Towards Better Laparoscopic Video Database Organization by Automatic Surgery Classification

Abstract

Minimally invasive surgery is an important breakthrough in the domain of medicine. Not only does it improve the quality of surgery, but the underlying digitization also provides invaluable information that opens up many possibilities for teaching, assistance during difficult cases, and quality evaluation. For instance, with a well-organized database, professors are one click away from showing and comparing various surgical procedures in their classes; surgeons can also retrieve and observe a video segment of a specific surgical task performed by another surgeon in varying conditions. However, to the best of our knowledge, database organization is done manually by experts. Considering the large number of surgical videos recorded, manual annotation is a tedious task. In this paper, we take the first step towards automatic surgical database organization by introducing the laparoscopic video classification problem, which consists of automatically identifying the type of abdominal surgery performed in a video. In spite of the visual challenges of such videos, such as blank frames, rapid movement, and sometimes incomplete recording, we show that we can rely on visual features alone to classify the videos with high accuracy. We use kernel Support Vector Machines (SVMs) for this classification task and compare their performance on different types of visual features. We also show that the result can be improved by combining the visual features using Multiple Kernel Learning approach. The classification pipeline demonstrates a classification accuracy of 91.39% on a database of 151 abdominal videos totaling over 200 hours of 8 different kinds of surgeries performed by 10 surgeons.