| Radiation pneumonitis (RP) is an acute lung toxicity induced by radiation. On average, 7% of patients who receive thoracic irradiation during radiation therapy (RT) treatment will develop RP, experiencing symptoms including cough, shortness of breath, or fever. RP decreases patient quality of life, may limit further cancer treatment, and in severe cases can result in patient mortality. RP may be treated with corticosteroids, which have been shown to reduce symptoms if administered immediately following RP onset. Thus, it is important to develop methods to aid in early diagnosis of RP, as these techniques would allow for improved patient surveillance and earlier intervention if RP develops.;This dissertation presents an imaging-based method to quantify acute radiation-induced lung injury and identify patients with RP based on changes in individual patient CT scans. The hypothesis was that the change in a set of mathematical descriptions of image texture ("texture features") between CT scans acquired before and after RT could be used to identify the presence of RP. The work began by developing a fully automated method for quantitative analysis of serial CT scans. Specifically, matched anatomic locations of serial scans were identified through deformable image registration, then the change in a set of texture features between these matched regions was measured. When tested using CT scans from 27 healthy patients, this method achieved sub-millimeter accuracy in the placement of matched regions between scans. Twenty texture features that demonstrated stability in the absence of radiation-induced change were identified. These 20 stable texture features were applied to measure radiation-induced lung damage in the CT scans of 25 patients who received RT for lung cancer treatment. For 19 of 20 features, a significant relationship between feature value change and the severity of radiation-induced damage, as identified by a radiologist, was observed. Subsequently, an automated method that used deformable registration to associate planned radiation dose with radiation-induced damage was implemented and evaluated. This method mapped dose between scans with high accuracy (e.g., errors in mapped dose less than 3% of the prescribed radiation dose) for the majority of regions.;Our automated texture analysis method was ultimately applied to characterize dose-dependent feature value change following RT for an independent cohort of 106 esophageal cancer patients, 19% of whom developed RP. For all 20 previously identified features, feature value change in the post-RT CT scan changed significantly with increasing radiation dose. Furthermore, 12 of the 20 features correlated with the presence of RP. A logistic regression-based classifier achieved moderate levels of performance (mean area under the ROC curve (AUC) of 0.71) when classifying RP status based on a linear combination of feature value changes, demonstrating the feasibility of using individualized, quantitative texture analysis to provide information about normal lung tissue toxicity following RT. This dissertation represents a first step in the development of methodology aimed towards aiding physicians in early RP diagnosis, which could subsequently improve the effectiveness of treatment. |
