Research On Computer-assisted Surgical Instrument Tracking And Postoperative Evaluation Based On Deep Learning

The real-time surgical instrument tracking algorithm is an important part of the computer-assisted minimally invasive surgery system,and it is the eyes of the entire system.In computer-assisted minimally invasive surgery systems,surgical instrument tracking algorithms can provide surgeons or surgical robots with real-time,accurate information,such as pose,position of the instruments,etc.,and then they make further decisions based on this information.In addition,the evaluation of the performance of the operator during the entire operation process can not only feedback to the operator and improve his surgical skills,but also can be used to train novice surgeons or surgical robots,which can reduce unnecessary risks from surgery.In this paper,we propose two easy surgical evaluation method.Therefore,this paper mainly focuses on the real-time tracking algorithm of minimally invasive surgical tools and surgical evaluation method.Due to the rapid development of computer technology,deep learning has become more and more popular in the fields of object detection,image classification and video understanding,especially convolutional neural network models.This paper proposes a real-time multi-instrument minimally invasive based on deep learning and surgical evaluation indexes.The main research contents are as follows:Firstly,the tracking algorithm proposed by the predecessors is introduced,and its advantages and disadvantages are analyzed.Based on this,we have proposed three detection algorithms for minimally invasive surgical tools and constructed corresponding convolutional neural models.They are the convolutional neural network and Hough transform based single surgical tool tip location method,the real-time multi-tool tracking algorithm based on heat map regression network and bounding box regression network,and the real-time multi-tool tracking algorithm based on RSSD and CONV-RNN.The first algorithm uses convolutional neural network and Hough transform to complete the edge detection of the surgical tool and then locate the tip position;the second algorithm uses two forms of network cascade to track the tools in real time;the third algorithm combines SSD network and the convolutional recurrent neural network solve the problem that the previous tracking method does not use the temporal and spatial context information.between video frames and frames,which effectively improves the tracking accuracy of the algorithm;On the premise of ensuring that the accuracy rate will not drop a lot,continuously optimizing the main network structure makes the tracking speed of the network model faster than other methods.Secondly,the datasets are constructed.The datasets used to train your own network and the comparison method are publicly available standard datasets.The first algorithm is tested on the EndoVisSub data set of a single tool and the standard dataset;the latter two algorithms are evaluated on the ATLAS da Vinci simulation surgery scene dataset and partial datasets in the MICCAI competition,and some of the datasets are annotated by us.Thirdly,comparison with other mainstream surgical tool tracking algorithms is done.We train each algorithm using the same training dataset,and use the same test dataset to compare the accuracy,speed of each algorithm to track the surgical instruments.Comparative experiments show that the first algorithm we proposed has higher accuracy than the other two single-tool positioning methods;the accuracy of the second algorithm is the highest among multi-tool tracking methods,and the speed of the third algorithm is fastest.At the same time,we also do a comparative experiment on convolutional recurrent neural network for the third algorithm.Finally,postoperative evaluation method is proposed.For this purpose,we specifically use output of surgical instrument tracking algorithm.By analyzing and processing this information,we get the trajectory and action timelines of the surgical each surgical instrument,and then we analyze whether the newly obtained information can be used as an index for postoperative evaluation.After analysis,this information has a certain relationship with the practitioner's intraoperative performance.We propose a postoperative evaluation method based on convolutional neural network,of which the input are the trajectory map and action timelines,but the proposed method could not be verified due to the lack of data sets with score.The preliminary experiments on the simulated dataset show that our method is feasible...