A Deep Reinforcement Learning Approach for Pre-Planning Problems in Robotic Intra-Operative Ultrasound: a Virtual Environment-Based Analysis
Robotic intra-operative ultrasound has the potential to improve the routine practice of diagnosis and procedure guidance that is currently performed manually. One of the common challenges towards standardization and intelligence in robotic ultrasound is pre-planning, which uses pre-operative data from the patient and computes the desired input and target pose of the robot given a desired view of the ultrasound image. In this paper, we present a deep reinforcement learning-based approach to solving the problem of pre-planning the ultrasound image for a four-degree-of-freedom intra-operative cardiac ultrasound robot. Evaluation in a virtual environment demonstrates the feasibility and robustness of the proposed method. The mean squared error representing the distance of the standard ultrasound view was found to be 5.70 ± 0.33 mm for the given task. As this is done in a semi-automatic way, it is believed that the proposed method can effectively reduce the operational difficulties for medical staff and improve the accuracy of future ultrasound view acquisition and scanning.
Paper
Full text
A Deep Reinforcement Learning Approach for Pre-Planning Problems in Robotic Intra-Operative Ultrasound: a Virtual Environment-Based Analysis
Semantic Scholar · Engineering · 2023
Abstract
Robotic intra-operative ultrasound has the potential to improve the routine practice of diagnosis and procedure guidance that is currently performed manually. One of the common challenges towards standardization and intelligence in robotic ultrasound is pre-planning, which uses pre-operative data from the patient and computes the desired input and target pose of the robot given a desired view of the ultrasound image. In this paper, we present a deep reinforcement learning-based approach to solving the problem of pre-planning the ultrasound image for a four-degree-of-freedom intra-operative cardiac ultrasound robot. Evaluation in a virtual environment demonstrates the feasibility and robustness of the proposed method. The mean squared error representing the distance of the standard ultrasound view was found to be 5.70 ± 0.33 mm for the given task. As this is done in a semi-automatic way, it is believed that the proposed method can effectively reduce the operational difficulties for medical staff and improve the accuracy of future ultrasound view acquisition and scanning.