Generative 3D Cardiac Shape Modelling for in-silico Trials

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computer Vision

Background:

• Accurate modeling of patient-specific anatomy is crucial for medical research and virtual trials.
• Generating diverse and realistic anatomical shapes presents a significant computational challenge.

Purpose of the Study:

• To propose a novel deep learning method for modeling and generating synthetic aortic shapes.
• To represent complex aortic geometries using neural signed distance fields.
• To enable the creation of patient-specific anatomical models for in-silico studies.

Main Methods:

• A deep learning approach representing shapes as the zero-level set of a neural signed distance field.
• Conditioning the model with trainable embedding vectors that encode geometric features.
• Training the network on aortic root meshes derived from CT images.
• Enforcing constraints on the neural field and its spatial gradient during training.

Main Results:

• The proposed model demonstrates high-fidelity representation of aortic shapes.
• Generated synthetic shapes closely resemble real patient anatomies.
• The method allows for the generation of novel aortic geometries through sampling embedding vectors.

Conclusions:

• The deep learning method effectively models and generates synthetic aortic shapes.
• The generated shapes are suitable for applications such as in-silico trials.
• This work advances the potential for data-driven anatomical modeling in medicine.