Selected work

Representative Papers

Dual-stream Spatio-Temporal GCN-Transformer Network for 3D Human Pose Estimation

Apr 19, 2026

This work addresses the limitations of existing Transformer-based methods for 3D human pose estimation, which struggle to effectively model local skeletal structures and inter-channel dependencies, leading to insufficient fusion of global and local features. To overcome this, the authors propose MixTGFormer, a novel dual-stream network that innovatively integrates Graph Convolutional Networks (GCNs) into the Transformer architecture. The core component is a spatio-temporal Mixformer module augmented with Squeeze-and-Excitation channel attention, enabling synergistic modeling of local-global relationships and efficient feature fusion. Evaluated on standard benchmarks, the method achieves state-of-the-art performance with P1 errors of 37.6 mm on Human3.6M and 15.7 mm on MPI-INF-3DHP.

0 citationsRead paper

Building evidence-based knowledge graphs from full-text literature for disease-specific biomedical reasoning

Mar 30, 2026

Existing biomedical knowledge resources are often limited to unstructured text or flat triples that omit critical contextual details such as study design, source provenance, and quantitative support, thereby hindering fine-grained reasoning. This work proposes a large language model–assisted end-to-end framework that extracts experimental findings from full-text literature as structured evidence nodes, standardizes biomedical entities, evaluates evidence quality, and constructs disease-specific knowledge graphs via typed semantic relations. To our knowledge, this is the first approach to systematically build structured evidence graphs from full-text articles while preserving study design, source attribution, and quantitative backing. The authors release two high-quality datasets, EvidenceNet-HCC and EvidenceNet-CRC, with component-level validation accuracy of 98.3%, demonstrating significant performance gains in retrieval-augmented question answering, link prediction, and target prioritization tasks.

0 citationsRead paper

A two-stream network with global-local feature fusion for bone age assessment

Dec 20, 2025

Existing bone age assessment methods struggle to simultaneously capture global skeletal structure and fine-grained epiphyseal details, limiting accuracy. To address this, we propose BoNet+, a novel dual-stream deep network that integrates Transformer-based global modeling with RFAConv-generated multi-scale adaptive local attention maps; feature fusion is further optimized via fine-tuned Inception-V3. This architecture overcomes the limitations of conventional single-stream CNNs by jointly modeling hierarchical anatomical cues. Evaluated on the RSNA and RHPE benchmarks, BoNet+ achieves state-of-the-art mean absolute errors (MAE) of 3.81 months and 5.65 months, respectively. The framework enables fully automated, high-accuracy, and clinically deployable bone age estimation—significantly reducing reliance on expert interpretation and advancing the practical deployment of intelligent bone age analysis in real-world clinical settings.

0 citationsRead paper

FMD-TransUNet: Abdominal Multi-Organ Segmentation Based on Frequency Domain Multi-Axis Representation Learning and Dual Attention Mechanisms

Sep 19, 2025

To address the low segmentation accuracy of small-sized, irregular, and anatomically complex abdominal organs, this paper proposes a novel deep learning framework integrating multi-axis frequency-domain representation with a dual-attention mechanism. Methodologically, we design a Multi-Axis External Weighting Block (MEWB) to jointly model global structural and local boundary information in the frequency domain, and enhance the Dual-Attention module (DA+) to strengthen cross-domain feature fusion between spatial and frequency domains, thereby mitigating the encoder-decoder semantic gap. Built upon the TransUNet architecture, our method incorporates frequency-domain multi-axis representation learning, depthwise separable convolutions, and joint spatial-channel attention. Evaluated on the Synapse dataset, it achieves a mean Dice Similarity Coefficient (DSC) of 81.32% and Hausdorff Distance (HD) of 16.35 mm—improving DSC by 3.84% and reducing HD by 15.34 mm over the baseline—and outperforms state-of-the-art methods.

0 citationsRead paper

EvidenceMap: Unleashing the Power of Small Language Models with Evidence Analysis for Biomedical Question Answering

Jan 22, 2025

Large language models (LLMs) exhibit excessive dependency and insufficient utilization of multi-source evidence in biomedical question answering. Method: We propose EvidenceMap, a framework based on small language models (SLMs) that introduces the first evidence graph modeling mechanism—decoupling and jointly learning supportive evidence assessment, logical relational reasoning, and summary generation across heterogeneous sources. It employs a dual-SLM collaborative architecture (Analyzer + Generator) to enable analysis-augmented autoregressive QA. Contribution/Results: EvidenceMap is the first to demonstrate that evidence-analyzed SLMs outperform GPT-4, Claude-3, RAG, and Chain-of-Thought on multiple biomedical QA benchmarks, achieving significant accuracy gains, 3.2× higher inference efficiency, and >90% parameter reduction. This establishes a new paradigm for lightweight, interpretable, and high-accuracy domain-specific QA.

0 citationsRead paper

Recent publications

Latest Papers

Dual-stream Spatio-Temporal GCN-Transformer Network for 3D Human Pose Estimation

Apr 19, 2026

This work addresses the limitations of existing Transformer-based methods for 3D human pose estimation, which struggle to effectively model local skeletal structures and inter-channel dependencies, leading to insufficient fusion of global and local features. To overcome this, the authors propose MixTGFormer, a novel dual-stream network that innovatively integrates Graph Convolutional Networks (GCNs) into the Transformer architecture. The core component is a spatio-temporal Mixformer module augmented with Squeeze-and-Excitation channel attention, enabling synergistic modeling of local-global relationships and efficient feature fusion. Evaluated on standard benchmarks, the method achieves state-of-the-art performance with P1 errors of 37.6 mm on Human3.6M and 15.7 mm on MPI-INF-3DHP.

0 citationsRead paper

Building evidence-based knowledge graphs from full-text literature for disease-specific biomedical reasoning

Mar 30, 2026

Existing biomedical knowledge resources are often limited to unstructured text or flat triples that omit critical contextual details such as study design, source provenance, and quantitative support, thereby hindering fine-grained reasoning. This work proposes a large language model–assisted end-to-end framework that extracts experimental findings from full-text literature as structured evidence nodes, standardizes biomedical entities, evaluates evidence quality, and constructs disease-specific knowledge graphs via typed semantic relations. To our knowledge, this is the first approach to systematically build structured evidence graphs from full-text articles while preserving study design, source attribution, and quantitative backing. The authors release two high-quality datasets, EvidenceNet-HCC and EvidenceNet-CRC, with component-level validation accuracy of 98.3%, demonstrating significant performance gains in retrieval-augmented question answering, link prediction, and target prioritization tasks.

0 citationsRead paper

A two-stream network with global-local feature fusion for bone age assessment

Dec 20, 2025

Existing bone age assessment methods struggle to simultaneously capture global skeletal structure and fine-grained epiphyseal details, limiting accuracy. To address this, we propose BoNet+, a novel dual-stream deep network that integrates Transformer-based global modeling with RFAConv-generated multi-scale adaptive local attention maps; feature fusion is further optimized via fine-tuned Inception-V3. This architecture overcomes the limitations of conventional single-stream CNNs by jointly modeling hierarchical anatomical cues. Evaluated on the RSNA and RHPE benchmarks, BoNet+ achieves state-of-the-art mean absolute errors (MAE) of 3.81 months and 5.65 months, respectively. The framework enables fully automated, high-accuracy, and clinically deployable bone age estimation—significantly reducing reliance on expert interpretation and advancing the practical deployment of intelligent bone age analysis in real-world clinical settings.

0 citationsRead paper

FMD-TransUNet: Abdominal Multi-Organ Segmentation Based on Frequency Domain Multi-Axis Representation Learning and Dual Attention Mechanisms

Sep 19, 2025

To address the low segmentation accuracy of small-sized, irregular, and anatomically complex abdominal organs, this paper proposes a novel deep learning framework integrating multi-axis frequency-domain representation with a dual-attention mechanism. Methodologically, we design a Multi-Axis External Weighting Block (MEWB) to jointly model global structural and local boundary information in the frequency domain, and enhance the Dual-Attention module (DA+) to strengthen cross-domain feature fusion between spatial and frequency domains, thereby mitigating the encoder-decoder semantic gap. Built upon the TransUNet architecture, our method incorporates frequency-domain multi-axis representation learning, depthwise separable convolutions, and joint spatial-channel attention. Evaluated on the Synapse dataset, it achieves a mean Dice Similarity Coefficient (DSC) of 81.32% and Hausdorff Distance (HD) of 16.35 mm—improving DSC by 3.84% and reducing HD by 15.34 mm over the baseline—and outperforms state-of-the-art methods.

0 citationsRead paper

EvidenceMap: Unleashing the Power of Small Language Models with Evidence Analysis for Biomedical Question Answering

Jan 22, 2025

Large language models (LLMs) exhibit excessive dependency and insufficient utilization of multi-source evidence in biomedical question answering. Method: We propose EvidenceMap, a framework based on small language models (SLMs) that introduces the first evidence graph modeling mechanism—decoupling and jointly learning supportive evidence assessment, logical relational reasoning, and summary generation across heterogeneous sources. It employs a dual-SLM collaborative architecture (Analyzer + Generator) to enable analysis-augmented autoregressive QA. Contribution/Results: EvidenceMap is the first to demonstrate that evidence-analyzed SLMs outperform GPT-4, Claude-3, RAG, and Chain-of-Thought on multiple biomedical QA benchmarks, achieving significant accuracy gains, 3.2× higher inference efficiency, and >90% parameter reduction. This establishes a new paradigm for lightweight, interpretable, and high-accuracy domain-specific QA.

0 citationsRead paper