Para-sagittal Midclavicular Plane Localization in 3DCT Using Multi-Agent Dueling Network

Locating the para-sagittal on the left portal vein in the abdominal enhanced 3DCT data is a critical step in clinical diagnosis and multi-modality fusion. The process of manually annotating landmarks on the anatomical structure is timeconsuming, laborious, and requires a wealth of professional knowledge. In recent years, reinforcement learning (RL) has developed rapidly, making it possible to deal with complex medical data. This paper is based on model-free RL to train a joint multi-agent detection system. The sharing part of the network provides implicit communication for collaborative work between agents, and independent output helps each agent achieve the purpose of locating specific critical points, respectively. Results show that this method can identify three anatomical landmarks simultaneously on the para-sagittal on the left portal vein correctly, introducing a novel formulation for locating the standard plane of volumes and can deal with complex situations such as intraoperative image fusion. Also, the multi-model convergence time is 30% shorter than that of the single-agent model.