A Zero-Shot Domain Adaptation Framework for Computed Tomography Via Reinforcement Learning and Volume Rendering

Ang Li, Yongjian Zhao, Fan Bai, Jiayi Han, Max Q.-H. Meng, Li Liu

Abstract: The Domain gap is an important issue when applying AI-based approaches to clinical use. Many recent approaches introduce domain adaptation (DA) to eliminate the influence of the domain gap. However, these approaches usually require modifications of network weights and losing the ability in the source (training) domain. To perform zero-shot domain adaptation without forgetting the source domain, in this work, we propose a novel reinforcement learning (RL) based framework which modifies the distribution of the input data instead of the weights of the pre-trained model. The RL agent observes the rendering results of the CT volume and uses a look-up table to perform non-linear mapping on the input to the standard distribution without tuning the original model weights. Since the pre-trained weights are fixed, the model can still generalize well on the source domain. Experiments show that the proposed approach dramatically improves the generalization of the proposed model (Dice score from 73.26% to 77.84%) without forgetting the source domain.