Evaluation of Keypoint Detectors and Descriptors in Arthroscopic Images for Feature-Based Matching Applications

Knee arthroscopy is a very challenging surgical procedure that would strongly benefit from systems that can continuously map the inside of the knee, localize the arthroscope and surgical tools, and control instruments using visual information. A fundamental requirement of most of these systems is the correct and fast matching of visual features. Feature-based systems have been demonstrated in laparoscopy but have yet to be extended to the context of arthroscopy. As an essential initial step, this letter proposes the first detailed experimental evaluation of the performance of state-of-the-art feature detection and description methods on arthroscopic images. We first evaluate the behavior of eight keypoint detectors under 133 setting combinations using four different metrics in a dataset with 100 in-vivo images. We then combine the previous detectors with six feature descriptors and evaluate the matching performance for the resulting features (detector+descriptor) across five different image transformations. A validation is performed using in-vivo images acquired under varying camera motion and illumination. The results show that the best-performing feature in knee-arthroscopy images is DoG+SIFT, while features BRISK+SURF and BRISK+BRISK are recommended for viable implementations in real time.