Arthroscopic Tool Classification using Deep Learning

Shoulder arthroscopy is a common surgery to diagnose and treat tears to improve patient's quality of life. Quality of cleaning the tear during shoulder arthroscopy significantly affects the outcome of the surgery. Appropriate cleaning is necessary to reduce healing time and avoid feature pain in the area. In this paper, we used convolutional neural networks to automatically differentiate between two tools-electrocautery and shaver tools- that are used during the cleaning phase of a shoulder arthroscopy. We captured images from the actual shoulder arthroscopy videos. We used 8,691 images that contain the shaver tool, 7,773 images that contain the electrocautery tool, and 4,834 images that contain no tools. Our results showed that average accuracy of our model is 99.1(+/- 0.49) %. For the electrocautery tool precision and sensitivity was calculated as 0.988 and 0. 988, respectively. For the shaver tool precision and sensitivity was calculated as 0.993 and 0. 988, respectively. For the no tool scenes precision and sensitivity was calculated as 1.0 and 1. 0, respectively.