Type and Shape Disentangled Generative Modeling for Congenital Heart Defects

Congenital heart diseases (CHDs) encompass a wide range of cardiovascular structural abnormalities, and CHD patients exhibit complex and unique malformations. Analysis of these unique cardiac anatomies can greatly improve diagnosis and treatment planning. However, CHDs are often rare, making it extremely challenging to acquire patient cohorts of sufficient size. Although generative modeling of cardiac anatomies capturing patient variations can generate virtual cohorts to facilitate in-silico clinical trials, prior approaches were largely designed for normal anatomies and cannot readily model the vast topological variations seen in CHD patients. Therefore, we propose a generative approach that models cardiac anatomies for different CHD types and synthesizes CHD-type specific shapes that preserve the unique topology. Our deep learning (DL) approach represents whole heart shapes implicitly using signed distance fields (SDF) based on CHD-type diagnosis, which conveniently captures divergent anatomical variations across different types. We then learn invertible deformations to deform the learned type-specific anatomies and reconstruct patient-specific geometries. Our approach has potential applications in augmenting the image-segmentation pairs for rarer CHD types for cardiac segmentation and generating CHD cardiac meshes for computational simulations. Our source code is available at https://github.com/fkong7/SDF4CHD .