Session: 12-26-01: Data-Driven Modeling and Simulation for Computational Biomedicine

Paper Number: 68056

Start Time: Friday, 03:15 PM

68056 - Data-Driven Modeling of Aortic Dissection Using Deeponet 

Thoracic aortic dissection is responsible for significant morbidity and mortality, yet the underlying process remains poorly understood. Increasing evidence shows localized accumulations of glycosaminoglycans hyaluronan (GAGs) can lead to the delamination of lamellar structures within the intima, where negatively charged aggregates of GAGs imbibe more water and further pressurize the intramural space. The \textit{in-vitro} liquid injection test is used to study the impact of GAGs on progressive delamination, where a fluid is infused into the media at a constant flow rate, while recording the pressure-volume relation and displacement of the specimen. Furthermore, recent \textit{in-silico} studies demonstrated that the spatial distributions of interlamellar struts that connect adjacent elastic lamellae can affect the differential propensity of dissection, resulting in different damage progression in the path of least resistance. More specifically, diverse histological microstructures may lead to differential mechanical behavior during dissection, including the pressure-volume relationship of the injected fluid and the displacement field of the intima layer. Traditional phase-field finite element models can provide a quantitative prediction of the delamination path, yet, the computational cost is expensive.

In this study, we develop a data-driven surrogate model for the delamination process of pressurized aortic walls with differential struts distribution using DeepONet, a powerful operator-regression neural network. The surrogate model is trained to predict the pressure-volume curve of the injected fluid and the displacement field of the wall given a spatial distribution of struts, with synthetic data generated from a validated phase-field finite element model. We use 80\% of data to train the model and 20\% of data as the testing data. Each training case includes the snapshots of progressive delamination at different stages and the interlamellar struts distribution, which is generated by random sampling to mimic the microstructure in the human thoracic artery,

Our results show that the DeepONet model can provide accurate predictions of the quantities of interest (pressure-volume curve and displacement/damage field) for different struts distributions, indicating that the model can effectively extract the underlying functional relationship between distinctive microstructures and their mechanical properties. We also show using our surrogate model can dramatically decrease the computational cost on the phase-field finite element model from days to seconds. Further parametric studies on the generalization ability of our model show that testing error reaches a plateau when the training dataset is sufficiently large.

In the future, our DeepONet model can facilitate surrogate model-based analysis on quantifying biological variability, inverse design, and it has the potential for predicting mechanical properties based on experimental multi-modality data.

Presenting Author: Minglang Yin Brown University

Authors:

Minglang Yin Brown University
Ehsan Ban Yale University
Enrui Zhang Brown University
Bruno Rego Yale University
Jay Humphrey Yale University
George Karniadakis Brown University