Sliced Online Model Checking for Optimizing the Beam Scheduling Problem in Robotic Radiation Therapy

Abstract

In robotic radiation therapy, high-energy photon beams from different directions are directed at a target within the patient. Target motion can be tracked by robotic ultrasound and then compensated by synchronous beam motion. However, moving the beams may result in beams passing through the ultrasound transducer or the robot carrying it. While this can be avoided by pausing the beam delivery, the treatment time would increase. Typically, the beams are delivered in an order which minimizes the robot motion and thereby the overall treatment time. However, this order can be changed, i.e., instead of pausing beams, other feasible beam could be delivered. We address this problem of dynamically ordering the beams by applying a model checking paradigm to select feasible beams. Since breathing patterns are complex and change rapidly, any offline model would be too imprecise. Thus, model checking must be conducted online, predicting the patient's current breathing pattern for a short amount of time and checking which beams can be delivered safely. Monitoring the treatment delivery online provides the option to reschedule beams dynamically in order to avoid pausing and hence to reduce treatment time. While human breathing patterns are complex and may change rapidly, we need a model which can be verified quickly and use approximation by a superposition of sine curves. Further, we simplify the 3D breathing motion into separate 1D models. We compensate the simplification by adding noise inside the model itself. In turn, we synchronize between the multiple models representing the different spatial directions, the treatment simulation, and corresponding verification queries. Our preliminary results show a 16.02 % to 37.21 % mean improvement on the idle time compared to a static beam schedule, depending on an additional safety margin. Note that an additional safety margin around the ultrasound robot can decrease idle times but also compromises plan quality by limiting the range of available beam directions. In contrast, the approach using online model checking maintains the plan quality. Further, we compare to a naive machine learning approach that does not achieve its goals while being harder to reason about.