Machine Learning Model for Force Prediction in a Soft Robotic Colonoscope Sleeve

Colonoscopies are important for early detection of colorectal cancer, polyps, and other abnormalities. However, due to an inability to monitor internal contact forces with current endoscopic tools, the procedure can be painful and cause tissue perforation and bleeding. A soft robotic sleeve with embedded optical sensors and pneumatic actuators was developed to monitor and minimize contact forces between the colon and the colonoscope.

In this study, a feed-forward neural network was developed in Matlab that can be trained on calibration data to monitor contact forces with greater accuracy and less sleeve-specific tuning. Network topology and hyperparameters were explored and optimized. A calibration machine was designed and developed to automatically obtain bending and force training datasets that were uniformly distributed over the parameter space. Force validation testing was performed to evaluate the performance of the neural network.

Results from initial testing show that the algorithm can accurately classify the force as above or below the actuator deployment threshold. The neural network was further shown to be capable of real-time force predictions and was incorporated into a graphic user interface that can be used by surgeons to operate the system.