Smart Tumour Targeting by Reinforcement Learning

By drawing stimulating parallels between the optimization and nanobiosensing processes, a novel framework of computational nanobiosensing (CONA) employs nanorobots as computing agents to perform “smart” searching of tumours, greatly improving targeting performance. Optimal routing in the vascular network and learning the biological gradient field (BGF), a peritumoral field triggered by the tumour, are two fundamental problems in CONA. In this paper, we present a novel reinforcement learning (RL)-based CONA strategy for tumour targeting, which can address these two problems together. The overall reward of the Q-Learning mission includes the Markov reward and BGF reward. The Markov reward is employed to train nanorobots to avoid colliding vessel obstacles during the searching process, resulting in learn the structure of the vascular network. While the BGF reward, learning prior information of BGF, benefit in faster convergence of searching process. In other words, this hybrid strategy, taking advantage of feasible path planning of the vessels and the information of BGF together, accelerates finding the optimal path planning for the tumour in a complicated vessel network. By introducing different weights between the Markov and BGF rewards, it is found that the best weight is roughly the same for different scenarios.