Using Artificial Intelligence to Segment & Predict Low Grade Glioma Growth*

Session Number

3

Advisor(s)

Dr. Ankush Bhatia, University of Wisconsin-Madison

Location

A119

Discipline

Medical and Health Sciences

Start Date

15-4-2026 2:15 PM

End Date

15-4-2026 3:00 PM

Abstract

For the first half of my two-year SIR project, I focused on learning about previous research adjacent to my project as I underwent the onboarding and other necessary tasks required to take my project from the preparation phase and begin my work, which started in March. For the  majority of the year, I read research papers and familiarized myself with the artificial intelligence segmentation algorithms that I will eventually employ on the datasets I have received from UW-Madison. Specifically, I trained an nnU-Net 3D segmentation model on 200 deidentified MRIs (800 total images) from a publicly available dataset (70% training, 20% Validation, 10% testing), I was able to achieve a modest DICE score of 0.7. With more MRI to train from, along with further optimizations in the initial parameters, and input from my mentor and his team, I am confident that I can use the knowledge and experience gained this year to increase the accuracy of the model, and begin to create a prediction model once the accuracy is satisfactory.

Share

COinS
 
Apr 15th, 2:15 PM Apr 15th, 3:00 PM

Using Artificial Intelligence to Segment & Predict Low Grade Glioma Growth*

A119

For the first half of my two-year SIR project, I focused on learning about previous research adjacent to my project as I underwent the onboarding and other necessary tasks required to take my project from the preparation phase and begin my work, which started in March. For the  majority of the year, I read research papers and familiarized myself with the artificial intelligence segmentation algorithms that I will eventually employ on the datasets I have received from UW-Madison. Specifically, I trained an nnU-Net 3D segmentation model on 200 deidentified MRIs (800 total images) from a publicly available dataset (70% training, 20% Validation, 10% testing), I was able to achieve a modest DICE score of 0.7. With more MRI to train from, along with further optimizations in the initial parameters, and input from my mentor and his team, I am confident that I can use the knowledge and experience gained this year to increase the accuracy of the model, and begin to create a prediction model once the accuracy is satisfactory.