Surgical Phase Recognition of Short Video Shots based on Temporal Modeling of Deep Features

Abstract

Recognizing the phases of a laparoscopic surgery (LS) operation form its video constitutes a fundamental step for efficient content representation, indexing and retrieval in surgical video databases. In the literature, most techniques focus on phase segmentation of the entire LS video using hand-crafted visual features, instrument usage signals, and recently convolutional neural networks (CNNs). In this paper we address the problem of phase recognition of short video shots (10s) of the operation, without utilizing information about the preceding/forthcoming video frames, their phase labels or the instruments used. We investigate four state-of-the-art CNN architectures (Alexnet, VGG19, GoogleNet, and ResNet101), for feature extraction via transfer learning. Visual saliency was employed for selecting the most informative region of the image as input to the CNN. Video shot representation was based on two temporal pooling mechanisms. Most importantly, we investigate the role of 'elapsed time' (from the beginning of the operation), and we show that inclusion of this feature can increase performance dramatically (69% vs. 75% mean accuracy). Finally, a long short-term memory (LSTM) network was trained for video shot classification based on the fusion of CNN features with 'elapsed time', increasing the accuracy to 86%. Our results highlight the prominent role of visual saliency, long-range temporal recursion and 'elapsed time' (a feature so far ignored), for surgical phase recognition.