Item Details

title

A NOVEL PHYSICS-INFORMED NEURAL NETWORK FRAMEWORK TO MODEL LESION GROWTH FOR DISTINGUISHING TUMOR RECURRENCE FROM RADIATION NECROSIS

author

Gorlewski, Samantha

abstract

In recent years, metastatic brain tumors have been treated increasingly with stereotactic radiosurgery for local tumor control. While this treatment is effective and confers fewer side effects than whole brain radiation, it is not without its risks. One potential side effect is radiation necrosis (RN), in which nearby vasculature and brain parenchyma are killed. Additionally, stereotactic radiosurgery does not prevent tumor recurrence (TR) after treatment. Although they often have indistinguishable radiographic findings, the two pathologies have disparate treatments. Many models are capable of modeling tumor behavior, but these models are held back from clinical translation by long computational times. Integrating neural networks would expedite the solution process, but traditional neural networks offer little insight into how the network found its conclusion. Physics-informed neural networks (PINNs) seek to remedy this by directly incorporating physical laws into the training process. As such, we developed a novel machine learning methodology that incorporates PINNs into a reaction-diffusion tumor growth model. We also explored the feasibility of using PINNs to implement a biomechanical variant of the reaction-diffusion model. We found that with optimal network hyperparameters, our PINN framework could accurately replicate reaction-diffusion tumor growth. We also demonstrated that, even with non-ideal hyperparameters, our PINN’s forward pass could recreate mechanically coupled growth behavior. This study shows the capability of PINNs in modeling complex lesion behaviors as an expedited alternative to traditional computational methods.

subject

Cancer research

Machine learning

Neural networks

Physics-informed neural networks

contributor

Weis, Jared A. (advisor)

Bahrami, Mohsen (committee member)

Zhao, Dawen (committee member)

date

2024-09-13T08:36:55Z (accessioned)

2024 (issued)

degree

Biomedical Engineering (discipline)

embargo

2025-09-12 (terms)

2025-09-12 (liftdate)

identifier

http://hdl.handle.net/10339/109863 (uri)

language

en (iso)

publisher

Wake Forest University

type

Thesis