Topology-oriented foreground focusing network for semi-supervised coronary artery segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate coronary artery (CA) segmentation from coronary-computed tomography angiography (CCTA) is vital for diagnosing coronary diseases and surgical planning.
• Current segmentation methods struggle with maintaining topological consistency, distinguishing CAs from other tubular structures, and limited labeled data.

Purpose of the Study:

• To develop a novel semi-supervised deep learning network, TOFF-Net, for robust and accurate automatic CA segmentation.
• To address challenges in topological consistency, irrelevant vessel interference, and data scarcity in CCTA image segmentation.

Main Methods:

• Proposed a topology-oriented foreground focusing network (TOFF-Net) incorporating a vascular connectivity preservation (VCP) loss.
• Introduced an irrelevant vessels removal (IVR) module for enhanced local and global vessel detail integration.
• Implemented a foreground label migration and focusing (FLMF) module with Pioneer-Imitator learning for semi-supervised learning.

Main Results:

• TOFF-Net achieved state-of-the-art performance in CA segmentation across multiple datasets.
• Demonstrated high topological consistency and significantly reduced false positives from irrelevant tubular structures.
• Validated effectiveness on in-house and public CCTA datasets.

Conclusions:

• TOFF-Net effectively addresses critical challenges in automatic coronary artery segmentation from CCTA.
• The proposed methods show significant improvements in accuracy and topological integrity.
• TOFF-Net exhibits potential for segmenting other vascular structures beyond coronary arteries.