Convolutional neural network-based surgical instrument detection

Abstract

BACKGROUND: Minimally invasive surgery (MIS), unlike open surgery in which surgeons can perform surgery directly, is performed using miniaturized instruments with indirect but careful observation of the surgical site. OBJECTIVE: Instrument detection is a crucial requirement in conventional and robot-assisted MIS, which can also be very useful during surgical training. In this paper, we propose a novel framework of using two three-layer convolutional neural networks (CNNs) in a series to detect surgical instrument in in-vivo video frames. METHODS: The two convolutional neural networks proposed in this paper have different tasks. (i) The former CNN is trained to detect the edges points of the instrument shaft directly from images patches. (ii) The latter is trained to locate the instrument tip also from images patches after the former detection finishes. RESULTS: We validated our method on the publicly available EndoVisSub dataset and a standard dataset, and it detected tools with an accuracy of 91.2% and 75% respectively. CONCLUSION: Our two-step detection method achieves better performance than other existing approaches in terms of detection accuracy.