Learning to Predict Action Based on B-ultrasound Image Information

In the medical field, B-ultrasound is an important way to diagnose diseases. However, due to the lack of professional sonographers, patients have to queue for a long time for examination. Or due to some easily contagious diseases, sonographers cannot directly contact the patient for examination. Therefore, it is necessary to use robotic arms to perform automated B-ultrasound examinations on patients. In our work, the strategy of how to move the probe to detect the kidney is studied. The sonographer is required to hold a special probe instrument to collect the demonstration data, including the B-ultrasound image, as well as the posture and force information of the probe. Then, we leverage the data learning to realize the guidance of the B-ultrasound probe action. In this paper, supervised learning is firstly used to predict actions according image inputs. In other words, the supervised network is input with the B-ultrasound image and output posture and force that the probe should reach at the next moment. Based on the supervised learning, an actor-critic reinforcement learning algorithm that uses asymmetrical data is proposed to improve the utilization of data and enhance the generalization of neural networks.