Poster D6 - AI-Enabled Predictive Modeling: Tracking the Morphological and Stress Response Changes in Abdominal Aortic Aneurysms

Thursday, October 24, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Aakash Kottakota (he/him/his)

Undergraduate Student Researcher
University of Pittsburgh
Gibsonia, Pennsylvania, United States

Co-Author(s)

Jason Lee

Research Assistant
University of Pittsburgh Vascular Bioengineering Laboratory
Santa Clara, California, United States
Pete Gueldner, BS

NSF Graduate Research Fellow
University of Pittsburgh Vascular Bioengineering Laboratory, United States
Nathan Liang, MD MS

Associate Professor of Sugery
UPMC, United States

Primary Investigator(s)

David Vorp

Senior Associate Dean for Research & Facilities
University of Pittsburgh, United States

Last Author(s)

Timothy Chung

Visiting Research Assistant Professor
University of Pittsburgh, United States

Introduction: Abdominal aortic aneurysm (AAA), the 13th leading cause of death in the United States, is an often-fatal condition where the terminal aorta begins to balloon and can rupture[1]. Clinicians typically intervene on patients when the diameter of the AAA reaches a maximum diameter (5.0 cm for women and 5.5 cm for men)[2]. However, the maximum diameter criterion is an antiquated system as 5.6% of men and 11.5% of women suffered ruptures below the operable limits[2]. To address this, our research group is leveraging artificial intelligence tools to track and analyze the morphological and stress changes of the aneurysm and generate a comprehensive forecasting model of the risk of rupture over time. Establishing a neural network trained with patient data and medical imaging allows for the creation of a novel method for tracking the progression of AAA growth over time. This approach leverages data-driven capabilities to establish a more precise predictive model. Afterwards, the resulting forecasting model would be coupled with a neural net trained to associate wall stresses with aortic geometries to create a machine learning model capable of depicting the 3D morphological and stress changes of the aortic aneurysm over time. The model, through machine learning, accurately maps and predicts aneurysm growth and stress changes based on patient-specific data, guiding the timing for optimal medical intervention.

Materials and

Methods: A custom MATLAB (MathWorks, Inc., Natick, MA) script converted a longitudinal database of previously segmented AAA scans into 3D point clouds at each timepoint. To enhance training data, intermediate point clouds were created by interpolating growth stages, using a fixed number of nodes per point cloud for consistency. Cartesian coordinates (X, Y, Z) and metrics (nearest neighbors, unit normal vectors, distance to centerline, etc.) for each node were extracted and appended to a spreadsheet with data for five patients.

The preprocessed data was read into a Python environment (Python Software Foundation, Fredericksburg, VA) and fed into two TensorFlow models on the University of Pittsburgh Center for Research Computing (C2C) Graphics Processing Unit (GPU) cluster. The first model, designed to predict morphology changes over time, used an 80/20 train/test split of the data. It employed Long Short-Term Memory (LSTM) layers to learn and remember long-term dependencies, capturing metric correlations as nodes moved over time.

The second model predicted wall stresses by analyzing spatial, geometric, and biomechanical metrics of nodes across the aneurysm surface with their corresponding MISES stress values. This model was trained on 274 AAA patients using an 80/20 train/test split and utilized three Dense layers to learn intricate patterns in the data.

Both models were unified to use projected surfaces and predict wall stress at intermediate steps, providing a unified analysis. This approach allowed for a comprehensive evaluation of morphology, mean wall stress, and peak wall stresses. Statistical analyses of these metrics were conducted to assess the models’ performance.

Results, Conclusions, and Discussions: Currently, the morphological TensorFlow model with 5 patients is still in the early stages of training, so predictions cannot be discussed yet. Each training cycle takes ~3 minutes, and with around ~20,000 cycles needed, the models have not yet converged. However, previously predicted geometries from a morphological model that needed to be retrained for each new patient will be discussed. These predictions, from 3 separate cases, were tested using the stress model.

Figure 1 illustrates two models from the same case: the ground truth aneurysm from the original patient scan and the predicted geometry from an initial morphological prediction. Both models have been subjected to stress predictions, and the results are represented as a colormap on the surface of each aneurysm. Qualitatively, the stress distributions exhibit a similar pattern on the aneurysm surfaces, with no significant differences.

Table 1 shows quantitative metrics across the 3 separate cases. A paired, one-tailed heteroscedastic t-test was performed for both the peak and mean wall stresses for the 3 cases. P-values of 0.57 and 0.28 for the peak and mean wall stresses, respectively, demonstrate no statistically significant differences between the ground truth and predicted metrics. This provides reasonable confidence in the model’s predictive capabilities.

Further work will be done on enhancing the efficiency of the 5-patient morphological model by adjusting the model architecture and taking advantage of more powerful hardware available through the CRC to facilitate quicker results.

Acknowledgements (Optional): Funding for this project was provided by the University of Pittsburgh Swanson School of Engineering through the SURI (Summer Undergraduate Research Internship) grant. This research was supported in part by the University of Pittsburgh Center for Research Computing, RRID:SCR_022735, through the resources provided. Specifically, this work used the H2P cluster, which is supported by NSF award number OAC-2117681. We would also like to acknowledge the NTA3CT trial for providing the longitudinal medical imaging database.

References:
[1] Umebayashi, Ryoko et al. “Abdominal aortic aneurysm in aged population.” Aging vol. 10,12 (2018): 3650-3651. doi:10.18632/aging.101702
[2] Kontopodis, Nikolaos et al. “The Obsolete Maximum Diameter Criterion, the Evident Role of Biomechanical (Pressure) Indices, the New Role of Hemodynamic (Flow) Indices, and the Multi-Modal Approach to the Rupture Risk Assessment of Abdominal Aortic Aneurysms.” Annals of vascular diseases vol. 11,1 (2018): 78-83. doi:10.3400/avd.ra.17-00115