Nobuyoshi Takeshita, Daichi Kitaguchi, Masaaki Ito. National Cancer Center Hospital East
Introduction: Tracking and analyzing surgical tool movements, anatomical structures and bleeding in laparoscopic-surgical views are beneficial for training and safety. Furthermore, automatic recognition of surgical workflow is useful for operating room management and prediction of adverse event during surgery. In this work, we made use of the video database of laparoscopic sigmoidectomy and introduced semantic segmentation and workflow recognition using convolutional neural networks (CNN).
Methods: We used a CNN architecture to perform automatic segmentation of surgical tools, anatomical structures and bleeding: U-NET which was designed for biomedical image segmentation. For the semantic segmentation purpose, we manually labeled surgical tools, inferior mesenteric artery (IMA) and bleeding in the videos of laparoscopic sigmoidectomy from 300 cases of the database. We pixelwisely annotated spatial bounds of the objects. As for surgical tools, we used 1326 annotated frames for training data and 80 frames for test data in machine learning. Finally, we evaluated precision rate, recall rate and F-measure. The datasets for semantic segmentation of IMA and bleeding were created in the same way and then used as training and test data for machine learning. For the workflow recognition purpose, we labeled the surgical phases (Preparation, Recto-rectal space, Median to lateral approach 1, Inferior mesenteric artery transection, Median to lateral approach 2, lateral approach, Rectal mobilization, Mesorectal dissection, Transection and anastomosis, Inferior mesenteric vein transection, Closing) for every frame in the surgical videos. We used 63 videos for training data and 8 videos for test data in machine learning.
Results: F-measures of the tip of grasper, linear dissector and Maryland were 0.873, 0.763 and 0.686, respectively. The results of automatic semantic segmentation of IMA and bleeding were also acceptable from clinical perspective. F-measures of automatic recognition of surgical workflow were around 0.900.
Conclusions: The results of semantic segmentation of surgical tools, IMA and bleeding seemed promising. Furthermore, the results of workflow recognition were acceptable for contribution in clinical fields. CNN approach must be feasible for automatic image recognition in laparoscopic surgery.
Presented at the SAGES 2017 Annual Meeting in Houston, TX.
Abstract ID: 94187
Program Number: P685
Presentation Session: Poster Session (Non CME)
Presentation Type: Poster