Contrast-Free Liver Tumor Detection Using Ternary Knowledge Transferred Teacher-Student Deep Reinforcement Learning

Abstract

Contrast-free liver tumor detection technology has a significant impact on clinics due to its ability to eliminate contrast agents (CAs) administration in the current tumor diagnosis. In this paper, we proposed a novel ternary knowledge transferred teacher-student DRL (Ts-DRL) as a safe, speedy, and inexpensive contrast-free technology for liver tumor detection. Ts-DRL leverages a teacher network to learn tumor knowledge after CAs administration, and create a pipeline to transfer teacher’s knowledge to guide a student network learning of tumor without CAs, thereby realizing contrast-free liver tumor detection. Importantly, Ts-DRL possesses a new ternary knowledge set (actions, rewards, and features of driven actions), which for the first time, allows the teacher network to not only inform the student network what to do, but also teach the student network why to do. Moreover, Ts-DRL possesses a novel progressive hierarchy transferring strategy to progressively adjust the knowledge rationing between teachers and students during training to couple with knowledge smoothly and effectively transferring. Evaluation on 325 patients including different types of tumors from two MR scanners, Ts-DRL significantly improves performance (Dice by at least 7%) when comparing the five most recent state-of-the-art methods. The results proved that our Ts-DRL has greatly promoted the development and deployment of contrast-free liver tumor technology.