorithms

OBJECTIVES: To construct a bone tumor classification model based on feature decoupling and fusion for processing modality loss and fusing multimodal information to improve classification accuracy. METHODS: A decoupling completion module was designed to extract local and global bone tumor image features from available modalities. These features were then decomposed into shared and modality-specific features, which were used to complete the missing modality features, thereby reducing completion bias caused by modality differences. To address the challenge of modality differences that hinder multimodal information fusion, a cross-attention-based fusion module was introduced to enhance the model's ability to learn cross-modal information and fully integrate specific features, thereby improving the accuracy of bone tumor classification. RESULTS: The experiment was conducted using a bone tumor dataset collected from the Third Affiliated Hospital of Southern Medical University for training and testing. Among the 7 available modality combinations, the proposed method achieved an average AUC, accuracy, and specificity of 0.766, 0.621, and 0.793, respectively, which represent improvements of 2.6%, 3.5%, and 1.7% over existing methods for handling missing modalities. The best performance was observed when all the modalities were available, resulting in an AUC of 0.837, which still reached 0.826 even with MRI alone. CONCLUSIONS The proposed method can effectively handle missing modalities and successfully integrate multimodal information, and show robust performance in bone tumor classification under various complex missing modality scenarios.
Humans ; Bone Neoplasms/diagnosis* ; Multimodal Imaging/methods* ; Magnetic Resonance Imaging ; Tomography, X-Ray Computed ; Image Processing, Computer-Assisted/methods* ; Algorithms

OBJECTIVES: We propose a myocardial infarction (MI) detection and localization model for improving the diagnostic accuracy for MI to provide assistance to clinical decision-making. METHODS: The proposed model was constructed based on multi-scale field residual blocks fusion modified channel attention (MSF-RB-MCA). The model utilizes lead II electrocardiogram (ECG) signals to detect and localize MI, and extracts different levels of feature information through the multi-scale field residual block. A modified channel attention for automatic adjustment of the feature weights was introduced to enhance the model's ability to focus on the MI region, thereby improving the accuracy of MI detection and localization. RESULTS: A 5-fold cross-validation test of the model was performed using the publicly available Physikalisch-Technische Bundesanstalt (PTB) dataset. For MI detection, the model achieved an accuracy of 99.96% on the test set with a specificity of 99.84% and a sensitivity of 99.99%. For MI localization, the accuracy, specificity and sensitivity were 99.81%, 99.98% and 99.65%, respectively. The performances of the model for MI detection and localization were superior to those of other comparison models. CONCLUSIONS The proposed MSF-RB-MCA model shows excellent performance in AI detection and localization based on lead II ECG signals, demonstrating its great potential for application in wearable devices.
Myocardial Infarction/diagnosis* ; Humans ; Electrocardiography/methods* ; Signal Processing, Computer-Assisted ; Algorithms ; Sensitivity and Specificity