Gait recognition is an important technology to perceive human motion patterns, which is widely used in intelligent medical, security monitoring and human-computer interaction. Traditional gait recognition methods mostly rely on image analysis or wearable sensors, which are restricted by environmental conditions and difficult to balance accuracy and convenience. This paper constructs an intelligent carpet gait recognition system using Optical Frequency Domain Reflectometry (OFDR) based distributed fiber optic sensing. By embedding two-dimensional optical fiber array inside the carpet, it achieves high-density perception of strains of the foot. A local strain and global contour dual-modality attention network is proposed for gait recognition. Firstly, a strain graph attention network is designed to process the original strain data, which can construct the dynamic correlation between fibers and extract key fiber strain features. Then, the strain signal is transformed into image and a pressure gradient guided non-local attention module is introduced to focus on the contact area and contour of the sole, so as to enhance the global spatial features of the gait. Finally, the dual-modality features are fused to realize gait recognition. The proposed method is tested on 16 kinds of gaits. The quantitative and qualitative experiment results show that the proposed method can effectively collect gait data and has excellent recognition accuracy.

Federated learning (FL) faces significant challenges in medical image analysis due to data heterogeneity among clients, where balancing personalization and generalization is challenging. Existing methods often struggle to achieve both objectives simultaneously, as excessive personalization reduces generalization, while over-generalization weakens adaptation to client-specific features. To address these challenges, we propose a Perceptual Hashing Hypernetwork Personalized Federated Learning (PHH-FL) to enhance both personalization and generalization. PHH-FL first uses a perceptual hashing algorithm to construct a similarity matrix that captures data distribution differences among clients and employs a hypernetwork to generate personalized parameters for each client. Meanwhile, the shared hypernetwork is introduced to promote knowledge transfer between clients, thereby enhancing the generalization ability of the local model. By selectively generating parameters for the initial layers of the target network, PHH-FL reduces computational and communication costs while maintaining performance. Experiments on medical image classification and segmentation tasks show that PHH-FL outperforms state-of-the-art methods. Ablation studies further demonstrate that the proposed framework effectively balances personalization and generalization.

The operating conditions in complex industrial processes are often dynamic, unpredictable, and difficult to label, leading to significant challenges in predicting industrial quality indices. To address these challenges and improve prediction accuracy under Multiple Operating Conditions (MOC), we propose a Coarse-Fine Meta-learning Framework (CFMF). Initially, multiple Coarse Models are established using historical operating condition data. We then introduce a MOC-Dynamic Time Warping (DTW) strategy, which utilizes small-batch data from new operating conditions to identify similar time-series characteristics from historical conditions. These similar conditions data are used to train a meta-learning model for the Coarse Models based on Stacking, ultimately resulting in a fine model for quality index prediction. In industrial experiments, we compare the CFMF with classical multi-model learning strategies, and the results demonstrate that the proposed CFMF achieves superior prediction performance on the target domain test set.

Ultrasound is an important routine screening modality for breast cancer. Breast ultrasound screening is a dynamic process, and clinical practice involves radiologists recording representative frames during dynamic breast scanning for subsequent diagnosis. However, existing computer-assisted diagnosis methods often concentrate on dull diagnostic results by analyzing these representative frames and ignore the valuable information in the dynamic examination process that facilitates diagnosis. Moreover, breast lesions could exhibit various characteristic differences during scanning, and effective learning of lesion representations is challenging and may affect the clinical interpretability of the methods. To this end, we draw insights from the behavior of radiologists during the dynamic breast examination and leverage the knowledge of breast anatomy to propose a clinical knowledge-aware framework for lesion detection and classification of breast lesions in ultrasound videos. It is equipped with global-local attentive aggregation and a dynamic allocation mechanism that simulates the behavior of radiologists searching for diagnostic clues, thus integrating local localization and global semantic information from the video into the feature representation of the lesion. An anatomically-aware transformer is also designed to refine the lesion feature representation using spatial relationships within and across different anatomical layers of the breast anatomy. Extensive experiments show that the proposed framework can achieve competitive performance in both lesion detection and video classification tasks while exhibiting good clinical availability and interpretability, with an average precision of 40.80% and an AUC of 85.86% on our constructed breast video dataset and an average precision of 39.79% and an AUC of 87.04% on a publicly available dataset.

Multivariate time series forecasting estimates future development by capturing variable relationships and constructing temporal regular, which is widely used in many scenarios, including industrial production, economic development, and disease prediction. Although the existing deep learning methods have achieved impressive results in multivariate time series forecasting tasks, the existing methods only emphasize the prediction performance and ignore the widespread issue of missing data in the real world. This article proposes a robust multivariate correlation self-distillation Transformer framework for incomplete time series data forecasting. The proposed method first decouples the interinference of historical series and the exter-inference of future series into two stages. The first stage focuses on the reconstruction of historical series, while the second stage focuses on the prediction of future series. Then, a novel multivariate correlation Transformer is designed as the basic component of the network, which can perform feature inference from both multivariate relationships and single-variate temporal regular. Finally, a variable correlation self-distillation method is proposed to self-distill the more complete variable relationship from the exter-inference stage to the interinference stage. The proposed method is verified on eight real-world datasets, and both qualitative and quantitative results show that the proposed method has good performance.

Quality prediction is essential for optimizing operations and making timely decisions in industrial processes. However, the dynamic nature of industrial data, characterized by different sampling periods, presents significant challenges for quality prediction. The relationship between quality indices and various industrial data with differing sampling periods is complex and dynamically correlated in both spatial and temporal dimensions. To address this issue, we propose an adaptive multiresampling learning (AMRL) network that performs deep spatial-temporal feature mining and extraction from dual-scale data, facilitating multistep industrial quality prediction. The AMRL network leverages principal component scores of fast-scale process data and an adaptive multiresampling module to construct multiple fast-scale resampling channels adaptively. The cross-scale deep convolutional neural network and the multichannel self-attention module are then employed to capture spatial-temporal features and selectively focus on critical regions within the multiresampling sequences. We compared and evaluated the proposed method against eight state-of-the-art methods using real industrial datasets. The comparison results demonstrate the superior performance of the AMRL in multistep industrial quality prediction.

Accurate remaining useful life (RUL) prediction of lithium-ion batteries is critical for energy supply systems. In conventional data-driven RUL prediction approaches, the battery's degradation mechanism is difficult into incorporate in the RUL prediction. Furthermore, there are notable limitations in reflecting the significance of different time instances, and the uncertainty in the degradation process. Consequently, a novel data-model interactive RUL prediction approach based on particle filter-temporal attention mechanism-bidirectional gated recurrent unit (PF-BiGRU-TSAM) is proposed. Specifically, BiGRU-TSAM is trained offline through historical data, which assigns corresponding significance to battery capacities at different time instances. Moreover, regarding the interactive data-model for the online prediction phase based on PF-BiGRU-TSAM, the advantages of data-driven and model-based approaches are integrated, which accomplishes the purpose of modifying each other. The proposed PF-BiGRU-TSAM approach is validated with a real-world battery dataset. Experimental results demonstrate the proposed approach is better than some published approaches. Taking the 50th operational cycle of the four batteries B0005, B0006, B0007, and B0018 in the dataset as an instance, the absolute errors of the proposed PF-BiGRU-TSAM are 0, 1, 3, 3, respectively, which represents the proposed approach has an excellent performance.

Knee segmentation and landmark localization from 3D MRI are two significant tasks for diagnosis and treatment of knee diseases. With the development of deep learning, Convolutional Neural Network (CNN) based methods have become the mainstream. However, the existing CNN methods are mostly single-task methods. Due to the complex structure of bone, cartilage and ligament in the knee, it is challenging to complete the segmentation or landmark localization alone. And establishing independent models for all tasks will bring difficulties for surgeon's clinical using. In this paper, a Spatial Dependence Multi-task Transformer (SDMT) network is proposed for 3D knee MRI segmentation and landmark localization. We use a shared encoder for feature extraction, then SDMT utilizes the spatial dependence of segmentation results and landmark position to mutually promote the two tasks. Specifically, SDMT adds spatial encoding to the features, and a task hybrided multi-head attention mechanism is designed, in which the attention heads are divided into the inter-task attention head and the intra-task attention head. The two attention head deal with the spatial dependence between two tasks and correlation within the single task, respectively. Finally, we design a dynamic weight multi-task loss function to balance the training process of two task. The proposed method is validated on our 3D knee MRI multi-task datasets. Dice can reach 83.91% in the segmentation task, and MRE can reach 2.12 mm in the landmark localization task, it is competitive and superior over other state-of-the-art single-task methods.

Retinal vessel image is an important biological information that can be used for personal identification in the social security domain, and for disease diagnosis in the medical domain. While automatic vessel image segmentation is essential, it is also a challenging task because the retinal vessels have complex topological structures, and the retinal vessels vary in size and shape. In recent years, image segmentation based on the deep learning technique has become a mainstream method. Unfortunately, the existing methods cannot make the best use of the global information, and the model complexity is high. In this article, a convolutional neural network integrated with the attention mechanism is proposed. The overall network structure consists of a basic U-Net and an attention module, and the latter is used to capture global information and to enhance features by placing it in the process of feature fusion. Experiment results on five public datasets show that the proposed scheme outperforms other existing mainstream approaches, and most of the performance indicators are in the leading positions. More importantly, the proposed method has a significant reduction in the number of parameters.

Skeletal fluorosis is a form of endemic disease caused by the excessive intake of fluoride. Bone deformation and periosteal calcification are the typical manifestations that can be observed in the images and are usually served as a basis of pathological grading. In the current medical systems, the diagnosis of skeletal fluorosis fully relies on doctors' knowledge and experience, and no research effort has been made in automatic image information diagnostic systems. According to the image information, the shape of the lesion is irregular, the boundary is unclear and the lesion related pixels only occupy a small part of the image. We take the lead in proposing a two-stage scheme that can achieve automated X-ray image diagnosis and severity grading. In the first stage, a Dense U-Net is proposed for reliable lesion determination, and a multitype feature fusion approach passes effective and comprehensive features to the subsequent stage. In the second stage, a novel classifier is designed with the integration of ensemble learning and multiple instance learning, which can ensure classification accuracy in case that the feature for diagnosis only takes up a small proportion of the whole image. Through plenty of experiments on the actual data collected from the hospitals, it is verified that the proposed strategy can achieve satisfactory results on skeletal fluorosis image diagnosis and severity grading.