Deep Learning Based Assisted Decision Making For Endoscopic Intubation In Narrow Human Cavity

The scene which an endoscopic intubation faced was a narrow human cavity.This cavity is dark,untextured and hard to identify.But the efficiency of intubation can be greatly improved if we realize an intelligent process of endoscopic intubation with the aid of computers.Therefore,in this paper,to research and explore the realization of computeraided decision making of endoscopic intubation and the intelligent decision-making of unmanned endoscopic intubation,endoscopic intubation was used as the application scene of intelligent medicine,endotracheal intubation was used as the main entry point,and deep learning technology was combined.Meanwhile,a set of data related to our research work was collected and made.In a word,the paper‘s main research contents are as follows:To solve the problem that the traditional endoscopic assisted method based on 3d reconstruction is inconvenient to use in real-time endotracheal intubation decision making,this paper propose an end-to-end decision-making method of endoscopic intubation based on deep learning.A set of discrete operation instruction sets for the movement direction of the endoscope intubation are designed,and a network is constructed to map the intubationassisted decision-making problem into an image classification problem.The classification network has two output branches,which respectively complete the end point judgment decision and the forward direction decision of the intubation process.This method has a high accuracy rate of 96.51% on the decision task in the forward direction,and a slightly lower accuracy rate of 91.25% on the decision task of the end point judgment.In view of the opacity and low interpretability of the end-to-end method to the decision process,an intelligent endoscope intubation decision method based on target detection is proposed in this paper.This method explicitly divides the entire endoscopic intubation decision-making process into two parts: perception and decision.First,the target forward position coordinate point in the video frame is output through the target point perception network,and then the endoscopic operation decision instruction is output through rule mapping.Among them,the output range of the target point-aware network is [0,459],and the RMSE on the x-axis and y-axis are 11.2337763 and 10.096154,respectively.In order to verify the possibility and feasibility of the end-to-end endoscope-assisted decision-making method based on deep learning in the application of actual medical scenarios,this paper starts from the actual needs,starting from the hardware and software parts,designing and building a complete set of a prototype of an endoscope-assisted decisionmaking system with a visual user interface.The prototype has passed the simulation test first,and the decision result of the professional doctor is evaluated as the true value,showing good performance.Then a actual scene test was conducted on a volunteer and some feedback data was obtained.