H. Bagher-Ebadian1,2, S. L. Brown1, P. Acharya3, J. R. Ewing4, I. J. Chetty5, B. Movsas1, and K. Thind1; 1Department of Radiation Oncology, Henry Ford Health, Detroit, MI, 2Department of Radiology, Michigan State University, East Lansing, MI, 3Oakland University, Rochester, MI, 4Department of Neurology, Henry Ford Health, Detroit, MI, 5Department of Radiation Oncology, Cedars-Sinai Medical Center, Los Angeles, CA
Purpose/Objective(s): Dynamic Contrast Enhanced (DCE) MRI information has shown promise as a surrogate endpoint for assessing tumor response to radiation therapy (RT) because the functional pharmacokinetic (PK) alterations occur earlier than morphological changes. This study proposes a probabilistic unsupervised mathematical model constructed from DCE-MRI information to predict the probability of RT-induced PK changes in a rat brain tumor model of human cancer. The changes in DCE-MRI collected within hours of RT have the potential to predict long term tumor response. Materials/
Methods: Twenty-four immune compromised RNU rats were implanted with human U-251N cancer cells to form an orthotopic glioma studied 28 days after implantation when tumors were 3 to 4 mm in diameter. For each rat, two DCE-MRI studies (Dual Gradient Echo) were performed 24h apart using a 7T MRI scanner. A single 20Gy stereotactic conformal radiation exposure, equal to the dose to cure half of a group of tumors, was performed before the second MRI (acquired 1-6.5 hours post RT). DCE-MRI PK analysis was performed using a probabilistic nested model selection (PNMS) technique to distinguish three different pathophysiological states of brain regions. The longitudinal relaxation time change (?R1) was calculated for each brain voxel, normalized to a fixed timespan (5min), and used to build three RT-based Kohonen Self Organizing Maps (KSOM, topology:5X5, competitive learning using a “best matching unit” strategy). The PNMS results were used to perform model averaging of the estimations provided by the RT-KSOMs. The trained KSOMs were applied to the pretreatment RT DCE-MRI information to predict the RT-induced alteration probability maps. Average values of the RT induced probability map for different model regions (PNMS maps at a 50% threshold) were calculated. Results: Non-leaky and highly permeable tissues showed less RT-induced effect (mean probabilities were: 0.51±0.05 and 0.57±0.04, respectively) compared to the peritumoral regions (mean probability: 0.66±0.04) pertaining to leaky tissues with no back flux to the vasculature compartment. Conclusion: This pilot study introduces an unsupervised predictive model combined with a PNMS technique to effectively capture subtle RT-induced alterations in the PK response of brain tissues according to their similarities/dissimilarities. Results of this study strongly agree with our previous findings, indicating that the contrast-enhanced rim of the tumors - the peritumoral zones correspond to infiltrative tumor borders that are preferentially affected by RT. To improve the accuracy of the prediction, an uncertainty analysis is warranted that is designed to estimate the intrinsic variations between the pre and post RT DCE-MRI signal and their influence on the KSOM feature space along with physiological-based signal calibration of the RT-effect.