Development of a machine learning-based tension measurement method in robotic surgery

Background Over 300,000 colorectal surgeries are performed annually in the U.S. with up to 10% complicated by anastomotic leaks, which cause significant morbidity and mortality. Despite its significant association with anastomotic leaks, tension is predominantly assessed intraoperatively using subjective metrics. This study aims to assess the feasibility of a novel objective method to assess mechanical tension in ex vivo porcine colons. Methods This research was conducted using the da Vinci Research Kit (dVRK). First, a machine learning algorithm based on a long short-term memory neural network was developed to estimate the pulling forces on robotic arms of dVRK. Next, two robotic arms were used to apply upward forces to five ex vivo porcine colon segments. A force sensor was placed underneath the colons to measure ground-truth forces, which were compared to estimated forces calculated by the machine learning algorithm. Root mean square error and Spearman's Correlation were calculated to evaluate force estimation accuracy and correlation between measured and estimated forces, respectively. Results Measured forces ranged from 0 to 17.2 N for an average experiment duration of two minutes. The algorithm's force estimates closely tracked the ground-truth sensor measurements with an accuracy of up to 88% and an average accuracy of 74% across all experiments. The estimated and measured forces showed a very strong correlation, with no Spearman's Correlation less than 0.80 across all experiments. Conclusion This study proposes a machine learning algorithm that estimates colonic tension with a close approximation to ground-truth data from a force sensor. This is the first study to objectively measure tissue tension (and report it in Newtons) using a robot. Our method can be adapted to measure tension on multiple types of tissue and can help prevent surgical complications and mortality.