Generation of virtual abdominal aortic aneurysm shape evolution using a physics-based G&R model and its applications...

The computational growth and remodeling (G&R) of aortic aneurysms has been extensively investigated to better understand vascular disease progression and improve surgical intervention strategies. This study introduces a framework to simulate abdominal aortic aneurysms (AAAs) shape evolution using a physics-based G&R model, addressing the limitation posed by the scarcity of large-scale clinical imaging datasets. By generating a virtual cohort of AAAs, this study bridges gaps in data availability and enables leveraging of physics-based G&R simulations and data-driven machine learning models for predicting aneurysm growth. The G&R modeling framework utilizes key factors influencing aneurysm disease progression, such as elastin damage function and collagen production rate, which significantly affect the growth and evolution of AAAs. A new elastin degradation formula within the arterial G&R model was developed to generate realistic geometries, capturing clinically relevant features such as asymmetry and tortuosity. By systematically varying coefficients for the elastin damage function and collagen production rate, 200 distinct G&R simulations were performed, yielding a diverse set of AAA geometries. This dataset is further extended using kriging, a statistical method for spatial interpolation, to generate a large in silico cohort. The synthetic dataset, combined with clinical data, was used to train and validate four different machine learning models: Deep Belief Network (DBN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) network, and Gated Recurrent Unit (GRU) network. These machine learning models were trained in a two-step process to quickly predict the maximum diameter and growth rate of AAAs based on a patient's aneurysmal geometrical characteristics at prior time points. Using medical images from 25 patients for model fine-tuning and validation, the results demonstrated promising predictive performance. Specifically, the RNN model achieved R² scores of 0.90 and 0.89 for the prediction of the maximum diameter and the growth rate of AAA, respectively. The DBN model yielded R² scores of 0.82 and 0.86 for the maximum diameter and growth rate, respectively. LSTM showed R² scores of 0.92 for the maximum diameter and 0.84 for the growth rate, while the GRU model attained R² scores of 0.89 for the maximum diameter and 0.76 for the growth rate. Overall, the findings indicate that the proposed machine learning approach, utilizing data from G&R simulations and clinical images, can predict AAA growth and maximum diameter with notable accuracy, offering valuable insights for clinical decision-making.