Graduation Year

2021

Document Type

Thesis

Degree

M.S.M.E.

Degree Name

MS in Mechanical Engineering (M.S.M.E.)

Degree Granting Department

Mechanical Engineering

Major Professor

Wenbin Mao, Ph.D.

Committee Member

Nathan Gallant, Ph.D.

Committee Member

Shaun Canavan, Ph.D.

Committee Member

Yu Sun, Ph.D.

Keywords

Cardiac mechanics, Clinical metrics, Constitutive modeling, Constitutive-based deep learning, Hyperelastic material, Machine learning

Abstract

Modern breakthroughs in biomedical engineering, computer science, and data mining have created new opportunities for detecting important mechanical properties of soft tissues that can be employed to identify possible signs of diseases or physiological difficulties. However, the scarcity of different mechanical properties obtained through noninvasive testing emphasizes the importance of incorporating authentic biological data into computer models capable of replicating the behavior of soft tissues.

The field of continuum theory of large deformation hyperactivity permits the formulation of highly descriptive mathematical research and computational models capable of perfectly describing the minute mechanical characteristics of soft materials. By including features about the tissue's morphology into its internal constitution, constitutive models effectively associate applied loading to the material's mechanical function, allowing for accurate reinterpretation and analysis of tissue behavior.

The advancement of sophisticated analytics techniques, such as machine learning and high-performance computational science, has sparked interest in data-driven computational modeling to extract fast and valid observations of complicated systems. Additionally, machine learning approaches are proving beneficial in a range of biological applications.

In this study, we present a physics-based deep learning approach for predicting material parameters for the active constitutive model based exclusively on a few clinical parameters. This model is also capable of predicting the fiber orientation of epicardium and endocardium of the myocardium wall which are used in an algorithm for computing the fiber distribution through a given geometry. The data used to train the deep learning model was gathered by finite element simulations developed with state-of-the-art passive and active constitutive modeling of the myocardium. Moreover, a variational autoencoder is presented and utilized to produce realistic PV loops based on set of cardiovascular metrics for testing our final model.

The results demonstrate that the deep learning model can estimate active material parameters and fiber orientations with high accuracy for any specified conditions. Moreover, the constitutive model was able to replicate material characteristics based on experimental data presented in the literature and follow expected behavior under diverse pre-set cardiac cycles. Further research is suggested to optimize the proposed deep learning model and its applicability to patient-specific scenarios to account for passive properties and non-idealized fiber distribution by extending its training dataset with data obtained by clinical procedures.

Share

COinS