Optimization of Decision Making in Personalized Radiation Therapy using Deformable Image Registration

Cancer has become one of the dominant diseases worldwide, especially in western countries, and radiation therapy is one of the primary treatment options for 50% of all patients diagnosed. Radiation therapy involves the radiation delivery and planning based on radiobiological models derived primarily from clinical trials. Since 2015 improvements in information technologies and data storage allowed new models to be created using the large volumes of treatment data already available and correlate the actually delivered treatment with outcomes. The goals of this thesis are to 1) construct models of patient outcomes after receiving radiation therapy using available treatment and patient parameters and 2) provide a method to determine real accumulated radiation dose including the impact of registration uncertainties. In Chapter 2, a model was developed predicting overall survival for patients with hepatocellular carcinoma or liver metastasis receiving radiation therapy. These models show which patients benefit from curative radiation therapy based on liver function, and the survival benefit of increased radiation dose on survival. In Chapter 3, a method was developed to routinely evaluate deformable image registration (DIR) with computer-generated landmark pairs using the scale-invariant feature transform. The method presented in this chapter created landmark sets for comparing lung 4DCT images and provided the same evaluation of DIR as manual landmark sets. In Chapter 4, an investigation was performed on the impact of DIR error on dose accumulation using landmarked 4DCT images as the ground truth. The study demonstrated the relationship between dose gradient, DIR error and dose accumulation error, and presented a method to determine error bars on the dose accumulation process. In Chapter 5, a method was presented to determine quantitatively when to update a treatment plan during the course of a multi-fraction radiation treatment of head and neck cancer. This method investigated the ability to use only the planned dose with deformable image registration to predict dose changes caused by anatomical deformations. This thesis presents the fundamental elements of a decision support system including patient pre-treatment parameters and the actual delivered dose using DIR while considering registration uncertainties