Anne van Dalen

A review on the current applications of artificial intelligence in the operating room I 225 8 CNN and hidden Markov models (HMMs), a popular application for ordinal or temporal data within AI, recognized different phases in the surgical procedure with an accuracy of 90%. 17 Tool Detection and Localization Tool detection and localization adds to the precision of phase recognition. By recognizing more subtle and detailed activities, tool detection and localization may be informative for predicting operative steps and length of operation. Deep learning techniques were used to research tool detection and localization in laparoscopic images and videos. Using a CNN, results show a mean average precision of 87% in tool detection and 88% in tool localization. 17 Human Detection and Pose Estimation Since the people are the main actors in the OR, detecting their position and estimating their poses by localizing their body parts can provide useful information for optimizing workflow. With the use of external videos, the ability to estimate the specific body poses of the people in the OR was investigated. The mean per joint position error (MPJPE) was used as a quantitative measure for 2D and 3D body part localization. Deep learning approaches yielded the best results in both 2D and 3D pose estimation with an average MPJPE of 17 and 5 cm, respectively. 17 Endoscopic Guidance System Weede et al. described an autonomous endoscopic guidance system based on ML. The system is capable of collecting and processing data on the movements of surgical instruments in recorded videos of surgical procedures. Subsequently, with the use of trajectory clustering, maximum likelihood classification, and HMMs, the system uses this information to predict trajectories that are used to guide the endoscope. The results show a hit rate of over 89% for predicting the movement of the surgeon’s instruments, leading to 29.2% less camera movements and improved visibility. 27,32