Cancer represents an important barrier to increasing life expectancy, and therefore its treatment and management are key challenges of modern society in the contexts of health and quality of life. Cancer of the head and neck (HaN) region accounts for around 5% of all cancers worldwide, for which radiotherapy (RT) is a widespread treatment modality that delivers a high radiation dose to the cancerous cells while sparing the healthy organs-at-risk (OARs). In the past decade, advances in RT have contributed to the preservation of function and reduced mortality of the HaN cancer patients, among which a considerable progress has been made in artificial intelligence (AI) that has a wide spectrum of applications spanning the field of RT. In particular, the process of partitioning medical images into three-dimensional spatial instances representing tumors and OARs, known as contouring or segmentation, that aids in producing optimal patient-specific radiation dose distribution plans has witnessed a revival with the introduction of automated AI-assisted computerized medical image analysis techniques that perform auto-segmentation. While most segmentation still relies on computed tomography (CT) images as they contain electron density information used for the calculation of the radiation beam energy absorption, the integration of complementary magnetic resonance (MR) images has been recognized as valuable, especially for OAR segmentation in regions such as the HaN. However, from the perspective of RT, OAR segmentation has direct clinical implications because it has to be evaluated not only for the geometric agreement against the ground truth, but also from the perspective of the dosimetric impact. Moreover, with the increasing availability of MR images in RT planning, the radiation oncology workflow has been lately focusing on the MR modality for the purpose of MR-only RT. Since CT images are required by the RT planning software for dose distribution calculations, the main idea of MR-only RT is to replace CT image acquisition by synthetic CT images, generated from given MR images, and then use only MR images in combination with synthetic CT images throughout the whole RT planning workflow. In the proposed research project, we aim to build upon knowledge and experience of our past projects to systematically measure and analyze the dosimetric impact on RT planning in the HaN region caused by different geometric configurations of OAR contours as obtained from CT images, MR images, a combination of CT and MR images, or synthetic CT images by either manual or auto-segmentation, or by manual editing of auto-segmentation contours. For this purpose, we will collect a database of RT planning cases from our clinical partners that will complement and augment our existing database of HaN CT and MR images. With such a database at our disposal, we plan to design and develop a state-of-the art AI framework for generating synthetic CT images given acquired MR images of the HaN region by using generative denoising diffusion models, a recent cutting-edge AI technology that demonstrated a high performance in learning complex data distributions, and evaluate their impact on MR-only RT. Finally, we aim to provide relevant feedback on the obtained outcomes to our clinical partners so as to provide novel applications and insights to clinicians that will, in turn, facilitate the radiation oncology workflow, improve the quality of RT planning, and, ultimately, benefit HaN cancer patients. Our objectives are therefore concentrated around the design and development of AI-supported state-of-the-art computerized medical image analysis techniques for synthetic CT image generation, and systematic evaluation of the geometric and dosimetric impact on RT cancer treatment planning in order to contribute to the development MR-only RT and leverage its application in the HaN region.
