Medical image segmentation plays a crucial role in clinical diagnosis, lesion quantification, and preoperative planning. However, existing Mamba-based architectures, which rely on fixed-direction sequence modeling and flatten images into one-dimensional (1D) sequences, struggle to capture hierarchical anatomical features and spatial dependencies, thereby limiting their representational capacity for complex medical structures. To address these limitations, we propose EccoMamba (Enhanced Cross-hierarchical Continuity Orthogonal Mamba), a U-shaped encoder--decoder framework designed for medical image segmentation. In the encoder's downsampling path, we introduce a Hierarchical Aggregation Enhancement (HAE) module that integrates multi-scale convolutions with hierarchical attention mechanisms. The attention branch further incorporates cross-channel interactions, allowing the model to selectively enhance semantically relevant features while suppressing irrelevant background responses. For skip connections, we design a Structural Continuity Orthogonal (SCO) module to preserve spatial continuity by modeling cross-dimensional dependencies via orthogonal Axial Shifts (AS), thereby mitigating directional bias and improving anatomical consistency. Extensive experiments on four benchmark datasets---ISIC 2018, ISIC 2017, Synapse, and ACDC---show that EccoMamba consistently outperforms state-of-the-art methods in both segmentation accuracy and structural fidelity.

The advent of Whole Slide Imaging (WSI) has revolutionised digital pathology by enabling computational analysis of gigapixel-scale images. To handle their large size, most deep learning models divide WSIs into patches and apply Multiple Instance Learning (MIL) for slide-level classification. However, MIL models often depend on pre-trained feature extractors, resulting in domain gaps between natural and pathological images. Parameter-Efficient Fine-Tuning (PEFT) via visual prompting has emerged to bridge this gap with minimal overhead. Nevertheless, existing visual prompts are typically attached at the image level and tightly coupled with specific architectures such as CNNs or ViTs, limiting generalisability and scalability in WSI tasks. To overcome these limitations, we propose Slide-aware Deep Feature Prompt (S-DFP), a novel visual prompting method which derives task-specific information directly from feature embeddings and is initialised with slide-specific cues, thereby enhancing compatibility with diverse feature extractors and MIL frameworks. Experiments on four benchmark datasets, CAMELYON16, BRIGHT, TCGA-IDH, and UniToPath, demonstrate that S-DFP consistently boosts MIL model performance by 2–5% in AUC while introducing less than 0.02% additional parameters. Furthermore, when integrated with recent pathology foundation models, S-DFP yields additional performance gains. The code is publicly available at S-DFP .

Designing new molecules with ideal properties is a critical task in drug discovery. In recent years, the accumulation of available molecular datasets has facilitated the widespread application of deep generative models in drug design. Nonetheless, a significant challenge remains in developing highperformance generative models that not only need to produce chemically valid molecular structures but also optimize the chemical properties of the generated molecules. In this study, we introduce TransFVAE, a graph Transformer-based flow Variational AutoEncoder tailored for molecular graph generation. Our approach employs VAE as the encoder and integrates a lightweight flow model as the decoder. The encoder is strategically designed to expedite the training process of the decoder, while the decoder reciprocally enhances the performance of the encoder. Unlike some existing models that only account for local node connections, our model leverages Transformer architecture in the encoder, ensuring comprehensive consideration of global information. This enables each atom to holistically interact with all other atoms, thereby enhancing molecular attribute constraint optimization in molecular optimization tasks. Validation of our model is conducted through three core tasks: molecule generation and reconstruction, latent space visualization, and molecular optimization. The results affirm the state-of-the-art performance of our model, underscoring its substantial potential in facilitating the generation of drug molecules endowed with desired properties. All source datasets and codes can be downloaded from: https://github.com/Biowust/TransFVAE.

Atopic dermatitis (AD) is a chronic inflammatory skin disorder characterized by recurrent itching, erythema, dryness, and eczematous lesions. Automated AD severity assessment is crucial for cost-effective and precision clinical decision-making but remains challenging. This is due to the subtle contrast variations between key dermatological signs and significant variations in lesion sizes across patients and disease stages. To address these issues, we propose ADSA-Net, which is designed to handle both intra- and inter-class variabilities. ADSA-Net first extracts multi-scale texture-aware features to effectively model variations in lesion size and texture. It then leverages contrastive learning to enhance intra- and inter-class differentiation, strengthening model's discriminatory ability for samples that are difficult to distinguish. Finally, ADSA-Net refines the learning process by leveraging a dynamic feature pool of correctly classified samples to guide the calibration of misclassified instances, enhancing overall accuracy. We further establish a dataset for AD severity assessment. Comprehensive experiments on this dataset show that ADSA-Net significantly outperforms existing state-of-the-art methods.

In real-world applications where knowledge systems continuously evolve with new relationships, current knowledge graph completion (KGC) methods face a fundamental limitation: their over-reliance on localized patterns and inability to capture structural semantic correlations lead to significant performance degradation when encountering unseen relations. This manifests particularly in their failure to effectively predict potential entity interactions for these novel relations. To bridge this gap, we propose BiRMet, a novel framework integrating bidirectional relational fusion with meta-learning. The framework innovates through relational graph structures that preserve semantic relationships, multi-head attention mechanisms modeling complex relational interactions, and bidirectional feature fusion enabling dynamic co-adaptation of entities and relations. When combined with meta-learning, these components collectively address the core challenge of generalizing to unseen relations. Experimental results across multiple benchmarks demonstrate consistent improvements over existing approaches in handling unseen relations. Furthermore, case studies on drug repurposing based on a biomedical knowledge graph further highlight the potential of our framework in accelerating real-world drug discovery. Codes are available at https://anonymous.4open.science/r/BiRMet-DE01

Micro-expression recognition is a challenging task of identifying hidden emotion, as micro-expressions have brief durations and involve small-scale facial muscle movements. Although deep learning-based methods, especially transformer-based methods, have achieved impressive performance in this task, these methods exhibit high computational complexity and struggle to learn effective representations in the context of typically small-scale micro-expression datasets, due to the excess of tokens in the multi-head self-attention. Moreover, most existing methods do not differentiate the importance of local features, especially in micro-expression recognition with subtle changes. Therefore, we propose a novel unified Patch Tokenization, Selection and Representation framework (PTSR) with vision Transformer for micro-expression recognition. Specifically, PTSR first presents a dual norm shifted patch tokenization (DNSPT) module to learn spatial relations between neighboring pixels of the face region, which is implemented by elaborating spatial transformation and dual norm projection. Then, we employ a local-global attention module (LAM) to extract the local-global image feature, incorporating a dynamic token selection module (DTSM) to select important patches/tokens, thereby capturing more discriminative representations for the input clip. Extensive experiments are conducted on 4 widely used public datasets, i.e., CASME II, SAMM, SMIC, CAS(ME)3, and the experimental results indicate that our method can achieve clear performance improvements over the state-of-the-art methods, such as 8.37% improvement on the CAS(ME)3 dataset in terms of UF1 and 3.1% improvement on the SMIC dataset in terms of UAR metric.

Recently, diffusion models which have achieved promising progress in text-to-image generation generally have also been generally explored for image captioning. However, these diffusion-based image captioning methods usually suffer from semantic inconsistency between image content and textual description, thus producing lagging results compared with Auto-Regressive (AR) ones. To this end, in this paper, we propose a novel dual diffusion-based framework namely MirrorDiff, to achieve semantic consistency with a symmetric image-to-text-to-image generation model, which acts like a mirror that maps the original input image into a regenerated image via the generated caption. Specifically, it first utilizes both pre-trained image encoder and text encoder to obtain image representation and textual representation respectively, then forwards the image representation and the noisy textual representation into a continuous diffusion model to output an intermediate sentence. To semantically align the intermediate sentence with the input image, a diffusion-based visual regenerator is employed to regenerate the input image conditioned on the intermediate sentence, resulting in a proposed visual regeneration loss. Different from most existing image captioning methods, MirrorDiff is a plug-and-play framework which can be plugged into many previous image captioning methods, and further evaluate the generated sentence via the visual similarity between the input image and the regenerated image. Extensive experiments on the MS COCO dataset show that our method achieves obvious improvements over state-of-the-art diffusion-based methods, up to 127.9 on CIDEr, and achieves competitive performance on multiple evaluation metrics over the auto-regressive methods trained on larger-scale datasets.

Radar imaging provides non-contact, privacy-preserving, and environmentally robust monitoring for continuous human motion recognition (HMR) by leveraging diverse information embedded in various radar signal domains. However, current research has not effectively integrated multi-radar and multi-domain imaging to fully exploit the benefits of distributed radar systems. To bridge this gap, we propose a multi-radar, multi-domain parallel cross-attention model with four key components: intra-domain cross-radar weight sharing encoders specific to each domain for consistent feature extraction and parameter reduction, domain-level parallel cross-attention (DLPCAN) modules to fuse domain-specific features and enhance feature representation robustness in each radar, a source-level attention fusion (SLAF) module to highlight significant features from multiple radar inputs, and two bi-directional gated recurrent unit (BiGRU) modules to capture temporal information. The model is trained using connectionist temporal classification (CTC) loss for effective sequence prediction. By integrating data from multiple radar nodes and domains, our approach significantly improves continuous HMR performance compared to single radar systems and single domain data. Comparative evaluations demonstrate that our model outperforms state-of-the-art radar imaging-based HMR solutions.

Molecular subtyping of cancer is recognized as a critical and challenging upstream task for personalized therapy. Existing deep learning methods have achieved significant performance in this domain when abundant data samples are available. However, the acquisition of densely labeled samples for cancer molecular subtypes remains a significant challenge for conventional data-intensive deep learning approaches. In this work, we focus on the few-shot molecular subtype prediction problem in heterogeneous and small cancer datasets, aiming to enhance precise diagnosis and personalized treatment. We first construct a new few-shot dataset for cancer molecular subtype classification and auxiliary cancer classification, named TCGA Few-Shot, from existing publicly available datasets. To effectively leverage the relevant knowledge from both tasks, we introduce a task-specific embedding-based meta-learning framework (TSEML). TSEML leverages the synergistic strengths of a model-agnostic meta-learning (MAML) approach and a prototypical network (ProtoNet) to capture diverse and fine-grained features. Comparative experiments conducted on the TCGA FewShot dataset demonstrate that our TSEML framework achieves superior performance in addressing the problem of few-shot molecular subtype classification.

Gliomas hold a prominent position in neurooncology due to their high malignancy and poor survival rates. Accurately predicting the prognosis and survival risk of glioma patients is crucial for clinical treatment. Recent advances in survival prediction methods emphasize the importance of integrating complementary information from diverse modalities while neglecting the significant modality gap between pathological images and genomic data. To address this issue, we propose a modality-specific knowledge interaction network (MSKI-Net), which integrates whole slide images (WSI), RNA-Seq gene expression data, and copy number variation (CNV) data for glioma survival analysis. The MSKI-Net consists of a modality-specific feature enhancement (MSFE) module, a modality-interactive cross-attention (MICA) module, and a modality-specific knowledge-guided representation learning (MSKR) module. The three modules collaborate by complementing modality-specific features with modality-agnostic knowledge to improve the learning capability of MSKI-Net. Furthermore, we construct a dataset named TCGAmm, which combines WSI, RNA-Seq, and CNV data from The Cancer Genome Atlas (TCGA) to address the issue of data scarcity. Extensive experiments demonstrate that MSKI-Net achieves superior performance in predicting the survival risk of glioma cancer.

Retinal vessel segmentation from fundus images is of significant importance for detecting and diagnosing common ocular diseases. Conventional deep learning-based methods for retinal vessel segmentation follow the U-Net framework with an encoder-decoder architecture and employ skip connections for the recovery of spatial information lost during downsampling. However, skip connections cannot consistently have positive contributions to segmentation performance, which is caused by the semantic incompatibility between encoder features and decoder features. Based on this observation, we propose CaFFNet, a Cross-attention Feature Fusion Network designed specifically for retinal vessel segmentation. Specifically, we improve skip connections by introducing a Cross-attention Feature Fusion (CaFF) module, which effectively mitigates the semantic gap between encoder and decoder feature maps by leveraging the cross-attention mechanism for feature fusion. Besides, we introduce a Dual-Branch Pooling Fusion (DBPF) module to address the loss of vessel spatial information during pooling and capture contextual details more effectively, so as to improve segmentation performance. Experimental results on three fundus image datasets demonstrate that our CaFFNet outperforms current representative methods for retinal vessel segmentation.

Accurate tumor segmentation is crucial for esophageal cancer radiotherapy treatment planning. The low contrast among the esophagus, tumors, and surrounding tissues, and irregular tumor shapes limit the performance of automatic segmentation methods. In this paper, we aim to exploit the irregular shapes of tumors to facilitate accurate segmentation. We propose a simple and pluggable shape-aware contrastive deep supervision network (SCDSNet) with shape-aware regularization and voxel-to-voxel contrastive deep supervision. Specifically, the shape-aware regularization with an uncertainty minimization strategy encourages the precise predictions of an additional shape-aware head. The voxel-to-voxel contrastive deep supervision enhances the multi-scale shape-tumor contrast for better voxel-to-voxel prediction of shapes. The proposed method is simple and highly pluggable, which can easily be extended to other frameworks. Further, we establish a large in-house dataset on esophageal cancer to validate the effectiveness of our proposed method. The quantitative and qualitative experimental results demonstrate the effectiveness of SCDSNet on the esophageal cancer dataset.

Since abnormal expression of long non-coding RNAs (lncRNAs) is associated with various human diseases, identifying disease-related lncRNAs helps reveal the pathogenesis of diseases. Existing methods for lncRNA-disease association prediction mainly focus on multi-sourced data related to lncRNAs and diseases. The rich semantic information of meta-paths, composed of multiple kinds of connections between lncRNA and disease nodes, is neglected. We propose a new prediction method, MGLDA, to encode and integrate the semantics of multiple meta-paths, the global topology of heterogeneous graph, and pairwise attributes of lncRNA and disease nodes. First, a tri-layer heterogeneous graph is constructed to associate multi-sourced data across the lncRNA, disease, and miRNA nodes. Afterwards, we establish multiple meta-paths connecting the lncRNA and disease nodes to derive and denote various semantics. Each meta-path contains its specific semantics formulated by an embedding strategy, and each embedding covers local topology formed by the diverse semantic connections among the lncRNA, disease, and miRNA nodes. We construct multiple graph convolutional autoencoders (GCA) with topology-level attention to learn global and multiple local topologies from the tri-layer graph and each meta-path, respectively. The topology-level attention mechanism can learn the importance of various global and local topologies for adaptive pairwise topology fusion. Finally, a convolutional autoencoder learns the attribute representations of lncRNA-disease pairs, which integrates the learnt detailed and representative pairwise features. Experimental results show that MGLDA outperforms other state-of-the-art prediction methods in comparison and retrieves more real lncRNA-disease associations in the top-ranked candidates. The ablation study also demonstrates the important contributions of the local and global topology learning, and pairwise attribute learning. Case studies on three diseases further demonstrate MGLDA's ability to identify potential disease-related lncRNAs.

Osteoporosis is the most common reason that causes the fracture among the elderly. For the purpose of convenience and safety, 2D texture analysis has been used to diagnose osteoporosis. In this study, a supervised method using proposed texture features to identify osteoporotic cases from healthy was proposed. We designed two groups of new features, Encoded GLCM and Encoded LBP, each of which contains two subgroups through encoding the Gabor and Hessian information into the Gray Level Co-Occurrence Matrix (GLCM) features and Local Binary Patterns (LBP) features respectively. These two groups of features, together with the raw feature group containing the GLCM and LBP features, totally 560 features, were categorized into various groups and used to train the Random Forest classifier. Classification performances using these features were compared inter-and intra-groups/subgroups. And the performance using each individual feature was also provided. We conducted feature selection based on Recursive Feature Elimination (RFE) inside a voting scheme to further increase the efficiency. The inter-and intra-groups/subgroups results indicate that the Encoded GLCM and Encoded LBP, are more discriminative than the raw GLCM and LBP features for the identification of the osteoporosis; The best individual feature is from the Encoded LBP group and can achieve 70% of balanced accuracy; Furthermore, using only ten of the proposed features through feature selection, the balanced accuracy can even be improved from 60% to 71%. This shows that the proposed method is promising to assist the early diagnosis of osteoporosis.