Trajectory Optimization for VMAT Treatment Aided by Artificial Intelligence

Abstract

Artificial Intelligence (AI) recently beat the top GO players. With the help of AI, here we tried to tackle one of the most difficult optimization problems in radiation oncology: designing an efficient trajectory to deliver an isocentric VMAT treatment with an optimum dose distribution on a C-arm machine, where the gantry and couch move in coordination. The state of art trajectory optimization still requires human expert to design heuristic algorithms such as those based on BEV or BEVD. We discretize the 4pi space into 2701 nodes, where the gantry and couch rotations are separated by 5 degrees. To search through the vast number of possible trajectories, we implemented Monte Carlo Tree Search (MCTS), the same method used by the AI which mastered the game of GO. Each simulated trajectory picks the nodes one by one. The next selected node would firstly need to be physically reachable, constrained by the rotation speed of the gantry and couch, as well as collision. Secondly it would have the maximum Upper Confidence Bounds for Trees (UCT), which is a function of average objective function value, the number of time this node has been selected and the total number of simulations. After forming a full trajectory, we run an inverse fluence map optimization with an infinity norm regularization term and the alternating direction method of multipliers (ADMM), using all the nodes on the trajectory. The obtained objective function value is then fed back to update the statistic of the nodes on the trajectory. The infinity norm regularization forces a solution with uniform fluence maps, close to that of VMAT treatment. The MCTS continues until the solution does not improve anymore or a predetermined time is reached. We tested the MCTS method with two challenging patient cases, one chest wall, and one skull and compared the dosimetry with coplanar plans. Without loss of generality, we confine the couch rotate uniformly from -90° to 90° with a speed of 0.5 degree/second and put no constraint on the gantry movement. The dose calculations are performed on the fly with the aid of GPU. For both patient cases, the AI with MCTS found optimum trajectories within an hour after ∼360 simulations with superior dosimetric qualities for OARs and same PTV coverages compared with coplanar plans. For the skull case, the brain V30, V20, and V10 are reduced by 21%, 38%, and 27%, respectively. The mean dose of the temporal lobe is decreased by 10%. For the chest wall case, the heart mean dose is reduced from 15 Gy to 7 Gy and V30 from 7.5% to 1.7%. The left and right lung mean dose decreased from 16 Gy and 8 Gy to 11 Gy and 3 Gy. The left and right lungs V20 are reduced from 27% and 4.4% to 15% and 0.9%. The estimated delivery time is ∼ 6 min. AI with MCTS successfully designed trajectories that are efficient to deliver and dosimetric superior compared with coplanar VMAT. The method can also be easily generalized to include non-isocentric treatments.