Date of Graduation
Doctor of Philosophy (PhD)
The purpose of this work was to determine if quantitative image features (QIFs) extracted from computed tomography (CT) and flourodeoxyglucose (FDG) positron emission tomography (PET) could provide prognostic information to improve outcome models. Our goal for this work was to determine if it may one day be feasible to incorporate QIFs into personalized cancer care. QIFs were used to quantitatively characterize patient disease as seen on imaging. A leave-one-out cross-validation procedure was used to assess the prognostic ability of QIFs extracted from CT and PET in addition to conventional prognostic factors (CPFs). QIFs were found to improve model fit for overall survival in contrast enhanced CT (CE-CT) (p = 0.027) and FDG-PET (p = 0.007). Correlations/associations were observed between QIFs from CE-CT, FDG-PET, and CPFs. However, our results indicate that while correlations/associations exist, QIFs provided additional prognostic information. QIFs from FDG-PET improved models using CPFs including GTV in terms of patient stratification, c-index, and log-likelihood more than QIFs from CE-CT alone. Various studies were performed assessing the reproducibility of FDG-PET based QIFs and found that reconstruction methods certainly impact the obtained QIF values. However, features maintain a reasonable reproducibility (mean CCC = 0.78) that may be improved when using similar reconstructions (e.g., 3D OSEM) (CCC = 0.93). The two FDG-PET features found to be prognostic were also able to isolate sub-cohorts of patients that demonstrated survival differences based on radiation dose.
QIFs were found to provide additional prognostic information beyond that found from CPFs. Initial evidence suggests that the examined FDG-PET based QIFs may have utility across cohorts and could potentially determine which patients may benefit from dose escalation.
Non-small cell lung cancer, quantitative image features, prognostic value, outcomes, modeling, texture