Selected work

Representative Papers

Clinical Risk-Aware Multi-Level Grading for Coronary Artery Stenosis through Curved Feature Reconstruction

Jun 29, 2026

This study addresses the challenges in existing methods for multi-class coronary artery stenosis grading using fused CCTA and 3D SCPR images, which suffer from inaccurate alignment and neglect of clinically relevant risk differences across stenosis severity levels. To overcome these limitations, the authors propose a Curved Feature Reconstruction (CFR) module that leverages geometric priors of vessel centerlines to achieve point-wise alignment and feature fusion between the two modalities. Additionally, a Clinical Risk-aware (CR) loss function is introduced to explicitly embed differential clinical risks into the training process. Integrated within a deep multi-class classification network, the proposed approach significantly outperforms current methods on an internal dataset. Ablation studies confirm the effectiveness of both the CFR module and the CR loss, demonstrating improved alignment with clinical requirements in stenosis grading.

0 citationsRead paper

MammoExpert: Benchmarking Chain-of-Thought Reasoning in Mammography Diagnosis

Jun 19, 2026

Existing mammography datasets lack large-scale, high-quality annotations with structured diagnostic reasoning, limiting the application of AI in breast cancer subtype identification and interpretable diagnosis. This study introduces the first mammography dataset annotated with a three-stage Chain-of-Thought (CoT) reasoning framework, encompassing 67 WHO pathological subtypes. Nine senior radiologists annotated each case with 42 imaging features, systematically organized into three reasoning phases: initial observations, factual assessments, and diagnostic synthesis. This structured CoT annotation establishes a new benchmark for interpretable AI, yielding a 4% improvement in classification accuracy on the MammoExpert test set. When combined with CBIS-DDSM for training, the approach achieves performance gains of 7.1%, and further enhances accuracy by 6.9% and 6.7% on INBreast and Vindr, respectively.

0 citationsRead paper

UMind-VL: A Generalist Ultrasound Vision-Language Model for Unified Grounded Perception and Comprehensive Interpretation

Nov 27, 2025

In ultrasound medicine, low-level perception (e.g., segmentation, detection) and high-level clinical interpretation (e.g., diagnosis, reasoning) have long remained disjointed. To bridge this gap, we propose the first vision-language unified foundation model tailored for ultrasound. Our method introduces: (1) a lightweight dynamic convolutional mask decoder that generates task-adaptive dynamic kernels conditioned on large language model outputs; (2) task-specific tokens enabling end-to-end joint modeling of segmentation, detection, biometric measurement, and diagnostic reasoning; and (3) a multimodal alignment training paradigm, pretrained and fine-tuned on the large-scale ultrasound dataset UMind-DS. Experiments demonstrate that our model surpasses general-purpose multimodal models across multiple benchmarks and matches or exceeds state-of-the-art task-specific models—while exhibiting strong generalization and clinical applicability.

0 citationsRead paper

A Chain-of-thought Reasoning Breast Ultrasound Dataset Covering All Histopathology Categories

Sep 21, 2025

Existing public breast ultrasound (BUS) datasets suffer from limited scale, coarse-grained annotations, and insufficient coverage of rare pathological subtypes, hindering the clinical deployment of interpretable AI. To address these limitations, we introduce BUS-CoT—the first high-quality, Chain-of-Thought (CoT)-enabled BUS benchmark dataset, comprising 11,439 expert-annotated images spanning all 99 histopathologically confirmed tissue types. BUS-CoT features a novel four-stage expert annotation schema—“Observation → Feature → Diagnosis → Histopathology”—integrating multi-level clinical knowledge with rigorous validation protocols. This structured reasoning framework significantly enhances model performance on rare lesion classification and improves cross-scenario generalizability. BUS-CoT establishes an open, authoritative evaluation benchmark for developing interpretable, robust AI systems in breast cancer diagnosis, enabling transparent, clinically grounded decision-making.

0 citationsRead paper

DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image

Jun 06, 2025

Fine vessel structures in 3D medical imaging are prone to fragmentation and discontinuity when extracted via pixel-wise segmentation. To address this, we propose a deformable centerline representation—a continuous, topology-aware geometric modeling paradigm. Our method introduces a novel node-edge graph-based centerline formulation that intrinsically ensures structural connectivity, noise robustness, and clinical interpretability. Integrated with graph neural networks, deformable convolutions, and differentiable curve parameterization, it forms an end-to-end generative and refinement framework. A cascaded training strategy is further designed to enforce geometric priors—including smoothness, continuity, and anatomical plausibility. Evaluated on four benchmark 3D vascular datasets, our approach achieves significant improvements over state-of-the-art methods in both centerline accuracy and completeness. Moreover, curved planar reformation (CPR) surface reconstruction demonstrates high anatomical fidelity and clinical utility, validating its practical applicability in interventional planning and diagnosis.

0 citationsRead paper

Recent publications

Latest Papers

Clinical Risk-Aware Multi-Level Grading for Coronary Artery Stenosis through Curved Feature Reconstruction

Jun 29, 2026

This study addresses the challenges in existing methods for multi-class coronary artery stenosis grading using fused CCTA and 3D SCPR images, which suffer from inaccurate alignment and neglect of clinically relevant risk differences across stenosis severity levels. To overcome these limitations, the authors propose a Curved Feature Reconstruction (CFR) module that leverages geometric priors of vessel centerlines to achieve point-wise alignment and feature fusion between the two modalities. Additionally, a Clinical Risk-aware (CR) loss function is introduced to explicitly embed differential clinical risks into the training process. Integrated within a deep multi-class classification network, the proposed approach significantly outperforms current methods on an internal dataset. Ablation studies confirm the effectiveness of both the CFR module and the CR loss, demonstrating improved alignment with clinical requirements in stenosis grading.

0 citationsRead paper

MammoExpert: Benchmarking Chain-of-Thought Reasoning in Mammography Diagnosis

Jun 19, 2026

Existing mammography datasets lack large-scale, high-quality annotations with structured diagnostic reasoning, limiting the application of AI in breast cancer subtype identification and interpretable diagnosis. This study introduces the first mammography dataset annotated with a three-stage Chain-of-Thought (CoT) reasoning framework, encompassing 67 WHO pathological subtypes. Nine senior radiologists annotated each case with 42 imaging features, systematically organized into three reasoning phases: initial observations, factual assessments, and diagnostic synthesis. This structured CoT annotation establishes a new benchmark for interpretable AI, yielding a 4% improvement in classification accuracy on the MammoExpert test set. When combined with CBIS-DDSM for training, the approach achieves performance gains of 7.1%, and further enhances accuracy by 6.9% and 6.7% on INBreast and Vindr, respectively.

0 citationsRead paper

UMind-VL: A Generalist Ultrasound Vision-Language Model for Unified Grounded Perception and Comprehensive Interpretation

Nov 27, 2025

In ultrasound medicine, low-level perception (e.g., segmentation, detection) and high-level clinical interpretation (e.g., diagnosis, reasoning) have long remained disjointed. To bridge this gap, we propose the first vision-language unified foundation model tailored for ultrasound. Our method introduces: (1) a lightweight dynamic convolutional mask decoder that generates task-adaptive dynamic kernels conditioned on large language model outputs; (2) task-specific tokens enabling end-to-end joint modeling of segmentation, detection, biometric measurement, and diagnostic reasoning; and (3) a multimodal alignment training paradigm, pretrained and fine-tuned on the large-scale ultrasound dataset UMind-DS. Experiments demonstrate that our model surpasses general-purpose multimodal models across multiple benchmarks and matches or exceeds state-of-the-art task-specific models—while exhibiting strong generalization and clinical applicability.

0 citationsRead paper

A Chain-of-thought Reasoning Breast Ultrasound Dataset Covering All Histopathology Categories

Sep 21, 2025

Existing public breast ultrasound (BUS) datasets suffer from limited scale, coarse-grained annotations, and insufficient coverage of rare pathological subtypes, hindering the clinical deployment of interpretable AI. To address these limitations, we introduce BUS-CoT—the first high-quality, Chain-of-Thought (CoT)-enabled BUS benchmark dataset, comprising 11,439 expert-annotated images spanning all 99 histopathologically confirmed tissue types. BUS-CoT features a novel four-stage expert annotation schema—“Observation → Feature → Diagnosis → Histopathology”—integrating multi-level clinical knowledge with rigorous validation protocols. This structured reasoning framework significantly enhances model performance on rare lesion classification and improves cross-scenario generalizability. BUS-CoT establishes an open, authoritative evaluation benchmark for developing interpretable, robust AI systems in breast cancer diagnosis, enabling transparent, clinically grounded decision-making.

0 citationsRead paper

DeformCL: Learning Deformable Centerline Representation for Vessel Extraction in 3D Medical Image

Jun 06, 2025

Fine vessel structures in 3D medical imaging are prone to fragmentation and discontinuity when extracted via pixel-wise segmentation. To address this, we propose a deformable centerline representation—a continuous, topology-aware geometric modeling paradigm. Our method introduces a novel node-edge graph-based centerline formulation that intrinsically ensures structural connectivity, noise robustness, and clinical interpretability. Integrated with graph neural networks, deformable convolutions, and differentiable curve parameterization, it forms an end-to-end generative and refinement framework. A cascaded training strategy is further designed to enforce geometric priors—including smoothness, continuity, and anatomical plausibility. Evaluated on four benchmark 3D vascular datasets, our approach achieves significant improvements over state-of-the-art methods in both centerline accuracy and completeness. Moreover, curved planar reformation (CPR) surface reconstruction demonstrates high anatomical fidelity and clinical utility, validating its practical applicability in interventional planning and diagnosis.

0 citationsRead paper