Gary GeFollow

Author ORCID Identifier

Date Available


Year of Publication


Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation




Radiation Science

First Advisor

Dr. Jie Zhang


Radiomics is a technique that extracts quantitative features, termed radiomic features, from medical images using data-characterization algorithms. These radiomic features can be used to identify tissue characteristics and radiologic phenotyping that are not observable by clinicians in a non-invasive, low-cost manner, potentially generating image biomarkers for clinical decision. To date, there are still many uncertainties involved in radiomics which limit its clinical implementation. Herein, we propose to explore the impact of each component in the radiomics pipeline on predicting clinical outcomes. In Chapter II, we conduct a thorough review of CT lung cancer radiomics studies to examine the typical feature selection methods and predictive models for radiomics, outlining current practices and identifying common factors that may negatively impact study outcomes. In Chapter III, we investigate radiomic feature uniqueness, a novel approach to evaluate the relationship between radiomic features and underlying tissue structure as a determinant of feature relevance. The results show approximately 10% of the examined radiomic features are not unique to a defined ROI for CT NSCLC. Features that are not unique should be removed prior to conducting radiomic analysis, reducing feature dimensionality in radiomic studies. In Chapter IV, we perform a multi-faceted examination of feature selection methods, predictive models and related factors including cohort size, cohort composition, number of input features, and training/validation methods. The results reveal the impact of these factors on radiomic model performance. In Chapter V, we investigate the impact of clinical features on predicting clinical outcome. The results show that clinical features or their combination with radiomic features can improve predictive performance over radiomic features alone. The uncertainties involved in CT NSCLC radiomics are likely to be present in other radiomic studies where the same or similar machine learning methods are applied. Our study provides a better understanding of the underlying factors in radiomics, which will help establish the foundation for clinical implementation of radiomics.

Digital Object Identifier (DOI)