Reinforcement Active Modeling for Flexible Needle Shape Prediction in Multilayer Tissues.

Area of Science:

• Robotics
• Medical Engineering
• Machine Learning

Background:

• Predicting flexible needle shape during tissue puncture is challenging due to complex tissue interactions.
• Existing methods struggle with accurate prediction, especially in multilayer tissues with non-Gaussian noise.
• Reliable needle shape prediction is crucial for minimally invasive robotic surgery.

Purpose of the Study:

• To develop a novel reinforcement learning-based active modeling scheme for predicting robotic flexible needle deflection.
• To enhance robustness against non-Gaussian noise during needle insertion into complex tissues.
• To improve the accuracy and reliability of flexible needle shape prediction.

Main Methods:

• Developed an active modeling scheme using an extended Kalman filter with the maximum correntropy criterion for noise insensitivity.
• Integrated reinforcement learning into the active modeling scheme, creating the reinforcement active modeling (RAM) framework.
• Proved theoretical convergence of the proposed scheme using the Banach fixed-point theorem.

Main Results:

• The proposed RAM scheme demonstrated superior performance in predicting flexible needle deflection.
• Under non-Gaussian noise, RAM achieved a 46.4% reduction in Root Mean Square Error (RMSE).
• RAM achieved over 76.1% reduction in variance (Var) during insertion into unknown multilayer tissue.

Conclusions:

• The novel RAM framework effectively predicts flexible robotic needle deflection in complex tissues.
• The method shows significant improvements in accuracy and robustness, particularly under non-Gaussian noise conditions.
• This approach offers a reliable solution for accurate needle shape prediction in robotic procedures.