Off-campus WSU users: To download campus access dissertations, please use the following link to log into our proxy server with your WSU access ID and password, then click the "Off-campus Download" button below.
Non-WSU users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Date of Award
Individualized medicine is possible due to the discovery that a patient’s gene expression can serve as a bio-marker and provide prognostic clinical knowledge. The genetic revolution increased interest in the field of “-omics” which investigates differing classifications of patient characteristics to increase clinician knowledge, which directly improves patient care. Glioblastoma (GBM) is the most aggressive brain cancer and very difficult to treat, resulting in a 15 month median survival. Researchers have discovered several patient specific markers, specifically IDH1 and MGMT methylation, which energized further research for GBM patient specific markers. A relatively new subfield of ‘-omics’ is Radiomics; which aims to glean prognostic information by quantitatively assessing a patient’s medical images and characterize tumor presentation. Radiomics research is growing in popularity due to the potential simplicity in clinical implementation. However, researchers must address the plethora of workflow heterogeneities as they are currently limiting robustness of radiomic features and therefore research lacks reproducibility. The goal of our work is to overcome these difficulties and create a reproducible and robust RSS derived from an automated workflow that can be deployed in the clinical settings.Our multi-institutional study employed a total of 445 patients’ diagnostic T2- weighted, T1-weighted, T1 Contrast enhanced, and FLAIR magnetic resonance images (MRI). Segmentation heterogeneity was initially addressed to decrease the intra/inter clinician differences. This was done by training a 2D U-Net to contour GBM sub-regions (Tumor Core (TC), Enhancing Tumor (ET), and Whole Tumor (WT)). To increase the U-Net performance, the training dataset was boosted by synthetic MRI created by a Generative Adversarial Network. The training cohort - obtained from BraTS 8,25,26 - is comprised 213 patients from nine different institutions. Alongside patient MR images BraTS also supplied physician generated GBM sub-region segmentations. The independent testing cohort is comprised of 231 patients, and segmented by our U-Net. Diminishing effects of the remaining heterogeneities was addressed by assessing the abilities of harmonization techniques both individually and in combinations. MRI processing techniques investigated were N4ITK Bias Correction, Z-Score and White Stripe. Feature extraction was performed by the Cancer Imaging Phenomics Toolkit (CaPTK). Radiomic feature harmonizations investigated were COMBAT, L1, L2, and Quantile normalization. Determination of which GBM sub-region resulted in the most robust RSS was done by analyzing each region independently and in combination. 246 individual datasets were created by application of differing combinations of these harmonizations. Changes in feature values for each of these datasets was assessed through the Intraclass Correlation Coefficient (ICC). To determine the best performing combination of harmonizations, a Radiomic Signature Sets (RSS) was generated for each one. Robustness of each of the 246 RSS was assessed through ability to maintain prognostic ability. Log-Rank p-value was used for statistical analysis. Robustness and prognostic ability was assessed by the receiver operating characteristic area under the curve metric. A trained support vector machine reported results for 5-fold cross validation on the training cohort and ability to correctly identify OS for the independent testing cohort. Benchmarking prognostic ability of the best performing RSS was done by direct comparison against MGMT methylation status, Age, and Karnfosky Performance Score (KPS) for the subset of patients who reported this additional information. The best performing RSS was generated when N4 and Z-Score normalizations were applied to the MRI, whole tumor contour is used, and features are harmonized through COMBAT followed by Quantile normalization. This RSS is composed of five features and shows statistical significance through the log-rank test, and ROC-AUC for both training and testing cohorts was 89%. It should be noted that the similarity in abilities between the cohorts also validates use of U-Net generated segmentation in the Radiomics workflow. It also shows high individual prognostic ability. Additionally, model performance using the known prognostic factors MGMT methylation, Age, and KPS is improved when the RSS is included. Generated RSS shows robustness through statistical and modeling results; and benchmark analysis shows ability to aid clinical markers. Additionally, the workflow can be fully automated due to acceptable U-Net performance. Combination of the ease of implementation of a fully automated workflow with high prognostic power of generated RSS indicates its potential clinical utilization.
Carver, Eric Nathan, "Automated, Robust Radiomics Biomarker Modeling For Glioblastoma" (2022). Wayne State University Dissertations. 3621.