Learning three-dimensional aortic root assessment based on sparse annotations.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Cardiovascular imaging

Background:

• Accurate analysis of the aorta and left ventricular outflow tract (LVOT) is essential for transcatheter aortic valve implantation (TAVI) risk assessment and planning.
• Current segmentation methods often require extensive annotated data, posing a challenge for supervised deep learning models.

Purpose of the Study:

• To develop a method that minimizes annotation complexity for training automated 3D segmentation networks of the aortic root and LVOT.
• To facilitate the generation of training data for patient-specific anatomical analysis.

Main Methods:

• Proposed a novel approach using 2D cross-sectional annotations and point cloud-based surface reconstruction.
• Trained a fully automatic 3D segmentation network for the aortic root and LVOT.
• Derived clinically relevant parameters for TAVI planning from segmentation results.

Main Results:

• Achieved high inter-observer agreement for 2D annotations (Dice similarity coefficient [DSC]: 0.94).
• The segmentation model demonstrated a DSC of 0.90 and an average surface distance of 0.96 mm.
• The maximum diameter difference for the aortic annulus between prediction and annotation was 0.45 mm.

Conclusions:

• The presented approach enables reproducible annotations, facilitating the training of accurate segmentation models.
• The segmentation results provide reproducible and quantifiable measurements crucial for TAVI planning.
• This method simplifies data generation for analyzing aortic and LVOT anatomy.