Deep learning for dose reduction and auto-segmentation in CBCT-guided online adaptive radiotherapy

Modern photon radiotherapy enables high precision dose delivery to the tumor while sparing adjacent organs-at-risk (OARs). Yet, in many cases, its potential is not fully exploited since, in current clinical practice, patients are treated with irradiation plans optimized at a single time-point prior to the start of therapy. Frequently occurring alterations of the patient anatomy during the course of fractionated therapy are only considered by the introduction of safety margins around the actual target volume, which inevitably increases the OAR dose burden. A substantially improved treatment becomes feasible by online adaptive radiotherapy (ART), where the dose is optimized prior to each fraction on basis of the daily patient anatomy as inferred from in-room imaging.

Such workflows have recently seen clinical adoption in the context of magnetic resonance imaging (MRI)-guided radiotherapy with combined MRI linear accelerators. However, these devices are still very sparsely available in clinics, while the vast majority of patients are treated using near ubiquitous gantry-mounted cone-beam computed tomography (CBCT) scanners. While the acquired CBCT images can be utilized for bony-anatomy or fiducial marker-based patient alignment, image quality is not sufficient to exploit the images for online treatment adaptation. In the last years, many approaches for improving the image quality and enabling accurate dose calculation for treatment plan adaptation have been discussed in the literature and might see clinical introduction in the near future. To enable CBCT-guided online ART, however, delineation of target and OAR structures on the CBCT images is also crucial and should be performed within few minutes in an online workflow with the patient waiting in treatment position. Besides accurate delineation, imaging dose is a second major concern in the scope of CBCT-guided ART, especially if considering daily pre-treatment imaging for adaptation.

This project aims at leveraging state-of-the-art deep learning techniques for low dose CBCT acquisition, intensity correction and segmentation to pave the way towards CBCT-guided online ART.

Project number 70114849