Predictive Constitutive Modelling of Atrial Appendage Tissue Using Deep Learning

The department specializes in the modeling and experimental characterization of soft biological tissues, with applications to the cardiovascular systems.

Context

Stroke is the third leading cause of mortality in France. Approximately one-third of strokes are associated with atrial fibrillation, in which the left atrial appendage (AA) becomes a primary site of thrombus formation. The mechanics of the AA remain poorly understood, particularly due to its highly trabeculated architecture, which induces strong and spatially heterogeneous thickness variations.

To better characterize the role of macro-structure in the mechanical response, previous experimental studies have combined full-field 3D thickness mapping, uniaxial mechanical testing, and numerical modelling to assess how structural variability influences the apparent anisotropy of atrial appendage tissue.

A recent hybrid modelling framework introduced by Holzapfel et al. [1] combines deep learning with mechanical testing, histology, and second-harmonic generation imaging. While their model was trained on microstructural features from 27 tissue samples, the present work aims to explore whether comparable predictive performance can be achieved using 3D scans alone, which contain information about the trabeculated structure (see Fig.). For this purpose, we will rely on a dataset of 80 samples, each including uniaxial mechanical tests and high-resolution 3D surface scans.

The aim of this internship is to develop a deep neural network capable of predicting the parameters of a continuum-mechanical constitutive law for atrial appendage tissue, based on local thickness maps extracted from 3D scans. Once established, this predictive capability would considerably improve the mechanical fidelity of cardiac computational models.

Tasks

• Familiarization with the biomechanics context: understanding soft tissue mechanics, constitutive laws, and uniaxial testing.
• Literature review: deep-learning approaches for constitutive modelling, physics-informed machine learning models.
• Data preparation: performing data augmentation.
• Deep learning model development: designing and training model to predict mechanical responses (stress–stretch curves) from thickness maps.
• Model evaluation and comparison: analysing performance compared to standard fitting methods.

Reference
[1] Holzapfel Gerhard A., et al. (2021), Predictive constitutive modelling of arteries by deep learning, J. R. Soc. Interface.1820210411, http://doi.org/10.1098/rsif.2021.0411

Interest and previous experience in machine learning approaches (e.g. student project). Solid programming skills; basic knowledge of continuum mechanics; interest in biomedical applications.