Feasibility of real-time workflow segmentation for tracked needle interventions.

Abstract

Computer-assisted training systems promote both training efficacy and patient health. An important component for providing automatic feedback in computer-assisted training systems is workflow segmentation: the determination of what task in the workflow is being performed. Our objective was to develop a workflow segmentation algorithm for needle interventions using needle tracking data. Needle tracking data were collected from ultrasound-guided epidural injections and lumbar punctures, performed by medical personnel. The workflow segmentation algorithm was tested in a simulated real-time scenario: the algorithm was only allowed access to data recorded at, or prior to, the time being segmented. Segmentation output was compared to the ground-truth segmentations produced by independent blinded observers. Overall, the algorithm was 93% accurate. It automatically segmented the ultrasound-guided epidural procedures with 81% accuracy and the lumbar punctures with 82% accuracy. Given that the manual segmentation consistency was only 84%, the algorithm's accuracy was 93%. Using Cohen's d statistic, a medium effect size (0.5) was calculated. Because the algorithm segments needle-based procedures with such high accuracy, expert observers can be augmented by this algorithm without a large decrease in ability to follow trainees in a workflow. The proposed algorithm is feasible for use in a computer-assisted needle placement training system.