A Reinforcement Learning Approach for Real-Time Articulated Surgical Instrument 3-D Pose Reconstruction

Abstract

3D pose reconstruction of surgical instruments from images stands as a critical component in environment perception within robotic minimally invasive surgery (RMIS). The current deep learning methods rely on complex networks to enhance accuracy, making real-time implementation difficult. Moreover, diverging from a singular rigid body, surgical instruments exhibit an articulation structure, making the annotation of 3D poses more challenging. In this paper, we present a novel approach to formulate the 3D pose reconstruction of articulated surgical instruments as a Markov Decision Process (MDP). A Reinforcement Learning (RL) agent employs 2D image labels to control a virtual articulated skeleton to reproduce the 3D pose of the real surgical instrument. Firstly, a convolutional neural network is used to estimate the 2D pixel positions of joint nodes of the surgical instrument skeleton. Subsequently, the agent controls the 3D virtual articulated skeleton to align its joint nodes’ projections on the image plane with those in the real image. Validation of our proposed method is conducted using a semi-synthetic dataset with precise 3D pose labels and two real datasets, demonstrating the accuracy and efficacy of our approach. The results indicate the potential of our method in achieving real-time 3D pose reconstruction for articulated surgical instruments in the context of RMIS, addressing the challenges posed by low-texture surfaces and articulated structures.