PROJECT SUMMARY Pancreatic cancer has the highest mortality rate of all cancers, with a 5-year survival rate of only 9%. Surgery still represents the only curative treatment option, though less than 20% of patients are candidates for resection. Approximately 30-40% of patients present with locally advanced unresectable tumors with no significant chance of long-term survival through standard treatments. The use of ablative radiation doses (biologically equivalent doses of 100Gy) produces results that are comparable to surgical resection in patients with inferior prognostic features. However, organ motion, due to respiratory motion, must be managed to minimize toxicity in the gastrointestinal tract. In this project, we will develop novel real-time volumetric MRI technology that can guide radiotherapy to enable the use of ablative doses with minimal risk. Our technique, called MR SIGnature Matching (MRSIGMA), pre-learns 3D motion states and assigns unique motion signatures during an offline learning phase and performs fast signature acquisition and matching during an online matching phase. We have demonstrated real-time tracking of liver tumors with an imaging latency (acquisition plus reconstruction) of about 250 ms using MRSIGMA. We will collaborate with Elekta to implement MRSIGMA on the Unity MR-Linac system and to link the output of MRSIGMA with the multileaf collimator (MLC) system to enable the radiation beam to track the 3D position and shape of the moving tumor in real-time. Specific Aims are as follows: 1. Develop deep learning reconstruction of undersampled dynamic MRI data for rapid motion database generation during offline learning and adaptation during online matching a. Develop a convolutional neural network for rapid reconstruction of motion-resolved data (< 10 seconds) b. Detect anatomical changes, such as motion baseline drifts, and adapt the motion database accordingly c. Perform initial validation on a dynamic MRI phantom and ten volunteers 2. Validate the potential of MRSIGMA for real-time volumetric tumor motion imaging on fifty patients with locally advanced unresectable pancreatic cancer a. Accuracy hypothesis: real-time MRSIGMA is noninferior to a non-real-time XDGRASP reference b. Reproducibility hypothesis: two MRSIGMA scans present equivalent real-time imaging performance 3. Develop and validate on dynamic phantoms the proposed MRSIGMA-guided MLC tracking in collaboration with Elekta a. Develop software to control the MLC with the output of MRSIGMA b. Evaluate tracking latency, geometric error, reproducibility and dosimetric accuracy

PROJECT NARRATIVE This project will develop novel real-time 3D MRI technology, using continuous radial sampling, motion-resolved reconstruction, and deep learning, for MRI-guided adaptive radiotherapy of unresectable pancreatic cancer. Our new approach for real-time 3D MRI, named MR signature matching (MRSIGMA), pre-learns 3D motion states and corresponding signatures during an offline learning phase and then performs fast signature acquisition and matching to one of the pre-learned 3D motion states to minimize latency and offer a real-time solution. The novel real-time MRI technology will unlock the potential of the MR-Linac for pancreatic cancer and enable the use of curative doses rather than palliative doses, with minimal risk in patients who currently do not have a chance for long-term survival.