Anatomically Guided Cascaded U-Net Ensemble for Coronary Artery Calcification Segmentation in Cardiac CT

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Accurate segmentation of coronary artery calcifications (CAC) in cardiac CT is crucial but challenging due to issues like class imbalance and small lesion sizes.
• Existing methods often struggle with anatomical ambiguity, leading to suboptimal performance in clinical settings.

Purpose of the Study:

• To develop and evaluate an anatomically guided, cascaded framework for robust and vessel-aware CAC segmentation.
• To improve the accuracy and reliability of CAC detection in cardiac CT imaging.

Main Methods:

• A cascaded framework employing ResU-Net models for heart region isolation and coronary artery segmentation using Frangi vesselness enhancement.
• A heterogeneous U-Net ensemble (U-Net, ResU-Net, Attention U-Net, UNet++) for calcification segmentation within a vessel-constrained region of interest.
• A rank-based selective fusion strategy integrating vessel priors to prioritize accurate predictions and suppress false positives.

Main Results:

• The proposed ensemble model achieved superior performance compared to individual models on the Stanford COCA dataset (Dice 84.25%, sensitivity 87.10%, specificity 98.00%).
• Integrating vessel priors into selective fusion further improved results (Dice 85.50%, sensitivity 88.53%), demonstrating enhanced boundary sharpness and small-lesion detectability.
• The framework showed improved anatomical plausibility, supporting reliable CAC segmentation.

Conclusions:

• Combining dataset-specific anatomical priors with selective ensembling significantly enhances CAC segmentation accuracy and reliability.
• The developed framework offers a promising solution for clinical imaging workflows, improving the detection of coronary artery calcifications.