Segmentacija večmodalnih medicinskih slik z uporabo globokega učenja

Cancer remains one of the major socioeconomic challenges of modern times. It is among the three leading causes of death globally, accounting for an estimated 9.7 million deaths annually. Over the years, various treatment modalities have been developed, with surgery, chemotherapy, and RT being the most widely adopted. While each treatment modality has its limitations, RT offers several advantages: unlike surgery, which is invasive and applicable to only a subset of tumor types, and chemotherapy, which often causes systemic side effects, RT is a localized or regional treatment that can better preserve anatomical structures. The objective of RT is to deliver a high dose of ionizing radiation to the target volumes (tumor and associated lymph nodes) while sparing nearby healthy structures critical to bodily function, known as organs-at-risk (OARs). A standard RT workflow begins with acquiring a planning CT scan, followed by manual delineation of OARs and target volumes. These delineations are then used in a treatment planning process that optimizes beam configurations within the linear accelerator (LINAC), ensuring that the prescribed dose is delivered to target volumes while satisfying the dosimetric constraints for the surrounding OARs. Manual contouring is laborious, time-consuming, and subject to both intra- and inter-observer variability. Automating the OAR delineation step has the potential to significantly streamline the RT planning process, reduce variability, and enable a broader adoption of adaptive RT, an approach often hindered by current time and resource constraints. However, given the high clinical stakes, auto-segmentation methods need to achieve a high level of accuracy. This is particularly challenging in anatomically complex regions such as the head and neck, where CT images often lack sufficient soft-tissue contrast. To address this issue, clinicians routinely supplement CT with MR imaging to support manual delineation. Similarly, leveraging complementary information from both modalities in the design and development of automatic methods may enhance segmentation quality, especially for structures that are poorly defined on CT alone. Motivated by the aforementioned issues and challenges, the primary focus of this thesis is to investigate the hypothesis that multi-modal image segmentation can improve the quality of OAR segmentation for RT planning. We approach this problem comprehensively: first, by constructing a paired CT-MR dataset with expert annotations