Physics Colloquium: Hybrid quantum-deep reinforcement learning for Decision Support in Response-adapted Lung Cancer Radiotherapy

Subtle physiological and genetic heterogeneity among patients can result in diverse treatment responses and outcomes even when treated under identical protocols. Thus, the current population-based one-size-fits-all clinical practice in radiation oncology, where all similar staged cancer patients are treated with almost identical radiotherapy regimens, is sub-optimal. Therefore, personalizing radiotherapy according to a patient’s estimated treatment response is crucial for optimizing radiotherapy outcomes and ultimately increasing the rate of cancer survivorship.

In this talk, the speaker will introduce a framework for clinical decision support system that can quantitatively estimate treatment outcomes and recommend a proper patient-specific dose adaptation. The decision support system is based on model-based reinforcement learning framework. At first, an artificial radiotherapy environment was modeled as a Markov decision process using advanced data-driven approach which given a patient’s state and radiation dose can predict patient’s next state and estimate their treatment outcome. Then, a reinforcement learning algorithm was trained using the artificial radiotherapy environment to solve a sequential dose decision-making problem. The objective of the learning problem was to learn an optimal dose decision that optimizes the estimated treatment outcome, i.e., maximize the probability of local tumor control and minimize the probability of radiation induced pneumonitis.

As a novel approach, the clinical dose decisions were modeled as indeterministic quantum states and a practical quantum controller circuit was designed for the decision selection process, which was then paired with the double deep Q-Net algorithm in creating a hybrid quantum-deep reinforcement learning framework. Intrinsically probabilistic quantum state can represent the indeterminism in the clinical decision-making process during RT treatment which arises due to the unavailability of complete information on the treatment course and outcomes. The decision selection quantum controller circuit was designed to be shorter than the quantum computers’ decoherence length and was run in 15 qubit IBMQ 16 Melbourne quantum computer.

As a next step, the artificial radiotherapy environment will be modeled as a quantum Markov decision process to reflect the uncertainty/noise in patient data.

Related Work: Dipesh Niraula, Jamalina Jamaluddin, Martha M Matuszak, Randall K Ten Haken, and Issam El Naqa, Quantum deep reinforcement learning for clinical decision support in oncology: application to adaptive radiotherapy. Sci Rep 11, 23545 (2021). https://doi.org/10.1038/s41598-021- 02910-y