Reinforcement Learning to Enhanced Microwave Imaging for Accurate Tumor Detection in Breast Images

Microwave imaging represents a promising technique for non-invasive tumor detection, but yet facing challenges for the accurate identification of tumor areas, given the background noise and the scattering complexities. Machine learning methods can be helpful at this purpose, to improve the reconstruction ability of microwave imaging. In particular, a novel framework is proposed in this work to integrate Reinforcement Learning (RL) with microwave imaging techniques, with the aim to enhance the tumor reconstruction process. The Quadratic Born Iterative Method (BIM) is used in conjunction with U-Net architectures, the proposed approach optimizes reconstruction parameters through, and a Deep Q-Network (DQN) is applied to realize an improved adaptive tumor identification. A numerical validation is performed by using breast images to demonstrate the effectiveness of the proposed framework. Both single-tumor as well as multi-tumor detection exhibits high identification precision, with a slight better accuracy in terms of evaluation metrics for the first case.