Patient-informed generative modeling of cardiac anatomy based on normalizing flows

Konstantinos Kevopoulos, Mathias Peirlinck, Alexander Heinlein

Keywords: conditional generative model, cardiac anatomy, normalizing flows, virtual cohort generation, dimensionality reduction

Abstract:

It is known that cardiac shapes vary for individuals with different characteristics, having complex variation patterns. Understanding these differences in cardiac anatomy is critical, as the heart’s geometry directly influences its functional performance, and deviations from expected shape patterns may serve as early indicators of cardiovascular disease.
Consequently, several deep learning-based methodologies have emerged to study cardiac shape variability for a wide range of applications, such as cardiovascular disease detection, generation of virtual patient cohorts, patient-specific modeling, and cardiac shape analysis.
In this work, we propose a novel, conditional generative framework based on normalizing flows, to model cardiac shape variations in healthy populations and generate synthetic anatomies according to patient-specific characteristics. This model is integrated in a more general pipeline, including a geometry-aware dimensionality reduction stage based on diffeomorphic registration. We conduct two experiments to evaluate the performance of the framework: the first compares the myocardial mass between real and synthetic shapes, and the second offers a mathematically concrete manner to compare the overall structure of real and generated anatomies.
In both experiments, we show that the proposed framework outperforms baseline models, such as variational autoencoders and generative adversarial networks with similar architecture, in terms of Wasserstein distance. We find that the developed method approximates the patient-informed distribution of cardiac shapes more closely, and generates virtual cohorts that capture real shape variability more effectively than the aforementioned baseline models. Additionally, we explore the interpretability of the latent space derived from our pipeline, and demonstrate that the resulting reduced representations encode meaningful shape information. Finally, we showcase that the framework offers efficient training and robust performance across hyperparameter variations, supporting its applicability in clinical settings.

Submission Number: 175