| Along with the development of the computer technique and medical treatment technique,computer-assisted interventions technique has a rocketing development since its first day,now it has become a hot research field and begins to take shape in industrialization.In the 21st century,the computer's computing power and speed have a exponential growth,the image processing,machine learning and artificial intelligence are lifting off,minimally invasive-surgery is becoming commonplace,modern surgery emphasis on precise and little contact,medical informatization becomes more and more important in modern medical technique and clinical medicine,all of which make the computer assisted minimally invasive-surgery technique have entered foreland of clinical medicine rapidly.One core technology of the computer assisted minimally invasive-surgery is to track the surgical tools in endoscopic videos in real time,using related algorithms,which can help the head surgeons perform the surgery or provide some information needed to make decisions by offering the spatial positions of the surgical tools,such as Da Vinci Si.The basis of the spacial tool tracking is to accurately obtain the two-dimensional position of the surgical tools in real time,to achieve it,we propose a new algorithm of real-time tracking of surgical instruments based on spatio-temporal context and deep learning.The article firstly summarizes the development of the computer-assisted interventions and the computer assisted surgical tool tracking algorithms.According to the advantages and disadvantages of the current computer assisted surgical tool tracking algorithms and the superiority of the deep learning algorithms,we propose a real-time algorithm using a rough localization network and an accurate localization network to detect the surgical tools jointly and the spatio-temporal context learning algorithm to track the tools frame by frame.The main contents of this paper are as follows:1.We construct two convolutional neural networks as accurate localization network,and train the network by using the endoscopic surgical tool tracking datasets,then the network can recognize the surgical tools in the video images by classifying images.Convolutional neural networks are good at image classification,as they can efficiently extract the two-dimensional features.The architectures of our convolutional neural networks are inspired by the classical networks architectures,such as AlexNet,GoogleNet and VGG.We use some methods to improve our networks,such as introducing the spatial transformer network into our architecture,which can correct the spatial pose of the object to decrease the validation errors in the network training,mitigate the impacts of over-fitting based on the limited data,to improve the architecture of the network.2.In order to save the detection time,we propose and achieve the rough localization network.We can get the results of the region proposal network and intersection-over-union by the classification of the rough localization network,and make the detection faster.On the other hand,the cooperationof the two localization networks can reduce the detection region of a more accurate network,so it can use shorter stride to traverse the region and get a more accurate position.3.We use the spatio-temporal context learning algorithm to track the tools frame by frame in real time.In cooperation of the two localization networks,the spatio-temporal context learning algorithm can automatically track the computer assisted surgical tools in real time.4.In order to show the superiority of our algorithm's accuracy and speed,we compared our method experimentally to four other visual tracking methods using eight existing online and in-house datasets.The comparison includes mean error,standard deviations,accuracy,the error curves and FPS,which confirm the superiority of our strategy of the detection and tracking and the feasibility of our research direction,also demonstrate that our algorithm is accurate and in real time. |
PDF Full Text Request