Ziyuan Yang

Multi-modality (MM) semi-supervised learning (SSL) based medical image segmentation has recently gained increasing attention for its ability to utilize MM data and reduce reliance on labeled images. However, current methods face several challenges: (1) Multiple networks assignment hinder scalability to scenarios with more than two modalities. (2) Focusing solely on modality-invariant representation while neglecting modality-specific features, leads to incomplete MM learning. (3) Leveraging unlabeled data with generative methods can be unreliable for SSL. To address these problems, we propose a novel unified all-in-one framework for MM-SSLmedical image segmentation. To address challenge (1), we propose a modality all-in-one segmentation network based on standard U-Net architecture that accepts data from all modalities, removing the limitation on modality count. To address challenge (2), we design two learnable plug-in banks, Modality-Level Modulation bank (MLMB) and Modality-Level Prototype (MLPB) bank, to capture both modality-invariant and modality-specific knowledge. These banks are updated using our proposed Modality Prototype Contrastive Learning (MPCL). Additionally, we design Modality Adaptive Weighting (MAW) to dynamically adjust learning weights for each modality, ensuring balanced MM learning as different modalities learn at different rates. Finally, to address challenge (3), we introduce a Dual Consistency (DC) strategy that enforces consistency at both the image and feature levels without relying on generative methods. We evaluate our method on a 2-to-4 modality segmentation task using three open-source datasets, and extensive experiments show that our method outperforms state-of-the-art approaches. The source code is publicly available at https://github.com/CYYukio/Double-Knowledge-Banking.

The rise of harmful speech on social media has become a significant challenge, making hate speech detection (HSD) more crucial than ever. The HSD methods focus on enriching contextual information to enhance detection performance, but they are hindered by the inherent uncertainty of hate speech and lack the capability for reasoning from the speech. To alleviate these concerns, we propose a novel and general HSD paradigm, named trustworthy hate speech detection method through visual augmentation (TrusV-HSD). Specifically, we firstly visualize hate speech to introduce visual imaginary cues that enhance text understanding by encouraging the model to reason from the imagery. Then, to jointly model images and speech, we design a novel HSD that connects the two modalities and efficiently capture long-range interactions while maintaining low computational overhead. Considering the specificity of the task, each speech is influenced by subjective of speakers, resulting in different distributions of speeches. Therefore, an uncertainty loss is introduced to emphasize predictions with high reliability while weakening unrelated ones. In this way, the proposed TrusV-HSD learns semantic representations by effectively extracting trustworthy information through multi-modal connections without paired data. Our proposed is a novel and general HSD learning paradigm, which can be embeded with different textual or image feature extractors. Experimental results on three public HSD datasets, covering both explicit and implicit tasks, demonstrate that the TrusV-HSD significantly outperforms state-of-the-art methods. The code will be released upon acceptance of the paper.

DNA data storage is an emerging data storage method with an extremely long storage duration and high density, yet it currently lacks adequate consideration for data security. Existing encryption methods for DNA storage are based on bit-level encryption, which makes them incompatible with DNA encoding methods based on data characteristics. In this study, we propose a method for directly encrypting DNA sequences via automata cryptography for secure DNA storage (DNA-AC) to solve the above challenges. In DNA-AC, sequence-level encryption is achieved through diffusion and rotation of bases, making the base distribution more uniform and random, improving the security of DNA storage data. Security analysis demonstrates that DNA-AC exhibits a high key space, an information entropy close to 2, a Number of Bases Change Rate (NBCR) of 75% and Bases Average Changing Intensity (BACI) of 42%, showing strong resistance to malicious attacks. Compared to representative works, DNA-AC supports parallel processing to accommodate high-throughput parallel DNA synthesis and sequencing. In general, DNA-AC offers a secure, high-performance, and scalable solution for DNA storage encryption, with the potential to be an ideal data encryption approach for DNA storage devices.

Dorsal hand imagery is an emerging biometric modality offering a promising avenue for re-identification (Re-ID) in criminal investigations. Progress, however, is hampered by the scarcity of datasets. Compounding this, video-surveillance acquisition typically occurs in uncontrolled settings, adding confounders that degrade accuracy. To mitigate these gaps, we curate a corpus of dorsal hand images at a certain scale, captured from real-world surveillance, with no constraints on posture or spatial placement. Building on this resource, we present a video-based dorsal hand Re-ID pipeline comprising a frame-proposal module and an identification module. The frame-proposal module localizes dorsal hands and selects keyframes. To facilitate identification, we introduce the Channel Group-wise Exciting Network (CGENet), which models latent relations among grouped channels. Within CGENet, Channel Group-wise Attention and a Channel Group-wise Soft Mask enrich representations by amplifying intra-group cues while probing inter-group dependencies. Extensive experiments and head-to-head comparisons with state-of-the-art systems validate the effectiveness of our approach. Results consistently show improved robustness under unconstrained conditions and reliable identification across varied surveillance scenarios, underscoring the overall effectiveness of our framework.

Palmprint recognition systems experience a significant performance decline in cross-domain scenarios due to domain shift caused by non-identity factors such as capture devices and lighting conditions. To address this issue, this paper introduces a novel deep decoupling framework, the Identity and Style Feature Decoupling Network (ISFDNet), designed to improve the model’s cross-domain generalization. ISFDNet explicitly separates stable identity-related information from variable domain-related style information within palmprint features. The framework incorporates two innovative mechanisms: at the feature level, the Spatially-Aware Separation Module (SASM) adaptively produces complementary spatial attention masks to decouple mixed features into identity and style components; at the image level, the Low-Frequency Disturbance Module (LFDM) creates stylized training samples by perturbing the low-frequency parts of images, encouraging the network to learn identity representations that are insensitive to style variations. Additionally, a carefully designed collaborative supervision strategy combines multiple losses to ensure effective decoupling. Extensive experiments on four publicly available palmprint datasets demonstrate that ISFDNet achieves top performance in both cross-domain and in-domain tests, while significantly enhancing the generalization capabilities of existing networks. The code is released at https://github.com/20201422/ISFDNet.

Due to the frequent co-occurrence of multiple diseases in patients, automated multi-label diagnosis (MLD) of medical images remains a challenging yet clinically important task, particularly due to two major challenges: (1) feature confusion arises from the loss of disease localization information due to pooling operations, and (2) suboptimal probabilistic modeling capacity and feature expressiveness using vanilla MLPs (multi-layer perceptrons) as the predictors. In this work, we propose a novel MLD framework that integrates a Spatial-Disease Feature Condensation (SDFC) mechanism and Kolmogorov–Arnold Layers (KALs). The SDFC module preserves the full spatial resolution of image features by condensing channel-wise information into disease-aligned spatial representations . These spatial-disease features are subsequently processed by a KALs-based classifier, which offers enhanced expressiveness for modeling complex interdependencies among co-occurring diseases. Extensive experiments conducted on two publicly available clinical datasets, ODIR and NIH ChestX-ray, demonstrate that the proposed method consistently outperforms state-of-the-art MLD approaches in terms of diagnostic performance, generalization, and reliability. By explicitly preserving spatial disease information and leveraging the superior function modeling capability of Kolmogorov–Arnold, the proposed framework provides an effective and robust solution for multi-label medical image diagnosis, with potential value for clinical decision support systems.