Aortic Valve Leaflet Shape Synthesis With Geometric Prior From Surrounding Tissue.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning

Background:

• Classical medical imaging struggles to visualize thin, mobile structures like aortic valve leaflets.
• Accurate biomechanical simulations and personalized therapies require precise knowledge of individual patient anatomy.
• Estimating shapes of inaccessible structures from surrounding tissue geometry presents a significant domain adaptation challenge with limited data.

Purpose of the Study:

• To develop a novel framework for estimating the shape of inassessible anatomical structures using surrounding tissue morphology.
• To address the domain adaptation problem in anatomical shape prediction, particularly in limited data scenarios.
• To enable personalized therapy and improve clinical research through accurate shape estimation.

Main Methods:

• Proposed a deep representation learning framework to predict latent shape representations instead of direct image-to-image mapping.
• Developed and evaluated two distinct approaches within this framework.
• Conducted a proof-of-concept study using ex-vivo porcine data, correlating volumetric ultrasound images with high-resolution leaflet images.

Main Results:

• Demonstrated robust prediction of aortic valve leaflet shape using only surrounding aortic root tissue information.
• Validated the effectiveness of deep representation learning and latent space domain mapping.
• Showcased successful application even in limited data scenarios, highlighting model hyperparameter analysis.

Conclusions:

• The proposed framework successfully bridges the domain gap for anatomical shape estimation.
• Deep representation learning offers a viable solution for predicting shapes of inassessible structures from surrounding data.
• This approach has broad applicability to various anatomical modeling tasks beyond cardiac imaging.