Kun Wang

Large Language Models (LLMs) often suffer from factual hallucinations when generating medical summaries, such as entity substitutions and relational misalignments. Existing active knowledge-injection approaches struggle to enforce strict topological constraints during sequence generation. To address this, we propose a multi-level graph alignment framework (called MedAlign) that transforms passive knowledge fusion into an integrated four-stage closed-loop mechanism to enforce factual faithfulness in document-level medical summarization. MedAlign first constructs a noise-free, self-consistent document-level medical knowledge graph (DMKG) topology directly from the source text using self-consistency knowledge fusion (SKF) and entity normalization. It then leverages the serialized DMKG alongside the source document to convert probabilistic sequence decoding into a topology-constrained generation process. MedAlign performs multi-level graph alignment to identify residual information inconsistencies. It contrasts entity, relation, and path overlaps between the summary graph and the source DMKG, generating a graph alignment score (GAS) for fine-grained structural error diagnosis. Finally, an alignment-guided selective revision module performs targeted, localized edits on inconsistent sentences. This further eliminates hallucinations without causing global rewrite latency or introducing secondary errors. Experimental evaluations across four benchmark datasets show that MedAlign significantly reduces overall hallucination rates to 2.9%-6.5%. It also achieves state-of-the-art medical summarization performance in factual faithfulness, structural alignment, and linguistic fluency.

Accurate identification of T-cell receptor (TCR) epitope binding specificity is pivotal for cancer immunotherapy. Prevailing unified representation methods are often constrained by unidirectional interaction modeling and neglect the intrinsic 3D spatial conformation of TCR molecules, leading to a loss of biophysical semantics. Furthermore, integrating multidimensional spatial coordinates into deep learning frameworks often faces compatibility challenges due to the dynamic variance in atomic dimensions. To address these issues, we propose BASIC-TCR, an end-to-end prediction framework that deeply integrates bidirectional synergistic co-attention and spatial structure awareness. Our model breaks the limitation of traditional unidirectional attention by introducing a symmetric interaction mechanism to achieve peer-to-peer modeling between peptides and TCRs. We refactor the conventional spatial feature pipeline and design a dynamic atom padding and calibration strategy, resolving the feature misalignment caused by variable amino acid atom counts. Through a tri-modal dynamic gating fusion mechanism, the model adaptively integrates sequential semantics, deep interaction cues, and spatial geometry. Extensive experiments on standard benchmark datasets demonstrate that BASIC-TCR significantly outperforms existing state-of-the-art baselines in terms of AUC and F1-score, demonstrating improved accuracy for TCR-epitope binding prediction on benchmark datasets.

In computed tomography (CT), metal artifacts pose a persistent challenge to achieving high-quality imaging. Despite advancements in metal artifact reduction (MAR) techniques, many existing approaches have not fully leveraged the intrinsic a priori knowledge related to metal artifacts, improved model interpretability, or addressed the complex texture of CT images effectively. To address these limitations, we propose a novel and interpretable framework, the wavelet-inspired oriented adaptive dictionary network (WOADNet). WOADNet builds on sparse coding with orientational information in the wavelet domain. By exploring the discriminative features of artifacts and anatomical tissues, we adopt a high-precision filter parameterization strategy that incorporates multiangle rotations. Furthermore, we integrate a reweighted sparse constraint framework into the convolutional dictionary learning process and employ a cross-space, multiscale attention mechanism to construct an adaptive convolutional dictionary unit for the artifact feature encoder. This innovative design allows for flexible adjustment of weights and convolutional representations, resulting in significant image quality improvements. The experimental results using synthetic and clinical datasets demonstrate that WOADNet outperforms both traditional and state-of-the-art MAR methods in terms of suppressing artifacts.

In sixth-generation (6G) cognitive communications, deep learning (DL)-driven electromagnetic signal identification (ESI) emerges as a pivotal sensing paradigm to fortify physical layer security in reconfigurable intelligent surface (RIS)-aided uncrewed aerial vehicle (UAV) networks through intelligent waveform learning while preserving spectral efficiency. This paper proposes a robust RIS-aided UAV ESI framework under multiple channel impairments, addressing key challenges in practical deployment scenarios. Specifically, the methodology integrates a real-world RIS signal acquisition system comprising a Rohde & Schwarz vector signal generator (R&S-SMBV100B) emulating UAV transmissions, an FPGA-controled $16\times 16$ programmable RIS metasurface for signal reflection, and an R&S-FSW 67 GHz spectrum analyzer for receiving data capture, augmented by UAV 3D modeling and emulation to construct composite channel fading impairments. The proposed Spatio-Temporal Liquid Learning ESI Network (STL-ESINet) innovatively integrates spatial convolutional operations with temporal liquid learning mechanisms to extract discriminative high-dimensional spatio-temporal features from electromagnetic signals. A hierarchical classification module then synthesizes these spatio-temporal features for precise decision-making, while inter-class and intra-class feature statistics techniques rigorously validate the model interpretability. Experimental validation demonstrates that STL-RSINet outperforms existing DL-based approaches by achieving 97.38% and 100% recognition accuracy at SNR= 2dB and 12dB, respectively, while exhibiting robustness against composite channel impairments including path loss, multipath fading, and Doppler shifts. These results establish a new paradigm for secure RIS-UAV communications, offering both theoretical insights into waveform-learning security mechanisms and practical guidelines for hardware implementation in 6G cognitive communications.

In computed tomography (CT) imaging, highly attenuated metallic implants often result in strong metal artifacts, which significantly compromise image quality and pose considerable challenges for precise clinical diagnosis and effective treatment. Recently, deep learning-based approaches have shown promising results in metal artifact reduction (MAR) for CT images. Nonetheless, these approaches often address the problem as a broad image restoration challenge. The networks employed are generally constructed intuitively, lack interpretability, and struggle with metal artifacts of varying sizes. To alleviate this problem, we propose an innovative MAR model called the learning weighted convolutional dictionary network (LWCDNet). The LWCDNet architecture comprises multiple iterative blocks, with each iteration featuring two distinct modules dedicated to optimizing the tissue attenuation component and the metal artifact component. First, we advance traditional convolutional learning by introducing a reweighted variant and develop a learning weighted convolutional dictionary unit on the basis of the iterative soft thresholding algorithm. This approach is designed specifically for the representation of artifact features. Second, a lightweight residual network (ResNet) model attempts to improve the restoration quality of the tissue attenuation component. Finally, via a metal mask, we realize discriminative representation and constraints on different components in metal artifact CT images. Comprehensive experimental investigations are carried out to validate the efficacy of the proposed method. The results on both synthetic and clinical datasets indicate that LWCDNet significantly outperforms traditional and state-of-the-art MAR methods, effectively enhancing image quality.