Robotic transesophageal echocardiography: system design and deep learning-based kinematic modeling

Area of Science:

• Robotics
• Medical Imaging
• Machine Learning

Background:

• Manual transesophageal echocardiography (TEE) requires specialized skills and can be challenging in various patient positions.
• Existing robotic TEE systems have limitations in mechanical design and kinematic modeling accuracy.
• Cable-driven mechanisms in TEE probes often suffer from nonlinearities like friction and tension asymmetry.

Purpose of the Study:

• To develop a robotic transesophageal echocardiography (TEE) system with enhanced mechanical design and deep learning-based kinematic modeling.
• To replicate essential degrees of freedom (DOF) of manual TEE procedures for improved control and versatility.
• To create a foundation for semi-autonomous TEE systems capable of supporting both diagnostic and interventional procedures.

Main Methods:

• Integration of a teleoperated UR5 manipulator with a TEE probe system featuring 6 DOF at the handle and 2 DOF at the gastroscope tube.
• Development of a data-driven kinematic model using recurrent neural networks with LSTM units, trained on 42,000 pose-command pairs across three bending configurations (0°, 45°, 90°).
• Validation of the system's performance through experimental tracking of probe position and orientation, assessing accuracy and real-time inference capabilities.

Main Results:

• The deep learning kinematic model achieved high position tracking accuracy with root-mean-square errors (RMSE) below 1.267 mm across all configurations.
• Mean orientation errors were below 8.503°, with a notable 4.947° error at the critical 90° bend configuration.
• The system demonstrated real-time performance with an inference time of 1.8 ms and coordinate frame independence, confirming true kinematic learning.

Conclusions:

• The developed robotic TEE system, enhanced by deep learning kinematic modeling, successfully replicates manual TEE DOF with high precision.
• The system's ability to handle various patient positions and its real-time performance are crucial for clinical deployment.
• This integration represents a significant step towards semi-autonomous TEE systems for both diagnostic examinations and image-guided structural heart interventions.