Towards robotic arthroscopy: ‘Instrument gap’ segmentation

Abstract

This paper evaluates the ability of visual segmentation algorithms to detect the space inside the knee joint; as recorded by a surgeon’s arthroscopic video camera, during minimally invasive surgery. We call this space the ‘instrument gap’. Video data was obtained during cadaver experiments, and three segmentation algorithms were tested and compared against a thousand marked-up frames of the instrument gap, prepared by an expert surgeon. Algorithms tested include adaptive thresholding, watershed, and level set active contours. Each algorithm has unique capabilities, but for the instrument gap the adaptive thresholding segmentation was found to execute faster on the test platform, and achieved similar or more accurate results relative to the other algorithms across all data sets.