A Fast Unsupervised Approach for Multi-Modality Surgical Trajectory Segmentation

Abstract

To improve the efficiency of surgical trajectory segmentation for surgical assessment and robot learning in robot-assisted minimally invasive surgery, this paper presents a fast unsupervised method using video and kinematic data, followed by a promoting procedure to address the over-segmentation issue. An unsupervised deep learning network called dense convolutional encoder-decoder network (DCED-Net) is first proposed to extract more discriminative features from videos in an effective way. DCED-Net has several advantages. It compresses the encoding-decoding structure, strengthens the feature propagation, and avoids the manual annotation. To further improve the accuracy of segmentation, on one hand, a modified transition state clustering model is employed with a strategy of reducing the redundancy of transition points. On the other hand, the segmentation results are promoted by identifying the over-segmented trajectories based on predefined similarity measurements. Extensive experiments on the public data set JIGSAWS show that with our method, the percentage increase in accuracy is 20.3% and the percentage decrease in time cost is 92.6%.