Date of Graduation


Document Type

Dissertation (PhD)

Program Affiliation

Medical Physics

Degree Name

Doctor of Philosophy (PhD)

Advisor/Committee Chair

Edward F. Jackson, Ph.D.

Committee Member

Lei Dong, Ph.D.

Committee Member

Valen E. Johnson, Ph.D.

Committee Member

Anita Mahajan, M.D.

Committee Member

R. Jason Stafford, Ph.D.


Quantitative imaging biomarkers (QIBs) are increasingly being incorporated into early phase clinical trials as a means of non-invasively assessing the spatially heterogeneous treatment response to anticancer therapies, particularly as indicators for early response. MR QIBs are derived from the analysis of in vivo imaging data, such as that acquired via dynamic contrast enhanced (DCE), dynamic susceptibility enhanced (DSC), and diffusion tensor imaging (DTI). To date, preclinical and clinical applications of such QIBs have provided strong evidence for potential efficacy, but efforts to create meaningful estimates of localized treatment response using multiple QIBs have been stifled by the need for rigorous characterization of biases and variances inherent in MR equipment and analysis tools and a suitable means of associating QIB changes with treatment response. This research sought to develop such a framework, incorporating multiple MRI QIBs associated with the microvascular environment, e.g., permeability, flow, and volume, and the cellular environment, e.g., water diffusion, into a single classification model to generate maps of predicted locoregional response. To ensure treatment associated changes measured in vivo exceeded equipment related levels of bias and variance, two phantoms were developed. Weekly assessment of the MR imaging data from which the QIBs were derived resulted in coefficients of variation less than 15% for QIBs assessed, well below the expected treatment related changes (approximately 40%). Bias and variance associated with the software tools developed to facilitate longitudinal assessments of treatment response, QUATTRO, was also assessed using synthesized imaging data mimicking clinically relevant acquisitions schema, and found to introduce negligible levels of bias and variance. Finally, to develop an integrated approach to assessing response using multiple QIBs, two experienced radiation oncologists contoured regions of partial response (PR), stable disease (SD), and progressive disease (PD) on rigidly co-registered high grade brain tumor patient data sets, which included DCE, DSC, and DTI acquisitions. Response matched voxel-by-voxel QIBs were trained using an ordinal regression classifier. Using leave-one-out cross-validation, the prediction accuracies of the best model (single DTI QIB) were found to be, mean (standard error), 69.0 (11.1)% for SD, 35.2 (11.7)% for PD, and 52.3(9.7)% overall. In summary, this work resulted in the development of a comprehensive framework for predicting voxelwise radiological treatment response, including the development of phantoms and associated acquisitions for MR equipment quality control and establishment of system-related bias and variance, and a comprehensive software package for performing related image analyses and outcome prediction.


Glioblastoma, ordinal regression, software, quantitative imaging