Learning-based computational strategies for treatment adaptation in image-guided radiotherapy.
Radiotherapy (RT) is one of the main treatment modalities for cancer patients. Increasing the delivered dose to the tumor improves the probability of local tumor control, but also increases the risk of radiation-induced damage to the surrounding tissues. Treatment delivery in RT is based on a treatment plan that is derived from a pre-treatment CT scan of the patient through a complex, time-consuming process. During the course of treatment, significant anatomical changes may occur such that the original treatment plan may be invalidated. Such anatomical uncertainties are currently accounted for by extending the target volume with a geometric margin, but this inevitably leads to an increase of the dose to the healthy tissues. Hence, there is a strong clinical interest for developing effective strategies for optimal treatment plan adaptation in response to anatomical changes over time, based on images acquired immediately prior to each treatment session. A major challenge for such adaptation is that is has to be performed on-line, i.e. with the patient on the treatment table, which imposes strong time constraints on any adaptation protocol. In this project, novel strategies for treatment adaptation in RT will be developed by making use of machine learning to exploit the rich information that is provided by previous, similar cases, aiming at predicting the required adaptation directly from the images in a statistical way without need for a full conventional replanning.