Date of Graduation
Doctor of Philosophy (PhD)
Mary K. Martel, Ph.D.
Tina Marie Briere, Ph.D.
Laurence E. Court, Ph.D.
Arvind Rao, Ph.D.
Francesco Stingo, Ph.D.
Improving outcomes for non-small-cell lung cancer patients treated with radiation therapy (RT) requires optimizing the balance between local tumor control and risk of normal tissue toxicity. In approximately 20% of patients, severe acute symptomatic lung toxicity, termed radiation pneumonitis (RP), still occurs. Identifying the individuals at risk of RP prior to or early during treatment offers tremendous potential to improve RT by providing the physician with information to assist in making clinical decisions that enhance therapy. Our central goal for this work was to demonstrate the potential gain in predictive accuracy of normal tissue complication probability models for RP by considering CT-based image features extracted from the normal lung volume.
To accomplish this, a software framework was first built to facilitate CT image feature extraction using multiple image analysis methods. Subsequently, we applied the implemented methods towards understanding the temporal change in the normal lung volume during treatment. After identifying a subset of highly reproducible and non-redundant image features, we investigated change in lung features on weekly CT image sets acquired during treatment. While multiple features exhibited significant association with dose, no temporal response was identified and we were unable to produce a predictive model that could outperform simple treatment-related factors.
CT-based image features calculated in regional subvolumes and on a voxel-wise basis in the normal lung were explored in the context of RP incidence. There was no clear spatial variation in the considered regionally extracted features or voxel-based feature maps. However, a limited subset of features were significantly associated with RP which may be a useful finding to consider in development of predictive models to assess toxicity risk.
We also considered the utility of pre-treatment total normal lung CT features for predicting RP using LASSO logistic regression and were able to successfully demonstrate improved discrimination of RP using such features relative to models constructed with clinical and dosimetric variables only. This is a significant step towards building robust models of RP with image based features that can subsequently be used to achieve personalized RT.
Radiation pneumonitis, radiomics, texture analysis, computed tomography, lung, predictive modeling