Force classification during robotic interventions through simulation-trained neural networks.

Area of Science:

• Ophthalmology
• Robotics
• Medical Imaging
• Machine Learning

Background:

• Intravitreal injections are common for chronic eye diseases, but increased use leads to adverse effects.
• Robot-assisted injection systems aim to improve drug delivery and safety.
• Current robotic systems require enhanced safety features to manage needle-tissue interaction forces.

Purpose of the Study:

• To develop a vision-based method for improving the safety of robotic intravitreal injection systems.
• To classify the force range applied by an injection needle on the sclera using a combination of 2D OCT data, numerical simulation, and machine learning.
• To enable real-time force estimation for prompt needle withdrawal.

Main Methods:

• A neural network was designed to classify force ranges directly from optical coherence tomography (OCT) images of the sclera.
• The network was trained on simulated scleral deformation images generated using a finite element method.
• A Bayesian filter parameterized the finite element model using OCT deformation observations to create realistic training data.

Main Results:

• The developed approach was validated on ex vivo porcine eyes using a robotically guided needle.
• Simulations demonstrated strong agreement with real OCT data due to thorough parameterization.
• The system achieved 93% accuracy in predicting the applied force range on real OCT data.

Conclusions:

• A simulation-trained neural network can estimate the force applied by a robotic needle on the sclera from a single OCT slice.
• The real-time nature of this solution allows integration into robotic control loops.
• This facilitates prompt needle withdrawal, significantly enhancing procedural safety.