Fast, accurate and trustworthy dose calculation on magnetic resonance images for proton therapy using deep learning

• Applicants: Professor Dr. Armin Lühr; Dr. Liheng Tian
• Subject Area: Medical Physics, Biomedical Technology
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 557923081

Each year, about half a million new patients are diagnosed with cancer in Germany. More than 50% of them receive radiotherapy as an integral part of their treatment. Proton therapy is a highly effective radiation treatment method in which the prescribed dose can be delivered very precisely to the identified tumor volume. To fully exploit the benefits of proton therapy, high-quality imaging is required that allows precise delineation of the tumor volume and critical healthy structures. This is where magnetic resonance imaging (MRI) is extremely attractive as it provides anatomical information with the highest soft tissue contrast. However, state-of-the-art proton dose calculation relies on computed tomography (CT) images and, as a consequence, standard radiotherapy needs to combine both planning CT and MRI. This results in multiple imaging sessions for the patient, increased costs and, in particular, uncertainties in image registration, which lead to dose errors. In addition, a CT scan exposes a large portion of the patient to unwanted ionizing radiation, limiting daily pre-treatment CT scans that reflect the patient's anatomy of the day. Recent studies suggest that MRI-only radiotherapy (i.e. without CT images) can overcome these limitations and enable adaptive online radiotherapy on daily MRI images. These achievements in photon-based radiotherapy (e.g. MR-Linac) have inspired active research and development to make MRI-only proton therapy a reality. Especially, accurate proton dose computation on MRI and the ability to test if it is of sufficient quality are of great importance to the field, but also with respect to patient comfort and would enable different treatment paths for cancer patients. The aim of this project is to a) develop, b) integrate and c) test advanced deep learning (DL) based software solutions for accurate MRI-only proton dose calculation for clinical scenarios. A particular focus is on estimating and dealing with the uncertainty of the calculated dose and on demonstrating clinical applicability. For the development (a), we will establish advanced DL-based MRI-only dose calculation engines and, importantly, novel approaches to estimate the corresponding dose errors. For the integration (b), we will implement the superior DL-based MRI-only proton dose engines into a treatment planning system (TPS) and integrate the uncertainty estimates into scenario-based robust dose optimization. Two clinical applications are considered: treatment planning and recalculation of dose for daily treatment adaptation. For the clinical application test (c), we will establish MRI-only robust dose optimization also in an existing clinical TPS and evaluate by how much the overall dose error can be reduced compared to clinical CT-based proton planning.

• DFG Programme: Research Grants
• Co-Investigators: Dr. Michael Kamp; Professor Dr. Jens Kleesiek