Imaging is crucial for the preparation of radiotherapy. A patient that is going to be irradiated undergoes multiple imaging sessions to prepare the treatment and during the treatment. These images are processed, interpreted and manipulated by various users such as radiation oncologists, radiographers, etc. To facilitate image processing and, ideally, automatize the workflow, we are developing deep learning-based solutions aiming at reduce human efforts and support tailoring the treatment for each patient.
Synthetic-CT generation Computed tomography (CT) is the reference modality that enables planning of the irradiaiton to target the tumour and spare the surrounding healthy tissues. Along with CT, magnetic resonance imaging (MRI) facilitates with its superior soft tissue contrast the localization and delineation of tumor and other clinially relevant structures. In this theme we are working on using convolutional neural networs to derive synthetic-CT from MRI thereby eliminating CT leading to an MRI-only simulation workflow. In this way, the patient does not have to undergo a CT exam anymore and the problematic registration between CT and MRI can be eliminated. Currently, our work has helped to introduce MRI only planning in our clinic for prostate and rectum cancer patients validating commercial solutions. In addition, we are developing synthetic-CTs for more challenging sites such as pediatric, brain and lung radiotherapy patients.
Automatic contouring and registration Deep learning can streamline image post-processing. Currently, image registration and contouring can require considerable manual interaction, e.g. a radiographers spend on average 1 hour per patient on image processing. In total, this accumulates to a significant number of yearly men hours where personnel shortage is severe. We have developed in collaboration with computer scientists and medical physicists of the radiotherapy department several deep learning solutions that are now used clinically for automatic delineation of organs-at-risks for pelvis, lung, esophagus, abdomen, thorax, head and neck and pediatric patients both MRI and CT-based. Solutions for other anatomical sites and imaging modality are under development.
Adaptive cone-beam guided radiotherapy Cone-beam CT (CBCT) is being employed on the linear accelerators to guide patient positioning prior to irradiation. Currently, the image quality of CBCT is sub-optimal due to various hardware and reconstruction issues. The goal of this research theme is to use deep learning to improve the quality of CBCT bringing it on-par with standard CT. This would enable new adaptive treatment schemes, where a treatment plan can be devised based on the daily patient anatomy minimizing target margins and dose to healthy neighboring tissue.
Adaptive MR-guided radiotherapy The introduction of online adaptive MRI-guided radiotherapy using the Elekta Unity MRI-Linac system pioneered at the UMC Utrecht, sets the time constraints for treatment adaptation even more stringent. While the patient is lying on the treatment contouring of targets and organs-at-risks should be performed in a minimum amount of time. The development of fast deep Learning solutions for auto-contouring, registration, image reconstruction, motion estimation and contour-propagation is essential to facilitate shorter treatment fractions.
Education on Deep learning in Radiotherapy Since 2019 we are organizing yearly a popular workshop on deep learning in radiotherapy focusing on principles and applications of deep learning in radiotherapy. Check out www.DLinRT.org for more informations.