UCSF

Radiotherapy is one of the most common techniques used to treat cancer and is administered to over 60% of patients treated in the US. Computed tomography (CT) and magnetic resonance imaging (MRI) are powerful imaging modalities which have widespread applications in radiotherapy. Given the different underlying physics behind how image contrast is generated in kilovoltage CTs (kVCTs), megavoltage CTs (MVCTs) and MRIs, each has their own inherent strengths and limitations. This dissertation includes four primary projects which use advanced image processing and deep learning methods to exploit the unique advantages of MRI, kVCT, and MVCT for improving dose calculation accuracy of photon and proton radiotherapy. The first project includes a novel multimodal imaging method used to estimate proton stopping power ratio using a combination of MRI and CT. Results show that our multimodal imaging approach using MRI and MVCT provided results within 1% of physical measurements. The second project includes the first work to synthesize MVCT images from MRI using a deep learning model in which the feasibility of utilizing MRI-derived synthetic CTs for radiotherapy treatment planning of head and neck cancer was demonstrated. This work lays the foundation for a proposed paradigm shift whereby MRI is utilized for anatomical delineation while MRI-derived synthetic MVCT is used for radiotherapy dose calculation, which can all be performed on the treatment machine and could provide substantial improvements to combined MRI-linear accelerator workflow and accessibility. The third project includes the first known application of deep learning to reduce uncertainty in the relationship between kVCT and electron density and stopping power ratio through learning of MVCT data, promising improvements for dose calculation accuracy. The fourth project includes clinical considerations and recommendations for implementing a commercial algorithm for producing MRI-derived synthetic CTs in a dataset of patients treated for prostate cancer.