الفهرس | Only 14 pages are availabe for public view |
Abstract We present a method that uses visual information in a cholecystectomy procedure{u2019}s video to detect the surgical workflow. While most related work relies on rich external information, we rely only on the endoscopic video used in the surgery. We fine tune a convolutional neural network and use it to get mid-level features representing the surgical phases. Additionally, we train DPM object detectors to detect the used surgical tools, and utilize this information to provide discriminative high-level features. We present a pipeline that employs the mid and high level features by using one vs all SVMs followed by an HHMM to infer the surgical workflow. We present detailed experiments on a relatively large dataset containing 80 cholecystectomy videos. Our best approach achieves 90% detection accuracy in offline mode using only visual information |