Force classification during robotic interventions through simulation-trained neural networks

Intravitreal injection is among the most frequent treatment strategies for chronic ophthalmic diseases. The last decade has seen a serious increase in the number of intravitreal injections, and with it, adverse effects and drawbacks. To tackle these problems, medical assistive devices for robotized injections have been suggested and are projected to enhance delivery mechanisms for a new generation of pharmacological solutions. In this paper, we present a method aimed at improving the safety characteristics of upcoming robotic systems. Our vision-based method uses a combination of 2D OCT data, numerical simulation and machine learning to classify the range of the force applied by an injection needle on the sclera. We design a neural network to classify force ranges from optical coherence tomography (OCT) images of the sclera directly. To avoid the need for large real data sets, the network is trained on images of simulated deformed sclera. This simulation is based on a finite element method, and the model is parameterized using a Bayesian filter applied to observations of the deformation in OCT images. We validate our approach on real OCT data collected on five ex vivo porcine eyes using a robotically guided needle. The thorough parameterization of the simulations leads to a very good agreement between the virtually generated samples used to train the network and the real OCT acquisitions. Results show that the applied force range on real data can be predicted with 93% accuracy. Through a simulation-trained neural network, our approach estimates the force range applied by a robotically guided needle on the sclera based solely on a single OCT slice of the deformed sclera. Being real-time, this solution can be integrated in the control loop of the system, permitting the prompt withdrawal of the needle for safety reasons.