Multivariate Statistical Techniques for Accurately and Noninvasively Localizing Tumors Subject to Respiration - Induced Motion

Title of Document: MULTIVARIATE STATISTICAL TECHNIQUES FOR ACCURATELY AND NONINVASIVELY LOCALIZING TUMORS SUBJECT TO RESPIRATIONINDUCED MOTION Kathleen T. Malinowski, Doctor of Philosophy, 2012 Directed By: Warren D. D’Souza, Associate Professor, Fischell Department of Bioengineering and Director, Department of Medical Physics Tumors in the lung, liver, and pancreas can move considerably with normal respiration. The tumor motion extent, path, and baseline position change over time. This creates a complex “moving target” for external beam radiation and is a major obstacle to treating cancer. Real-time tumor motion compensation systems have emerged, but device performance is limited by tumor localization accuracy. Direct tumor tracking is not feasible for these tumors, but tumor displacement can be predicted from surrogate measurements of respiration. In this dissertation, we have developed a series of multivariate statistical techniques for reliably and accurately localizing tumors from respiratory surrogate markers affixed to the torso surface. Our studies utilized radiographic tumor localizations measured concurrently with optically tracked respiratory surrogates during 176 lung, liver, and pancreas radiation treatment and dynamic MR imaging sessions. We identified measurement precision, tumor-surrogate correlation, training data selection, inter-patient variations, and algorithm design as factors impacting localization accuracy. Training data timing was particularly important, as tumor localization errors increased over time in 63% of 30-min treatments. This was a result of the changing relationship between surrogate signals and tumor motion. To account for these changes, we developed a method for detecting and correcting large localization errors. By monitoring the surrogate-to-surrogate and surrogate-to-model relationships, tumor localization errors exceeding 3 mm could be detected at a sensitivity of 95%. The method that we have proposed and validated in this dissertation leads to 69% fewer treatment interruptions than conventional respiratory surrogate model monitoring techniques. Finally, we extended respiratory surrogate-based tumor motion prediction to the otherwise timeconsuming process of contouring respiratory-correlated computed tomography scans. This dissertation clarifies the scope and significance of problems underlying existing surrogate-based tumor localization models. Furthermore, it presents novel solutions that make it possible to improve radiation delivery to tumors without increasing the time required to plan and deliver radiation treatments. MULTIVARIATE STATISTICAL TECHNIQUES FOR ACCURATELY AND NONINVASIVELY LOCALIZING TUMORS SUBJECT TO RESPIRATION-INDUCED MOTION