Automated quality assurance of deep learning contours in head-and-neck radiotherapy

During radiotherapy, the anatomy of the body changes and this demands for regular modification of the boundaries (or contours) of tumors and organs on 3D medical scans. Automated algorithms (generally called as auto-contouring models) exist, however they make mistakes. Thus, there is a need for quality assessment techniques that enable fast and easy refinement of contours that enable precise radiotherapy and eventually improve quality of life for cancer patients.This work explores techniques like uncertainty, radiotherapy dose and contour relation, as well as interactive auto-contour refinement. A successful technique was proposed that aligns deep learning (i.e. artificial intelligence) uncertainty better with auto-contouring errors, thus allowing uncertainty to be used a proxy for errors. Also, an AI contour refinement model was built and tested with clinicians on a web interface showing substantial time improvements for auto-contour quality assessment. Finally, a retrospective study investigated the relationship between radiotherapy dose and contouring errors, potentially offering suggestions for contour sections that greatly influence dose and thus speeding up contour refinement.As the world adopts more AI techniques, there is a strong need to explore human-centric AI approaches that can take the best of both parties: human context and AI efficiency. This thesis is an effort to apply this to head and neck radiotherapy.