Comparison of Mechanistic and Learning-based Tip Force Estimation on Tendon-driven Soft Robotic Catheters

Area of Science:

• Robotics
• Medical Devices
• Machine Learning

Background:

• Accurate tip force estimation is crucial for controlling soft robotic catheters during medical procedures.
• Existing literature often contrasts the high accuracy of mechanistic models with the faster computation of learning-based models.

Purpose of the Study:

• To compare the accuracy and computational performance of a validated mechanistic tip force estimation method against four learning-based methods (SVR, RF, Ada, DNN).
• To evaluate the capability of learning-based models in detecting the onset of tip contact with soft robotic catheters.

Main Methods:

• A previously validated mechanistic model was compared with Support Vector Regression (SVR), Random Forest (RF), AdaBoost (Ada), and Deep Neural Network (DNN) models.
• Learning-based models were trained on experimental data from an in-house developed robotic catheter.
• Performance was evaluated in a teleoperated catheter manipulation test against ground truth forces.

Main Results:

• Learning-based models achieved lower Mean Absolute Error (MAE) in force estimation (5.1–5.6 mN) compared to the mechanistic model (8.8 mN).
• Learning-based models demonstrated superior accuracy (97.0–97.7%) and precision (97.8–98.8%) in contact detection versus the mechanistic model (89.2% accuracy, 91.7% precision).
• Hyper-parameter optimization significantly enhanced the performance of learning-based models.

Conclusions:

• Optimized learning-based models surpass traditional mechanistic models in both accuracy and precision for soft robotic catheter tip force estimation and contact detection.
• Both mechanistic and learning-based approaches demonstrate acceptable performance for catheter manipulation applications.
• This study highlights the potential of machine learning to advance robotic catheter technology.