Dissected aorta segmentation using convolutional neural networks.

Area of Science:

• Medical Imaging
• Cardiovascular Pathology
• Artificial Intelligence in Medicine

Background:

• Aortic dissection is a life-threatening condition characterized by a tear in the aorta's intimal layer.
• Accurate 3-D morphological understanding of the dissected aorta is crucial for effective treatment planning.
• Current segmentation methods require improvement for precise aortic dissection analysis.

Purpose of the Study:

• To develop a deep-learning-based algorithm for automatic segmentation of dissected aortas in computed tomography angiography (CTA) images.
• To enhance the accuracy and robustness of aortic dissection segmentation.
• To provide a tool for better understanding the 3-D morphology of dissected aortas.

Main Methods:

• A two-step deep learning approach utilizing 3-D and 2-D convolutional neural networks (CNNs).
• Initial 3-D CNN divides the aorta into anatomical portions, followed by 2-D CNNs (PSPnet-based) for detailed segmentation.
• Integration of an edge extraction branch to improve intimal flap segmentation accuracy.

Main Results:

• The proposed algorithm achieved an average Dice index exceeding 92%, demonstrating high segmentation performance.
• The combined 3-D and 2-D model approach outperformed 3-D or 2-D only models in accuracy and robustness.
• The edge extraction branch significantly improved the Dice index near aortic boundaries from 73.41% to 81.39%.

Conclusions:

• The developed deep learning algorithm effectively segments dissected aortas, accurately capturing complex structures.
• The method demonstrates high performance in identifying the intimal flaps while minimizing false positives.
• This tool aids in the precise morphological assessment of dissected aortas, supporting clinical decision-making.