Program Information
Quantum Leap in Reinforcement Learning for Adaptive Radiotherapy of Liver SBRT
I El Naqa*, R Ten Haken , University of Michigan, Ann Arbor, MI
Presentations
SU-E-FS4-1 (Sunday, July 30, 2017) 1:00 PM - 1:55 PM Room: Four Seasons 4
Purpose: Rapid growth of Big data in radiotherapy invites new opportunities to deliver on the promise of personalized clinical decision support systems. However, classical statistical learning methods lack in efficiency and ability to predict outcomes under conditions of uncertainty and incomplete information. Therefore, we are investigating new reinforcement learning (RL) approaches inspired by quantum physics principles for decision-making and adaptation of SBRT regimens in liver cancer treatment.
Methods: We evaluated a cohort of 88 liver SBRT patients with 35 on non-adaptive and 53 on adaptive protocols. Adaptation was based on liver function in a split course of 3+2 fractions with a month break. RL is composed of an engine that makes decisions based on dosimetric and biological characteristics of the tumor and normal tissue. RL reward is defined by the complication-free tumor control (P+=TCP × (1-NTCP)), where TCP is represented by a Poisson model and NTCP by an LKB model of ALBI changes. Actions were represented by a single bit in classical-RL (cRL) and a qubit in quantum-RL (qRL). The optimal action is found using a Q-learning in cRL and a modified Grover's search algorithm in qRL, which offers quadratic speed-up by using uniform superposition over the possible states and properly adjusting the phases using unitary operations.
Results: The TCP model estimated failure increased with larger volumes while the NTCP model risk increased with higher expression of TGF-β1. Classical/quantum RLs favored adaptation (probability ≥0.5) 79% of the time to adapt for the split-courses and 100% for the continuous-courses. However, the cRL yielded lower adaptation probability of 0.59±0.16 compared to the qRL’s 0.76±0.28. Moreover, in scenarios where adaptation failed, higher phase fluctuations were noted in the case of the qRL.
Conclusion: Our results demonstrate that quantum RL approaches provide a feasible and promising framework for sequential clinical decision-making in adaptive SBRT.
Contact Email: