Selected work

Representative Papers

Foundation Model-driven Key Anatomy Frame Selection for Blind-sweep Ultrasound Fetal Birth Weight Estimation

Jul 01, 2026

This study addresses the challenge of operator-independent fetal birth weight estimation from unguided (“blind-scan”) ultrasound videos in resource-limited settings. The authors propose a vision–language foundation model-based framework that automatically selects key anatomical frames from videos acquired within 48 hours before delivery and incorporates a redundancy-aware feature compression module to preserve task-relevant anatomical information for end-to-end weight regression. As the first method to estimate fetal weight directly from blind-scan videos, this approach achieves a mean absolute error of 161.3 grams on a prospective cohort of 839 cases, with 90.23% and 100% of estimates falling within 10% and 15% absolute percentage error, respectively—significantly outperforming the conventional Hadlock method and current strong baselines.

0 citationsRead paper

UniSurg: A Video-Native Foundation Model for Universal Understanding of Surgical Videos

Feb 05, 2026

This work addresses the limitation of existing foundation models for surgical videos, which overly focus on low-level visual artifacts such as smoke and specular reflections while failing to capture high-level semantic structures. To overcome this, the authors propose a native video foundation model tailored for surgical videos, shifting the learning objective from pixel-level reconstruction to latent motion prediction. Built upon the V-JEPA architecture, the method introduces three key innovations: motion-guided latent prediction, spatiotemporal affinity-based self-distillation, and feature diversity regularization. The model is pretrained on UniSurg-15M, a large-scale surgical video dataset, and demonstrates state-of-the-art performance across 17 benchmark tasks, including surgical phase recognition, action triplet understanding, skill assessment, polyp segmentation, and depth estimation.

0 citationsRead paper

BREATH-VL: Vision-Language-Guided 6-DoF Bronchoscopy Localization via Semantic-Geometric Fusion

Jan 07, 2026arXiv.org

This work addresses the challenges of six-degree-of-freedom endoscopic localization in real-world clinical settings—namely, data scarcity, the difficulty of fine-grained pose regression, and high computational latency in temporal modeling—by introducing BREATH, the largest in vivo bronchoscopic localization dataset to date, and the BREATH-VL framework. BREATH-VL uniquely integrates the semantic understanding of vision-language models with geometric constraints from visual registration, and incorporates a lightweight temporal context learning mechanism that encodes motion history into language prompts for efficient temporal inference. Experimental results demonstrate that the proposed method significantly outperforms existing purely visual approaches while maintaining low computational latency, achieving a 25.5% reduction in translation error and exhibiting superior accuracy and generalization capability.

0 citationsRead paper

VTCBench: Can Vision-Language Models Understand Long Context with Vision-Text Compression?

Dec 17, 2025

This work systematically investigates, for the first time, how visual-textual compression (VTC) impacts the long-context understanding capabilities of vision-language models (VLMs). To this end, we introduce VTCBench—the first long-context benchmark specifically designed for VTC evaluation—covering retrieval, reasoning, and memory tasks, along with its real-world extension, VTCBench-Wild. We propose a multi-dimensional evaluation framework integrating OCR encoding, 2D dense representation compression, cross-modal attention analysis, and dialogue memory tracking, enabling fine-grained, unified assessment of both open- and closed-source VLMs. Experimental results reveal that while state-of-the-art VLMs accurately decode OCR-extracted text, their long-range factual association and implicit reasoning abilities degrade substantially under VTC compression—causing an average performance drop of 42.7% across all three task categories. This exposes a critical semantic fidelity gap in current VTC methods, directly undermining VLMs’ contextual comprehension.

0 citationsRead paper

CellScout: Visual Analytics for Mining Biomarkers in Cell State Discovery.

Nov 24, 2025IEEE Transactions on Visualization and Computer Graphics

Current cell state discovery relies on dimensionality reduction, visualization, and manual clustering interpretation; however, intra-cluster heterogeneity frequently compromises biomarker identification accuracy, resulting in high trial-and-error costs and poor interpretability. To address this, we propose a novel framework integrating Mixture-of-Experts (MoE) modeling with interactive visual analytics: the MoE model automatically learns nonlinear associations between cell subpopulations and gene biomarkers without imposing rigid clustering assumptions; concurrently, the visual interface enables biologists to iteratively formulate, test, and refine state hypotheses while incorporating domain knowledge to guide model optimization. Case studies on real single-cell datasets demonstrate that our approach significantly improves biomarker detection accuracy and biological interpretability, successfully aiding the discovery of novel cell states and reducing analytical uncertainty by 42% (per expert assessment) compared to conventional methods.

0 citationsRead paper

Recent publications

Latest Papers

Foundation Model-driven Key Anatomy Frame Selection for Blind-sweep Ultrasound Fetal Birth Weight Estimation

Jul 01, 2026

This study addresses the challenge of operator-independent fetal birth weight estimation from unguided (“blind-scan”) ultrasound videos in resource-limited settings. The authors propose a vision–language foundation model-based framework that automatically selects key anatomical frames from videos acquired within 48 hours before delivery and incorporates a redundancy-aware feature compression module to preserve task-relevant anatomical information for end-to-end weight regression. As the first method to estimate fetal weight directly from blind-scan videos, this approach achieves a mean absolute error of 161.3 grams on a prospective cohort of 839 cases, with 90.23% and 100% of estimates falling within 10% and 15% absolute percentage error, respectively—significantly outperforming the conventional Hadlock method and current strong baselines.

0 citationsRead paper

UniSurg: A Video-Native Foundation Model for Universal Understanding of Surgical Videos

Feb 05, 2026

This work addresses the limitation of existing foundation models for surgical videos, which overly focus on low-level visual artifacts such as smoke and specular reflections while failing to capture high-level semantic structures. To overcome this, the authors propose a native video foundation model tailored for surgical videos, shifting the learning objective from pixel-level reconstruction to latent motion prediction. Built upon the V-JEPA architecture, the method introduces three key innovations: motion-guided latent prediction, spatiotemporal affinity-based self-distillation, and feature diversity regularization. The model is pretrained on UniSurg-15M, a large-scale surgical video dataset, and demonstrates state-of-the-art performance across 17 benchmark tasks, including surgical phase recognition, action triplet understanding, skill assessment, polyp segmentation, and depth estimation.

0 citationsRead paper

BREATH-VL: Vision-Language-Guided 6-DoF Bronchoscopy Localization via Semantic-Geometric Fusion

Jan 07, 2026arXiv.org

This work addresses the challenges of six-degree-of-freedom endoscopic localization in real-world clinical settings—namely, data scarcity, the difficulty of fine-grained pose regression, and high computational latency in temporal modeling—by introducing BREATH, the largest in vivo bronchoscopic localization dataset to date, and the BREATH-VL framework. BREATH-VL uniquely integrates the semantic understanding of vision-language models with geometric constraints from visual registration, and incorporates a lightweight temporal context learning mechanism that encodes motion history into language prompts for efficient temporal inference. Experimental results demonstrate that the proposed method significantly outperforms existing purely visual approaches while maintaining low computational latency, achieving a 25.5% reduction in translation error and exhibiting superior accuracy and generalization capability.

0 citationsRead paper

VTCBench: Can Vision-Language Models Understand Long Context with Vision-Text Compression?

Dec 17, 2025

This work systematically investigates, for the first time, how visual-textual compression (VTC) impacts the long-context understanding capabilities of vision-language models (VLMs). To this end, we introduce VTCBench—the first long-context benchmark specifically designed for VTC evaluation—covering retrieval, reasoning, and memory tasks, along with its real-world extension, VTCBench-Wild. We propose a multi-dimensional evaluation framework integrating OCR encoding, 2D dense representation compression, cross-modal attention analysis, and dialogue memory tracking, enabling fine-grained, unified assessment of both open- and closed-source VLMs. Experimental results reveal that while state-of-the-art VLMs accurately decode OCR-extracted text, their long-range factual association and implicit reasoning abilities degrade substantially under VTC compression—causing an average performance drop of 42.7% across all three task categories. This exposes a critical semantic fidelity gap in current VTC methods, directly undermining VLMs’ contextual comprehension.

0 citationsRead paper

CellScout: Visual Analytics for Mining Biomarkers in Cell State Discovery.

Nov 24, 2025IEEE Transactions on Visualization and Computer Graphics

Current cell state discovery relies on dimensionality reduction, visualization, and manual clustering interpretation; however, intra-cluster heterogeneity frequently compromises biomarker identification accuracy, resulting in high trial-and-error costs and poor interpretability. To address this, we propose a novel framework integrating Mixture-of-Experts (MoE) modeling with interactive visual analytics: the MoE model automatically learns nonlinear associations between cell subpopulations and gene biomarkers without imposing rigid clustering assumptions; concurrently, the visual interface enables biologists to iteratively formulate, test, and refine state hypotheses while incorporating domain knowledge to guide model optimization. Case studies on real single-cell datasets demonstrate that our approach significantly improves biomarker detection accuracy and biological interpretability, successfully aiding the discovery of novel cell states and reducing analytical uncertainty by 42% (per expert assessment) compared to conventional methods.

0 citationsRead paper