Automated Bleeding Identification in Surgical Videos Using Deep Learning

Background: Analysis of operative data using convolutional neural networks (CNN) is expected to improve surgeon knowledge and professional skill. Further, the identification of bleeding on surgical videos can lead to improved surgical assessment and navigation. In this study, we performed bleed detection modeling, which had been previously used in surgical videos taken during colorectal procedures, and evaluated the detection accuracy. Methods: A total of 250 objects were annotated in 140 images extracted from five colorectal surgical videos for model training, with 100 images clipped from other videos for validation. The images were annotated and segmented individually for modeling. IBM Visual Insights, including the most popular open-source deep learning framework, Detectron, was used for the CNN. Results: In total, 142/162 bleeds were correctly identified (87.7%) in 100 test images, with a precision of 98.6%. The bleeds were correctly identified in the videos, and graphs indicated the accurate time and duration of the bleed. Conclusions: We evaluated the identification of bleeds using a CNN, which resulted in accurate detection. Real-time high-quality assessment of bleed identification suggests the possibility of clinical application during surgery with simultaneous bleed detection and technical evaluation.