Kinematics Data Representations for Skills Assessment in Ultrasound-Guided Needle Insertion

摘要

Ultrasound-guided needle insertion is a difficult skill to learn and, in the context of competency-based medical education, requires continual monitoring of trainees' performance. This work investigates two standard neural network architectures, temporal convolutional networks and long short-term memory networks, for automated classification of skill level based on kinematics data. It examines which data representations are optimal for skills assessment using the proposed architectures in low data scenarios. The data representation had significant effect on the computed results. But given the optimal data representation, the proposed architectures achieve skills classification on two simulated ultrasound-guided needle insertion tasks with better performance than summary statistics. Thus, neural networks can be an effective tool for skills assessment in ultrasound-guided interventions; however, it is recommended to search over the space of data representations when limited data is available.