Selected work

Representative Papers

Anatomy Guided Coronary Artery Segmentation from CCTA Using Spatial Frequency Joint Modeling

Dec 13, 2025

To address segmentation instability in coronary artery CT angiography—caused by fine vessel calibers, complex branching patterns, ambiguous boundaries, and myocardial interference—this paper proposes a 3D wavelet-based encoder-decoder network with joint spatial-frequency modeling. The method innovatively integrates myocardium-anatomy-guided priors, residual attention enhancement, and 3D inverse wavelet transforms to achieve structural-consistent multi-scale feature encoding and decoding. Additionally, overlapping voxel block training and multi-scale feature fusion are introduced to improve representation of small distal branches. Evaluated on the ImageCAS dataset, the model achieves a Dice coefficient of 0.8082, sensitivity of 0.7946, precision of 0.8471, and a 95th-percentile Hausdorff distance (HD95) of 9.77 mm—significantly outperforming state-of-the-art 3D segmentation methods.

0 citationsRead paper

Blood Pressure Prediction for Coronary Artery Disease Diagnosis using Coronary Computed Tomography Angiography

Dec 11, 2025

To address the high computational cost and clinical deployment challenges of conventional CFD simulations in coronary artery disease (CAD) diagnosis, this paper proposes a diffusion-model-based end-to-end method for predicting pressure distribution directly from coronary computed tomography angiography (CCTA) images. The method automatically reconstructs coronary anatomy, synthesizes high-fidelity hemodynamic data, and trains a diffusion probabilistic model to regress continuous pressure fields—bypassing traditional CFD solvers entirely. It is the first work to adapt diffusion models for continuous physical field regression, integrating geometric reconstruction, deep feature learning, and generative modeling. Evaluated on a synthetic dataset, our approach achieves an R² of 64.42% and RMSE of 0.0974, significantly outperforming state-of-the-art baselines. Moreover, inference is accelerated by two to three orders of magnitude, demonstrating strong potential for large-scale clinical deployment.

0 citationsRead paper

Recent publications

Latest Papers

Anatomy Guided Coronary Artery Segmentation from CCTA Using Spatial Frequency Joint Modeling

Dec 13, 2025

To address segmentation instability in coronary artery CT angiography—caused by fine vessel calibers, complex branching patterns, ambiguous boundaries, and myocardial interference—this paper proposes a 3D wavelet-based encoder-decoder network with joint spatial-frequency modeling. The method innovatively integrates myocardium-anatomy-guided priors, residual attention enhancement, and 3D inverse wavelet transforms to achieve structural-consistent multi-scale feature encoding and decoding. Additionally, overlapping voxel block training and multi-scale feature fusion are introduced to improve representation of small distal branches. Evaluated on the ImageCAS dataset, the model achieves a Dice coefficient of 0.8082, sensitivity of 0.7946, precision of 0.8471, and a 95th-percentile Hausdorff distance (HD95) of 9.77 mm—significantly outperforming state-of-the-art 3D segmentation methods.

0 citationsRead paper

Blood Pressure Prediction for Coronary Artery Disease Diagnosis using Coronary Computed Tomography Angiography

Dec 11, 2025

To address the high computational cost and clinical deployment challenges of conventional CFD simulations in coronary artery disease (CAD) diagnosis, this paper proposes a diffusion-model-based end-to-end method for predicting pressure distribution directly from coronary computed tomography angiography (CCTA) images. The method automatically reconstructs coronary anatomy, synthesizes high-fidelity hemodynamic data, and trains a diffusion probabilistic model to regress continuous pressure fields—bypassing traditional CFD solvers entirely. It is the first work to adapt diffusion models for continuous physical field regression, integrating geometric reconstruction, deep feature learning, and generative modeling. Evaluated on a synthetic dataset, our approach achieves an R² of 64.42% and RMSE of 0.0974, significantly outperforming state-of-the-art baselines. Moreover, inference is accelerated by two to three orders of magnitude, demonstrating strong potential for large-scale clinical deployment.

0 citationsRead paper