Computer-assisted Workflow Recognition for Central Venous Catheterization

Loading...
Thumbnail Image
Date
Authors
Hisey, Rebecca
Keyword
Workflow recognition , Surgical tool recognition , Convolutional neural networks , Central Venous Catheterization
Abstract
Purpose: The transition to competency-based medical education from an apprenticeship model requires a greater amount of human resources to evaluate trainees. Therefore, there is an increased need for tools that can evaluate trainees and provide feedback without the need for an expert observer. This thesis describes the development of a system for training central venous catheterization. The procedure includes many tasks that must be performed in a predetermined order. This system, Central Line Tutor, is able to provide trainees with feedback about their performance by evaluating their compliance to proper workflow when practicing in a simulated setting. Methods: Central Line Tutor uses a combination of electromagnetic tracking and tool recognition from a webcam video to analyze trainees’ workflow compliance. Critical tasks such as inserting the needle into the vessel are recognized by tracking the tools’ positions. The remaining workflow tasks were recognized by identifying the tools used in the given tasks. In the proof of concept implementation of the system, we used a color-based approach. This method has some limitations including: insufficient accuracy, and a long setup time. To improve Central Line Tutor’s task recognition, we also implemented a more robust method for tool recognition that involved training a convolutional neural network. We evaluated two different networks, Inception-V3 and MobileNet, and compared their accuracy to the initial color-based approach. Results: Central Line Tutor was designed, implemented, and tested. The system was able to successfully recognize all tasks in the central venous catheterization workflow with a delay of 1.46 ± 0.81s compared to human reviewers. The system was able to recognize all tools using a color-based method, but the convolutional neural networks recognized tools with significantly higher accuracy. Inception-V3 and MobileNet had an average accuracy of 79% and 77% respectively. The color-based method was only able to achieve 8% accuracy. Conclusions: Central Line Tutor is capable of evaluating trainees based on their adherence to procedure workflow. While all tools can be recognized using a color-based method, the use of convolutional neural networks is far more accurate. Overall, Central Line Tutor shows promise as a useful adjunct in competency-based medical education.
External DOI