Intraoperative Segmentation Through Deep Learning and Mask Post-processing in Laparoscopic Liver Surgery

James Borgars, Jibran Raja, Abhinav Ramakrishnan, Abdul Karim Abbas, Aodhan Gallagher, Ahmad Najmi Mohamad Shahir, Theodora Vraimakis, Sharib Ali

Abstract: Laparoscopic liver surgery is a popular surgical approach due to its capabilities of minimising trauma, complications, and recovery times. The use of a laparoscope allows for developments in the field of machine-assisted surgery due to the availability of intraoperative imagery. Accurate landmark detection of the liver using laparoscopic footage is a dependency to many developments, such as 3D-2D registration. In this paper, we present experimental results measuring the suitability of popular segmentation models, and their compatibility with different loss functions when handling intraoperative images; we also present a pipeline in training models for this segmentation task, including a novel step of applying post-processing techniques to maximise accuracy. Our results are evaluated using precision, Dice similarity coefficient, and a symmetric distance metric. Our results show that through the use of our proposed pipeline, models retain their ability to generalise, and can lead to noticeably improved accuracy both quantitatively and qualitatively. We demonstrate the feasibility of utilising post-processing to improve predictions. Finally, possible future directions in this field following from our results are discussed. The code from this research has been made available and can be accessed here: https://github.com/ARMADILLO-VISION/SLiPPA.

External IDs:doi:10.1007/978-3-031-98694-9_15