Figure 2.

Quantum deep reinforcement learning algorithm for optimal decision making in knowledge-based adaptive radiotherapy. Schematic of a quantum deep reinforcement learning (qDRL) algorithm for optimal decision making in knowledge-based adaptive radiotherapy. qDRL employs deep q-net as a decision optimization algorithm and employs quantum state as the decision. Here, qDRL is a model-based algorithm that utilizes an artificial radiotherapy environment (ARTE) as the RL model. The qDRL artificially intelligent (AI) agent feeds in patient’s state $s_t$ in its memory (deep q-net) and obtains a set of q-values for a range of dose $(\left\{q_{{\left|d\right.〉}_t}\right\})$. The agent then selects the dose with the highest q-value and performs quantum amplification of that dose on a superimposed quantum dose decision state, $\left|D\right.〉$. A quantum measurement is performed on the amplified state. The obtained dose measurement, ${\left|d\right.〉}_t$, along with the state $s_t$ is fed into the ARTE. ARTE is composed of three functions in succession: (1) transition function, (2) RT outcome estimator, and (3) reward function, which predicts the patient’s next state $s_{t+1}$, RT treatment outcome in terms of probability of local control, $p_{\mathit{LC}}$, and probability of radiation induced pneumonitis of grade 2 or higher, $p_{RP2}$, and reward value, $r_{t+1},$ for the state-dose-decision pair. $s_{t+1}$, and $r_{t+1}$ are then used by the quantum agent to update its memory. This cycle is repeated until the agent finds a terminating state, after which a new cycle is initiated for a different patient. Five relevant biophysical features from radiomics, cancer and normal tissue radiation, cytokines, and genetics, were selected to represent the patient’s state based on our earlier work¹³.