The University of Texas MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences Dissertations and Theses (Open Access)
Detecting and Evaluating Therapy Induced Changes in Radiomics Features Measured from Non-Small Cell Lung Cancer to Predict Patient Outcomes
Author ORCID Identifier
Date of Graduation
Doctor of Philosophy (PhD)
Aaron Kyle Jones
The purpose of this study was to investigate whether radiomics features measured from weekly 4-dimensional computed tomography (4DCT) images of non-small cell lung cancers (NSCLC) change during treatment and if those changes are prognostic for patient outcomes or dependent on treatment modality. Radiomics features are quantitative metrics designed to evaluate tumor heterogeneity from routine medical imaging. Features that are prognostic for patient outcome could be used to monitor tumor response and identify high-risk patients for adaptive treatment. This would be especially valuable for NSCLC due to the high prevalence and mortality of this disease.
A novel process was designed to select feature-specific image preprocessing and remove features that were not robust to differences in CT model or tumor volumes. These features were then measured from weekly 4DCT images. These features were evaluated to determine at which point in treatment they first begin changing if those changes were different for patients treated with protons versus photons. A subset of features demonstrated significant changes by the second or third week of treatment, however changes were never significantly different between patient groups. Delta-radiomics features were defined as relative net changes, linear regression slopes, and end of treatment feature values. Features were then evaluated in univariate and multivariate models for overall survival, distant metastases, and local-regional recurrence. In general, the delta-radiomics features were not more prognostic than models built using clinical factors or features at pre-treatment. However one shape descriptor measured at pre-treatment significantly improved model fit and performance for overall survival and distant metastases. Additionally for local-regional recurrence, the only significant covariate was texture strength measured at the end of treatment. A separate study characterized radiomics feature variability in cone-beam CT images to increased scatter, increased motion, and different scanners. Features were affected by all three parameters and specifically by motion amplitudes greater than 1 cm.
This study resulted in strong evidence that a set of robust radiomics features change significantly during treatment. While these changes were not prognostic or dependent on treatment modality, future studies may benefit from the methodologies described here to explore delta-radiomics in alternative tumor sites or imaging modalities.
Radiomics, Non-Small Cell Lung Cancer, Quantitative Imaging, Texture, Outcomes, Modeling
Medical Biomathematics and Biometrics Commons, Multivariate Analysis Commons, Other Physics Commons