YOLACT++を用いた手術用鉗子の識別

抄録

Forceps tracking in laparoscopic surgery contributes to improved surgical outcomes. We identified forceps by deep learning. Since it is important to identify forceps in real-time, we selected YOLACT++ for fast and accurate segmentation and verified whether the detection speed can be maintained in the video. We annotated a total of 2537 images combining multiple datasets including various surgical environments and divided them into training, validation, and test data at a ratio of approximately 8:1:1. The training was conducted with a batch size of 32, iterations of 100106, and epochs of 1588, and the results showed that the forceps identification speed was 25.79 fps and accuracy was 84.31 %. The results of the test using the trained model with this hyperparameter showed that the forceps identification speed was 28.01 fps and accuracy was 71.42 % for images, and the forceps identification speed was 17.70 fps for the video.