Date of Award
Fall 2019
Document Type
Thesis - Restricted
Degree Name
Master of Science (MS)
Department
Biomedical Engineering
First Advisor
LaViolette, Peter S.
Second Advisor
Cochran, Elizabeth
Third Advisor
Gilat-Schmidt, Taly
Abstract
Glioblastoma (GBM) is an aggressive cancer with an average 5-year survival rate of about 5%. Following treatment with surgery, radiation, and chemotherapy, diagnosing tumor recurrence requires serial magnetic resonance imaging (MRI) scans. Infiltrative tumor cells beyond gadolinium enhancement on T1-weighted MRI are difficult to detect. This study therefore aims to improve tumor detection beyond traditional tumor margins. To accomplish this, a neural network model was trained to classify tissue samples as ‘tumor’ or ‘not tumor’. This model was then used to classify thousands of tiles from histology samples acquired at autopsy with known MRI locations on the patient’s final clinical MRI scan. This combined radiological-pathological (rad-path) dataset was then treated as a ground truth to train a second model for predicting tumor presence from MRI alone. Predictive maps were created for seven patients left out of the training steps, and tissue samples were tested to determine the model’s accuracy. The final model produced a receiver operator characteristic (ROC) area under the curve (AUC) of 0.70. This study demonstrates a new method for detecting infiltrative tumor beyond conventional radiologist defined margins based on neural networks applied to rad-path datasets in glioblastoma.