Uncertainty quantification in automatic delineation for radiation therapy

Recent advances in automatic segmentation using deep learning have led to the emergence of solutions for the delineation of structures in radiotherapy (organs at risk and tumors). These solutions help doctors by providing automatic contouring which can be used as a starting point for delineation. They also enable real-time adaptation of treatment by integrating automatic delineation from the patient’s positioning image during treatment, leading to truly adaptive radiotherapy, which is one of the most promising evolutions
of radiotherapy. At the end of 2023, ICANS acquired a first commercial solution (ART-Plan Annotate from TheraPanacea) allowing the automatic delineation of organs and tumors, aiming to improve the reproducibility of contours and to help doctors. In 2024, the ICANS technical platform will also see the arrival of the latest generation of radiotherapy system.
This system includes another commercial solution for auto-contouring, that allows direct online adaptive radiotherapy. In this context, it is crucial for doctors and physicists to know the expected performance of different tools to ensure the validity of delineation and the safety of the treated patients. Furthermore, from a practical standpoint, the ability to predict tool malfunctions would enable the implementation of proactive measures. The objective of the thesis is therefore threefold:
•determine the level of difficulty of automatic segmentation for different tumor locations and different organs at risk.
•determine classes of images or structures for which the algorithm fails to create a segmentation or produces an aberrant segmentation.
•determine the uncertainty associated with the segmentation.

More informations on the link below.