Test-Time Adaptive Segmentation (TTA-Seg) aims to adapt trained models to unlabeled test data with shifted semantic distributions, which is critical for reliable scene understanding in autonomous driving under varying weather, illumination, and geographic conditions. Existing methods typically rely on momentum updates driven by model predictions to estimate target-domain distributions. However, such prediction-dependent strategies often fail to capture accurate semantic structures, leading to biased representations and error accumulation under large domain gaps in safety-critical scenarios. To address these challenges, we propose a semantic-aware prototype optimization framework for TTA-Seg. Specifically, we first extract instance-level mixture-of-prototypes from source data as frozen semantic priors, providing stable and transferable knowledge without requiring retraining. Based on this, target-domain semantics are dynamically modeled through feature–prototype alignment, enabling adaptive perception in continuously changing environments. Furthermore, we introduce learnable global prototypes to capture class-level relationships and model both intra-class variations and inter-class dependencies, thereby enhancing semantic consistency. Finally, we develop a consensus-guided weighting mechanism to adaptively balance local and global knowledge for stable and efficient deployment. Extensive experiments demonstrate that our method achieves state-of-the-art performance, improving the average mIoU by + 1.48%, +0.9%, and + 0.5% under Sim-to-Real TTA, Cityscapes-to-ACDC TTA, and Cityscapes-to-ACDC CTTA settings, respectively.

Hand fractures, particularly phalangeal fractures, are often subtle, concealed, and complex, posing significant challenges for accurate and efficient diagnosis with traditional radiography. To address the limitations of existing deep learning methods in handling complex anatomy, multi-view uncertainty, and causal discrimination, we propose a novel two-stage context-aware multi-view fusion framework for precise fracture classification and localization. In the first stage, a Context-Aware Multi-View Classification Network (CAMVC-Net) is developed. A Dual-path Semantic-Spatial Alignment (DSSA) module aligns low-level spatial details with high-level semantics, enhancing fine-grained fracture discrimination. A Context Aware Pyramid (CAP) module captures multi-scale context, while a Counterfactual Attention Learning (CAL) loss guides the network to focus on discriminative regions. Multi-view uncertainty is modeled using a Dirichlet distribution, and decision-level fusion is achieved by Dempster's rule to mimic clinical reasoning. In the second stage, a Detail-aware Fracture Localization Network (DFL-Net) is designed. To adapt to irregular fracture geometry, DFL-Net integrates deformable convolutions and incorporates DSSA into the feature pyramid to preserve fine spatial details during deep downsampling. Experiments on the MURA and clinical datasets demonstrate strong performance: the classification model achieved an accuracy of 92.89%, precision of 93.05%, recall of 96.96%, and F1 score of 94.97%. The localization network obtained 73.8% AP 50 and 43.8% AP 75 , with a 10.4% improvement in AP 75 over Faster R-CNN. These results indicate that the proposed framework provides an accurate and efficient tool for computer-aided fracture diagnosis, with the potential to reduce misdiagnosis and improve clinical decision-making efficiency.

Video Temporal Grounding (VTG) is a fine-grained video understanding task that aims to ground the relevant video moments corresponding to given language queries. Most existing approaches utilize powerful Vision-Language Models (VLMs), augmented with additional network architectures or specialized modules to supplement temporal reasoning capabilities. Despite achieving impressive performance, these approaches tend to overlook a critical issue that pre-trained visual and textual representations are not specifically optimized for VTG. Particularly, such representations often suffer from inter-modal semantic misalignment referring to the inconsistency between different modalities, and intra-modal semantic confusion, caused by insufficient discriminability within visual modality. To address these limitations, we propose an efficient semantics refinement framework built upon pre-trained models, featuring two core components. First, the Modal Knowledge Bidirectional Propagation (MKBP) component promotes inter-modal semantic alignment via bidirectional enrichment of textual and visual semantics, exploiting their complementary strengths without introducing additional parameters. Second, the Content Context Contrast Learning (C3L) component alleviates intra-modal semantic confusion within visual modality by bringing query-specific visual features closer while separating irrelevant ones. Comprehensive experiments on six benchmark datasets demonstrate the superior performance of our proposed methods.

Test-Time Adaptive Segmentation (TTA-Seg) aims to adapt a trained segmentation model to test data under distribution shift in an unsupervised manner. Existing approaches typically utilize class-wise prototypes to capture and transfer the source distribution, but inevitably neglecting the diversity within source samples. In this paper, we propose a new test-time adaptation paradigm based on the mixture-of-experts (MoE), where domain experts are designed to 1) better capture the source distribution, and 2) dynamically adjust their contribution in test case prediction. Specifically, during source training, prototypes are derived as the class-wise average for source pixel features. We then generate multiple experts through clustering these prototypes, providing each class with several experts with enhanced representativeness. At test time, each pixel's prediction is drawn from all experts' knowledge in an adaptive manner, \ie, a gating network assigns weight according to pixel-expert correlation. To optimize the system, we devise a min-max entropy optimization scheme for the gating network but keeping the rest frozen, minimizing the entropy of model prediction but maximizing the entropy in expert selection. Consequently, the model is urged to derived confident predictions with effective utilization of domain experts, hence promoting the adaptation. Experiments on two scenarios, Test-time Adaptation (TTA) and the more challenging continual TTA, demonstrate that our approach achieves the new state-of-the-art.

Femoral nerve block is a commonly used technique in lower limb anesthesia, relying on precise identification of nerve structures in ultrasound images. However, significant variations in the size of the femoral nerve across individuals, combined with complex backgrounds and blurred tissue boundaries in ultrasound images, pose challenges for automatic segmentation. Existing deep learning models often struggle with capturing long-range dependencies and have high computational complexity, which makes them unsuitable for real-time surgery, especially on portable devices with limited computational power. To address these challenges, we propose a lightweight deep learning network, LFNS-Net. In the encoder stage, we introduce the Axial Part Convolutional Block (APCB) to effectively capture long-range spatial information while maintaining low computational cost. In the decoder, we build the Dilated Channel Attention (DCA) module to guide the model's focus on key femoral nerve regions while suppressing background noise interference. Additionally, we incorporate a Grouped Attention Gate (GAG) in the skip connections to optimize multi-scale feature fusion. We constructed a dataset of 3684 images for femoral nerve segmentation and conducted experiments. The results show that LFNS-Net outperforms existing methods with a Dice Similarity Coefficient (DSC) of 82.61%, mean Intersection over Union (mIoU) of 84.76%, and sensitivity of 83.93%. Moreover, LFNS-Net requires only 1.51 M parameters and 3.32 G FLOPs, reducing the parameter count by 22 times compared to U-Net, while maintaining a real-time processing speed of 67 FPS. These results demonstrate LFNS-Net's potential for providing decision support to anesthesiologists, enhancing the segmentation quality and efficiency of ultrasound-guided femoral nerve block procedures.

Unsupervised domain adaptation (UDA) aims to adapt a model trained on a labeled source domain to an unlabeled target domain. Despite its promising progress, most UDA methods operate under two idealized assumptions: a well-aligned label space between source and the target, and joint access to both source and target data. These two assumptions limit the applicability of UDA methods in real-world scenarios. In this paper, we propose a unified framework that leverages inter-class relationships to distinguish known and unknown classes during test-time adaptation. Specifically, we gather class prototypes for known classes during source training, so as to derive inter-class relationships. During test, each target sample will perform matching the inter-class relationships with prior relationship knowledge, in which novel classes can be identified due to their larger divergence on the relational distribution. Extensive experiments on four benchmarks demonstrate the effectiveness of the proposed approach.

Computer-aided diagnosis holds significant promise for automatically screening thyroid nodules, thereby reducing missed and misdiagnosed cases in clinical practice. However, current deep learning approaches for thyroid nodules typically treat image segmentation and diagnostic classification as independent learning tasks, overlooking their intrinsic connection. This results in insufficient feature interaction between tasks, consequently limiting diagnostic accuracy. Furthermore, this task decoupling limits model interpretability, hindering clinical adoption. To address this, we propose SCRNet, an end-to-end interpretable multi-task collaborative learning framework that deeply integrates nodule segmentation, benign/malignant classification, and diagnostic report generation. Within the segmentation task, the prior knowledge guided module directs the model to focus on features critical to segmentation that align with clinicians' expertise. For classification, the graph reasoning module simulates the physician's diagnostic process by explicitly modelling relationships between regions. The multi-task coordination module explicitly fuses anatomical structure information derived from segmentation with raw image features.Finally, the collaborative features extracted through multi-task learning are decoded into structured diagnostic reports, with Bi-LSTM and context attention employed to enhance report accuracy and relevance. Experiments on the TR-TN3K dataset demonstrate that our method outperforms single-task models in image segmentation and diagnostic classification, achieving excellent performance (Dice: 0.87, accuracy: 0.799). For diagnostic classification and report generation, our method achieves high accuracy and clinical relevance, with strong matching of key diagnostic terms. The results indicate that this model can assist in clinical thyroid nodule diagnosis, reducing the risk of missed diagnoses and misdiagnoses, particularly for high-risk populations, thereby holding significant clinical value.

Text data are difficult for computers to process directly due to their unstructured nature. In traditional text information extraction systems, predefined structural templates are typically used to extract structured data, such as event schema or attribute-entity pairs. However, such template-based designs limit the adaptability of models when dealing with unseen domains and weaken their ability to capture the complex semantics of natural language. In this study, we propose a deep learning-driven schema-free structured information extraction paradigm to eliminate the dependence on manually designed templates. Unlike traditional slot-filling methods, our framework first extracts entities appearing in unstructured text, and then employs a deep neural text generation model to dynamically infer their semantic roles or structured attributes based on contextual semantics. This paradigm provides a unified and data-driven approach to representing unstructured text in a structured form, and can be effectively applied to complex event extraction tasks. Comprehensive experiments conducted on two benchmark datasets demonstrate that the proposed deep learning-based method not only achieves higher accuracy than the baseline models, but also significantly alleviates information omission and semantic fragmentation, exhibiting strong generalisation and robustness.

Semi-supervised learning based on consistency learning offers significant promise for enhancing medical image segmentation. Current approaches use copy-paste as an effective data perturbation technique to facilitate weak-to-strong consistency learning. However, these techniques often lead to a decrease in the accuracy of synthetic labels corresponding to the synthetic data and introduce excessive perturbations to the distribution of the training data. Such over-perturbation causes the data distribution to stray from its true distribution, thereby impairing the model's generalization capabilities as it learns the decision boundaries. We propose a weak-to-strong consistency learning framework that integrally addresses these issues with two primary designs: 1) it emphasizes the use of highly reliable data to enhance the quality of labels in synthetic datasets through cross-copy-pasting between labeled and unlabeled datasets; 2) it employs uncertainty estimation and foreground region constraints to meticulously filter the regions for copy-pasting, thus the copy-paste technique implemented introduces a beneficial perturbation to the training data distribution. Our framework expands the copy-paste method by addressing its inherent limitations, and amplifying the potential of data perturbations for consistency learning. We extensively validated our model using six publicly available medical image segmentation datasets across different diagnostic tasks, including the segmentation of cardiac structures, prostate structures, brain structures, skin lesions, and gastrointestinal polyps. The results demonstrate that our method significantly outperforms state-of-the-art models. For instance, on the PROMISE12 dataset for the prostate structure segmentation task, using only 10% labeled data, our method achieves a 15.31% higher Dice score compared to the baseline models. Our experimental code will be made publicly available at https://github.com/slhuang24/RCP4CL.

Although recent cross-domain person re-identification approaches have obtained great progress, they still suffer from two core issues. The first one is the insufficient useful knowledge transfer, which means the beneficial knowledge learned from the source domain is not utilized fully due to the fine-tuning process of the two-stage training especially. The second problem is the inappropriate transfer of the source domain knowledge. Concretely, this knowledge is not distinguished before being transferred, leading to the domain-specific knowledge is detrimental to the target domain performance. To circumvent them, we design a novel collaborative learning method named Preserving Knowledge from the Source Domain (PKSD) from both instance and pixel levels, composed of Ranking-guided Instance Selection (RIS) and Projection based Gradient Selection (PGS). Firstly, the collaborative learning manner could safeguard sufficient knowledge transfer from the source domain. Additionally, RIS tries to select reliable and informative samples from the source domain dataset for training to provide sufficient domain-shared knowledge at the instance level. Subsequently, PGS fine-tunes the feature maps of the selected samples according to the gradient modifying at the pixel level of feature maps to suppress remaining domain-specific knowledge from the source domain. Experiments show that PKSD outperforms existing state-ofthe-art methods.

Text-to-Image Person Re-identification (TIReID) aims to retrieve the queried person using a textual description, which is facing significant challenges due to the huge modality gap. Existing approaches to address this challenge have primarily focused on designing cross-modality alignment strategies based on instance-level contrastive learning, but largely ignoring the limitation of this paradigm. Specifically, instance-level learning overly focuses on salient patterns of individual instances and inevitably leads to feature bias in instance- level feature distribution. Inspired by biological research, which suggests that population-level patterns can mitigate individual-level errors, we introduce a Structure-level Distribution Guiding (SDG) strategy based on the Expectation-Maximization (EM) algorithm. SDG can efficiently compute a confident structure-level feature distribution for guidance and employs inter- and intra-class guiding losses to achieve comprehensive representation learning and instance-level feature bias mitigating respectively. Our SDG achieves competitive Rank-1 accuracy of 74.68%, 66.56%, and 63.20% on CUHK-PEDES, ICFG-PEDES, and RSTPReid respectively.

Applying deep learning to predict patient prognostic survival outcomes using histological whole-slide images (WSIs) and genomic data is challenging due to the morphological and transcriptomic heterogeneity present in the tumor microenvironment. Existing deep learning-enabled methods often exhibit learning biases, primarily because the genomic knowledge used to guide directional feature extraction from WSIs may be irrelevant or incomplete. This results in a suboptimal and sometimes myopic understanding of the overall pathological landscape, potentially overlooking crucial histological insights. To tackle these challenges, we propose the CounterFactual Bidirectional Co-Attention Transformer framework. By integrating a bidirectional co-attention layer, our framework fosters effective feature interactions between the genomic and histology modalities and ensures consistent identification of prognostic features from WSIs. Using counterfactual reasoning, our model utilizes causality to model unimodal and multimodal knowledge for cancer risk stratification. This approach directly addresses and reduces bias, enables the exploration of 'what-if' scenarios, and offers a deeper understanding of how different features influence survival outcomes. Our framework, validated across eight diverse cancer benchmark datasets from The Cancer Genome Atlas (TCGA), represents a major improvement over current histology-genomic model learning methods. It shows an average 2.5% improvement in c-index performance over 18 state-of-the-art models in predicting patient prognoses across eight cancer types.

The brachial plexus nerves at the axillary level are small in size and have a tortuous course, intertwining with blood vessels and other tissues. Ultrasound images are affected by speckle noise and structural blurring, leading to errors in manual segmentation and reducing the accuracy and success rate of axillary nerve block localization. Current lightweight networks primarily focus on semantic segmentation, making it difficult to precisely distinguish the boundaries between adjacent nerve bundles and surrounding tissues. Additionally, they lack optimisation for axillary nerve instance segmentation, making it challenging to balance accuracy and real-time performance. To address these issues, we propose the Dynamic Feature Fusion Network, which enhances segmentation accuracy while maintaining lightweight architecture. This model combines a re-parameterised vision transformer with spatial pyramid pooling to compress parameter counts while maintaining feature expression capabilities. To achieve dynamic fusion of multi-scale features, this study proposes a SimAM-enhanced Bidirectional Fusion Network based on a similarity attention module, which improves segmentation accuracy in neurovascular intersection regions. Finally, a context-based mechanism is used to improve target localisation accuracy while completing pixel-level segmentation tasks, and a hybrid loss function is designed to enhance the model's localisation accuracy and optimise the training process. The method was evaluated on the Ultrasound Axillary Brachial Plexus dataset and experimental results showed that the proposed method achieved 57FPS and 0.604 mAP@0.5:0.95 on a single GPU. This significantly improves the segmentation accuracy and efficiency of the axillary brachial plexus, providing a reliable auxiliary tool for ultrasound-guided nerve block.

Spatially resolved transcriptomics enable comprehensive measurement of gene expression at subcellular resolution while preserving the spatial context of the tissue microenvironment. While deep learning has shown promise in analyzing SCST datasets, most efforts have focused on sequence data and spatial localization, with limited emphasis on leveraging rich histopathological insights from staining images. We introduce GIST, a deep learning-enabled gene expression and histology integration for spatial cellular profiling. GIST employs histopathology foundation models pretrained on millions of histology images to enhance feature extraction and a hybrid graph transformer model to integrate them with transcriptome features. Validated with datasets from human lung, breast, and colorectal cancers, GIST effectively reveals spatial domains and substantially improves the accuracy of segmenting the microenvironment after denoising transcriptomics data. This enhancement enables more accurate gene expression analysis and aids in identifying prognostic marker genes, outperforming state-of-the-art deep learning methods with a total improvement of up to 49.72%. GIST provides a generalizable framework for integrating histology with spatial transcriptome analysis, revealing novel insights into spatial organization and functional dynamics.

Pavement crack detection is crucial for maintaining safe driving conditions; thus, the timely and accurate detection of cracks is of considerable importance. However, although deep neural networks (DNNs) have performed well in pavement crack detection, their dependence on large-scale labeled datasets, excessive model parameters, and high computational costs limit their application at the edge or on mobile devices. The conventional approaches concentrate on domain adaptation to leverage unlabeled data but overlook the domain shift issue, which can lead to performance degradation and is noticeable in lightweight models. Therefore, we propose a lightweight deep domain-adaptive crack detection network (L-DDACDN) to address these issues. Specifically, a novel distillation loss method that incorporates domain information, which facilitates the transfer of knowledge from a teacher model to a student model, is introduced. Additionally, L-DDACDN imitates the feature responses of a teacher model near the object anchor locations, ensuring that the student model effectively learns crucial features, thus addressing the domain shift issue and maintaining performance in lightweight models. Experimental results show that compared with the deep domain-adaptive crack detection network (DDACDN) trained with a large-scale pre-trained model, L-DDACDN has an average loss of only 3.5% and 3.9% in F1-scores and Accuracy, respectively. In contrast, the model parameters and FLOPs are reduced by approximately 92%. Additionally, compared to the YOLOv5, L-DDACDN demonstrates a notable improvement in the F1-scores and Accuracy on the CQU-BPDD dataset, revealing an average increase of 5% and 1.8% in F1-scores and Accuracy, respectively.

AI-based histopathological image analysis has significantly advanced the field of computer-aided diagnosis. While labeled data can enhance model performance, manual annotation by pathologists is labor-intensive and time-consuming, with variability and reliance on coarse slide-level annotations often introducing noise. To address these challenges, we propose introduces BPAL (Beta Mixture Model and Penalized Regression for Active Learning), a novel active learning framework for histopathological whole-slide image analysis. BPAL aims to reduce expert annotation costs and mitigate the impact of noisy samples during training by autonomously managing highly informative samples in each active learning iteration. Our approach integrates two noise detection modules into active learning frameworks. By incorporating Penalized Regression (PR) with parallel computation capabilities into our framework, we enhance the efficiency of noisy sample detection. Leveraging a Beta Mixture Model (BMM) with prior loss knowledge further augments this process by enabling a comprehensive analysis from various angles within the merged feature and label spaces. This approach maximizes the utilization of information extracted from pathological image samples, ensuring a robust and thorough assessment of data quality. We propose a heuristic sampling strategy based on these enhancements. High-information samples identified by the module are categorized into three types: typical samples with high confidence levels that can receive pseudo labels for training, difficult samples requiring expert re-annotation due to complex features, and mislabeled noisy samples. The iterative addition of training sets retains high-information samples while mitigating the impact of noisy samples. Comparative evaluations demonstrate the superior performance of our approach on breast cancer and prostate cancer classification tasks.

In the domain of activity-based image-to-video retrieval, dynamically consistent semantics are crucial for effective cross-modal search tasks. Existing methods face significant challenges, particularly in addressing the issue of modality asymmetry, where images and videos exhibit differing semantic representations. A key solution to this challenge lies in enhancing the learning capacity of the image encoder by leveraging knowledge from video data. To this end, we propose a Cross-Modal Knowledge Transfer (CMKT) framework that improves the behavior modeling capability of the image encoder. This enhancement is achieved through both global and local information transmission: globally, the model assimilates rich semantic information from videos across abroad temporal spectrum, while locally, it captures semantics from frames closely resembling the query image. Specifically, we design the Global Temporal Structure Transmission (GTST) Model to ensure temporal distribution consistency between query image objects and video content. Additionally, the Local Temporal Relation Enhancement (LRTE) Module is introduced to pinpoint the most relevant action information within the video. We evaluate the effectiveness of our method on two widely adopted action recognition datasets, THUMOS14 and ActivityNet, and provide comprehensive ablation studies to substantiate the efficacy of our approach.

Novel view synthesis and realistic scene reconstruction is crucial for advancing data augmentation of autonomous driving systems. However, previous 3D Gaussian Splatting methods excel in rendering with similar viewpoints, but cannot handle significant viewpoint changes, while some methods introduce additional information to improve rendering capability at the cost of increased computational overhead. To address it, we propose the Gaussian splatting that uses the Motion Trajectory priors (GMT) method, a specialized tuning of 3D Gaussian Splatting aimed at novel view synthesis of road regions under sparse-view conditions. Our method applies geometric constraints to the 3DGS that represent the road regions using only the ego-motion trajectory as a prior, without introducing additional information, which saves the computational overhead. We achieve multi-lane simulation on Waymo [15] and KITTI [5] at a horizontal displacement of 3 m, our module optimization has increased FID by 20.6 points.Our approach applies geometric constraints to the Gaussians that represent the road regions.

Single-frame temporal action localization has garnered attention in the computer vision community. Existing methods address annotation sparsity by generating dense pseudo labels within individual videos, but disregard the variable representation from intra-class action instances, resulting in inferior completeness localization. In this paper, we propose to model intra-class relationships by using Summarized Knowledge Guidance (SKG). Specifically, we initially design a learnable memory bank to summarize annotated single-frame knowledge for each class. Then, we introduce two corresponding components, i.e., the knowledge propagation module (KPM) and the knowledge refinement module (KRM), for intra-class guidance. In KPM, we propagate summarized knowledge for feature-level enhancement through bipartite matching. In KRM, summarized knowledge is presented as confident pseudo positive samples for label-level refinement in a contrastive learning manner. Extensive experiments and ablation studies on the THUMOS14, GTEA and BEOID reveal that our method significantly outperforms state-of-the-art methods.