(D-150) AI-Driven Growth Predictions: Tracking the Evolution of Morphology in Abdominal Aortic Aneurysms Over Time

Abdominal aortic aneurysm (AAA) rupture, a condition in which a section within the abdominal region of the aorta swells and eventually bursts, is the 13^th leading cause of death in the United States¹. AAA ruptures have a mortality rate of approximately 40% when the patient is hospitalized accordingly, but that figure rises substantially (~80%) when the rupture occurs pre-hospitalization². AAA diagnoses have traditionally been surveilled and classified based on risk of rupture, with the maximum diameter criterion being used to determine if a sufficient threshold has been reached to warrant surgery. However, this antiquated approach is fallible with studies demonstrating that 5.6% of men and 11.5% of women had ruptured aneurysms that were under the maximum diameter criterion³. Our group aims to reduce these rupture rates by utilizing machine learning algorithms to generate a morphological growth tracking model of an aneurysm on a per case basis.  The use of machine learning has escalated drastically in recent years due to its ability to analyze nearly imperceptible patterns. By utilizing this technology, a unique approach where a neural network is trained with patient data and medical imaging datasets can be used to generate an accurate, personalized model of a patient’s aneurysm growth toward a clinical outcome.

A custom MATLAB (MathWorks Inc., Natick, MA) script was developed to extract Cartesian coordinates (X, Y, and Z) from a longitudinal database of patient scans that translates them into 3D point clouds per available timepoint⁴. To enhance the predictive capabilities of the model, the script generated intermediate point clouds between the original timepoints using linear interpolation to increase data for future predictions.

A TensorFlow (Alphabet Inc., Mountain View, CA) model was then initialized in a Python (Python Software Foundation, Fredericksburg, VA) script which was fed the coordinate data at each timepoint in a series. The TensorFlow model, utilizing an 80/20 train /test split when being trained, then activated a Long Short-Term Memory layer which utilizes a unique architecture to develop context as to how an individual node is moving and generates an algorithm for predicting where the node will be at future timepoints. The model was then validated using the last 5 weeks of available data with the predicted final timepoint compared against the original final timepoint to assess accuracy.

This prediction process was repeated for every node in the model that resulted in future predictions of the nodal coordinates based on temporal prediction of displacement. Afterwards, another MATLAB script read in all the predicted nodes and organized them based on their corresponding future timestep. It then generated a point cloud using the filtered points for each future timestep for easy visualization of the predicted growth of the aneurysm. The whole process is then repeated on a per case basis.

Results and Discussion: Figure 1a displays the actual final model with a color overlay indicating individual nodal growths from the initial scan for a case. Figure 1b, which shows the predicted final model with a similar color overlay, reveals a visual similarity between the two models.       

In Figure 1c, an overlay of the two models demonstrates minimal deviation, with an average displacement difference of 0.16 mm and standard deviation of 0.39 mm per node. Notably, the predicted model consistently underpredicts at the superior and inferior edges, while the region encompassing the main aneurysm exhibits almost identical displacement patterns.

Figure 2a presents a scatter plot demonstrating a strong linear relationship between the original and predicted nodal displacements from the initial scan, highlighted by the least squares regression line overlaid on top of the plot. Figure 2b confirms this linear relationship with a low p-value (< 0.0001) and high R² value (0.923), indicating a robust linear correlation between the predicted and original displacements.

Conclusions: The model’s accuracy is supported by the low p-value and high R² value. To improve the efficiency of model generation, which currently takes 30-34 hours, access to more powerful hardware will be beneficial. One limitation of the current model is its assumption of a linear time progression between each timestep, whereas the actual could be non-linear and should be considered.

There are plans to integrate an AI stress analysis currently being developed to create a more comprehensive predictive model which can track both morphological and stress-based changes across an aneurysm over time.

Future work will focus on determining how far into the future the model is capable of accurately predicting changes, contributing to a better understanding of its effectiveness.

Funding for this project was provided by the University of Pittsburgh Swanson School of Engineering through the SURI (Summer Undergraduate Research Internship) grant. The NVIDIA RTX A5000 GPU (Graphics Processing Unit) used in regression testing was provided by NVIDIA through an NVIDIA Academic Hardware Grant Award. We would also like to acknowledge the NTA3CT trial for providing the longitudinal medical imaging database.