Efficient Computer Vision Pipeline for Automated Anesthetic Injection Documentation

Abstract: Purpose: This study introduces a novel computer vision-based approach to auto- mate the documentation of anesthetic injection events in the operating room. The objective is to enhance the accuracy and reliability of documentation by pro- viding precise records of injection start and end times and the exact amount of anesthetic administered. Methods: We developed a comprehensive pipeline combining several computer vision models to analyze surgical videos. The pipeline includes stages for motion detection, frame selection, syringe segmentation, syringe parts segmentation, syringe size classification, volume estimation and sequence analysis. Key techniques include a few-shot segmentation method based on the Segment Anything Model (SAM) and an XGBoost classifier for syringe size classification. The pipeline was tested on a dataset of videos from 19 anesthesiologists performing 16 injections each, using syringes of four different sizes (3ml, 5ml, 10ml, and 20ml). Results: The proposed model achieved an 86.3% success rate in detecting and documenting injection events. In 63% of cases, it accurately predicted the exact amount of anesthetic injected, with accuracy increasing to 87% for a tolerance of up to 1ml and 97% for up to 2ml. Performance varied with syringe size, showing higher accuracy rates for larger syringes. Conclusion: This work presets the potential solution to significantly improve the documentation process of anesthetic injection events, enhancing patient safety and reducing the workload of anesthesiologists.