Surgical gesture recognition with time delay neural network based on kinematic data

Automatic gesture recognition during surgical procedures is an enabling technology for improving advanced assistance features in surgical robotic systems (SRSs). Examples of such advanced features are user-specific feedback during execution of complex actions, prompt detection of safety-critical situations and autonomous execution of procedure sub-steps. Video data are available for all minimally invasive surgical procedures, but SRS could also provide accurate movements measurements based on kinematic data. Kinematic data provide low dimensional features for gesture recognition that would enable on-line processing during data acquisition. Therefore, we propose a Time Delay Neural Network (TDNN) applied to kinematic data for introducing temporal modelling in gesture recognition. We evaluate accuracy and precision of the proposed method on public benchmark dataset for surgical gesture recognition (JIGSAWS). To evaluate the generalization capability of the proposed method, we acquired a new dataset introducing a different training exercise executed in virtual environment. The dataset is publicly available to enable other methods to be tested on it. The obtained results are comparable with other methods available in literature keeping also computational performance compatible with on-line processing during surgical procedure. The proposed method and the novel dataset are key-components in the development of future autonomous SRSs with advanced situation awareness capabilities.