Quantum Reinforcement and Deep Learning for Decision Support in Response-Adapted Lung Cancer Radiotherapy
D Niraula1*, J Jamaluddin2, M Matuszak2, R Ten Haken2, I El Naqa1, (1) Moffitt Cancer Center, Tampa, FL, (2) University of Michigan, Ann Arbor, MI
Presentations
WE-F-TRACK 6-6 (Wednesday, 7/28/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Purpose: To develop a robust decision support framework for personalized recommendations in response-adapted lung cancer Radiotherapy based on quantum computing and deep learning estimated tumor and normal tissue responses.
Methods: Combined quantum reinforcement learning (QRL) and deep learning (DL) techniques were implemented for creating the proposed framework. DL models were used to create an Artificial Radiotherapy Environment, consisting of a transition function for governing the patients’ state dynamics as a function of dose, and an RT outcome estimator of local control and grade 2 radiation induced pneumonitis for a given state. RT state features were selected from a multi-objective Bayesian network: IP10, GLSZM-ZSV, cxcr1-Rs2234671, Tumor-gEUD, and Lung-gEUD. Decisions were modeled as quantum states to reflect the associated uncertainty in treatment outcomes, and QRL was used for optimizing the decision-making process. Two new approaches were developed to improve the robustness of the framework,: (1) a quantum controller circuit to evaluate the framework on a quantum processor and (2) a constrained DL approach to guide the RT outcome estimator learning of radiation response. We trained our framework on 67 non-small lung cancer patients treated with RT and then validated it on the independent data of 20 prospective patients.
Results: The constrained DL modeling improved the learning of experimentally observed monotonic radiation response outcomes. The framework constructed with QRL and quantum controller using 5-qubit IBMQ quantum processor demonstrated improved performance over the ones trained in 3-qubit and 5-qubit quantum simulators. The RMSE values between the clinical dose decision and framework recommendations were 0.81 (0.78-0.84) Gy/frac, 0.93 (0.84-0.99) Gy/frac, and 0.86 (0.84-0.93) Gy/frac, respectively. The validation results showed reasonable performance in the institutional prospective dataset with an RMSE value of 0.96 (0.80-1.07) Gy/frac.
Conclusion: This work demonstrates a robust QRL-DL framework for response-adapted lung cancer radiotherapy.
Handouts
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email