Speaker
Description
At the UMCG radiotherapy department, cancer patients are treated with targeted radiation. To plan treatment, clinicians outline tumours and nearby healthy organs on CT scans. This step is crucial but time-consuming. Therefore, deep learning segmentation models are now used in the clinical workflow. However, their outputs still require human oversight, as accuracy varies between cases. The resulting time-gain of automated segmentation is therefore limited. To address this, we use the Hábrók computer cluster to explore two directions: (1) uncertainty quantification, enabling models to indicate their confidence and better guide human oversight, and (2) modelling the impact of local segmentation errors, since not all errors equally affect the final treatment plan and some may have lower priority for correction.
