基于无监督深度学习的多模态手术轨迹快速分割方法

A Fast Approach for Multi-Modality Surgical Trajectory Segmentation with Unsupervised Deep Learning

传统的手术机器人轨迹分割方法存在耗时长、分割准确度差且容易产生过度分割等问题.为解决上述问题,本文提出了一种基于特征提取网络DCED-Net(密集连接的卷积编码-解码网络)的多模态手术轨迹分割方法.DCED-Net采用无监督方法,不必进行十分耗时的人工标注,使用密集连接结构,使图像信息能更有效地在卷积层间传递,从而提高了特征提取质量.将特征提取后的视频数据和运动学数据投入转移状态聚类(TSC)模型得到预分割结果.为进一步提高分割精度,提出了一种基于轨迹段间相似性的合并后处理算法,通过衡量轨迹段间的4个相似性指标,包括主成分分析、互信息、数据中心距离和动态时间规整,将相似度高的分割段进行迭代合并,从而降低过度分割造成的影响.公开数据集JIGSAWS上的大量实验证明,与经典的轨迹分割聚类方法相比,本文方法的分割准确率最高提升了48.4%,分割速度加快了6倍以上.

Traditional trajectory segmentation approaches for surgical robot are time consuming, low-accuracy and prone to over-segmentation. For those problems, a multi-modality surgical trajectory segmentation approach is proposed based on DCED-Net(densely-concatenated convolutional encoder-decoder network) feature extraction network. DCED-Net adopts an unsupervised approach and a densely-concatenated structure, and the time consuming manual annotation is not required.Therefore, the image information can be transferred more effectively between convolutional layers, and the quality of extracted features is improved. The kinematic data and video data obtained after feature extraction are input into a transition state clustering(TSC) model to get pre-segmentation results. To further improve the segmentation accuracy, a post-merger processing algorithm based on the similarity between trajectory segments is proposed. By measuring four similarity indicators between trajectory segments, including principal component analysis, mutual information, data center distance, and dynamic time warping, the segments with high similarity are iteratively merged to reduce the impact of over-segmentation. A lot of experiments on the public data set JIGSAWS show that the proposed approach can increase the segmentation accuracy by up to 48.4% and accelerate the segmentation speed by more than 6 times, compared with the classical trajectory segmentation and clustering methods.