Multi-stage learning for segmentation of aortic dissections using a prior aortic anatomy simplification.

Area of Science:

• Cardiovascular Imaging and Intervention
• Medical Artificial Intelligence
• Computational Anatomy

Background:

• Aortic dissection (AD) is a critical cardiovascular condition with high mortality.
• Accurate 3D reconstruction of AD from CT-angiography is vital for clinical procedures but lacks efficient tools.
• Existing methods struggle with generalized and precise segmentation of AD structures.

Purpose of the Study:

• To develop and validate a novel multi-stage segmentation framework for type B AD.
• To accurately segment the true lumen (TL), false lumen (FL), and all branches (BR) of the dissected aorta.
• To enhance 3D reconstruction for improved clinical and surgical planning in AD patients.

Main Methods:

• A multi-stage segmentation framework utilizing two cascaded neural networks.
• Implementation of an aortic straightening method based on prior vascular anatomy to simplify geometry.
• Segmentation of the aortic trunk and branches, followed by dual lumen separation.

Main Results:

• Achieved high mean Dice scores: 0.96 for TL, 0.95 for FL, and 0.89 for BR.
• Outperformed end-to-end and non-straightening multi-stage methods in dual-lumen segmentation.
• Demonstrated superior performance on a multi-center dataset of 120 patients.
• Enabled better identification and quantification of global and local AD features.

Conclusions:

• The proposed straightening-based multi-stage framework significantly improves segmentation accuracy for type B AD.
• This novel approach offers advantages over existing deep learning methods for AD segmentation.
• The enhanced 3D reconstructions can effectively assist in clinical procedures and surgical planning for AD.