Research On Key Technologies Of Non-rigid Registration In Robotic Navigation System For Lung Puncture Surgery

In recent years,lung cancer has become one of the cancers with the highest incidence,seriously threatening people’s health.Brachytherapy is an important method for the treatment of lung cancer by precisely implanting radiotherapy particles into the target to kill cancer cells.The puncture surgery robot can improve the efficiency of brachytherapy surgery with the advantages of precise positioning and stable control,but the inability to accurately locate the lesion target is an important factor restricting its development.The main difficulty is that the lungs will be greatly deformed by the breathing movement.Non-rigid registration techniques are the main means to solve this problem.With the application of deep learning in the field of non-rigid registration,it has overcome the problems of low efficiency and poor robustness of traditional methods,and has become the mainstream algorithm for current non-rigid registration.However,in the registration of lung CT images,the methods based on deep learning still have the problems that the vector field is prone to collapse and cannot cope with large-scale deformation well.In this thesis,with the overall goal of improving the navigation accuracy of brachytherapy surgical robots,and the specific research goal of improving the accuracy and efficiency of lung CT image registration,a series of studies have been carried out around the unsupervised learning-based lung CT image registration algorithm.The main research contents and innovations are as follows:(1)Design and implement an end-to-end unsupervised deep learning-based registration algorithm for fully convolutional networks.Considering that it is difficult to provide the supervised data required for training in practical clinical applications,this thesis has designed an unsupervised network model,and has compared two different networks to explore the impact of different network structures on the registration accuracy and efficiency.The experimental results showed that the unsupervised learning registration model designed in this thesis can register lung CT images of any two phases without retraining or adjusting network parameters,which is more efficient and robust than traditional methods.(2)Design and implement a lung CT image registration algorithm based on multi-scale feature fusion.In view of the large deformation of the lungs with the breathing movement,we extracted the deformation features from the receptive fields of different scales,and realized the registration of large-scale deformations on the premise of ensuring that the registration of local details is not distorted.In addition,the training objective of the network is to minimize the similarity measure between images.Aiming at the folding of the deformation field after registration,an anti-folding regular term was added to suppress the deformation that does not meet the anatomical significance in the deformation field while ensuring the accuracy.Aiming at the degradation and gradient disappearance problems of the network with increasing depth,we introduced residual units to improve the network performance.(3)Build an experimental platform to verify the effect of the lung CT registration algorithm proposed in this thesis on the performance improvement of the puncture surgery robot system.In this thesis,based on the research topic,on the basis of the original experimental platform of the puncture surgery robot system,we used the registration algorithm to predict the movement trajectory of the target point,thus adding the function of breathing movement follow-up to the original system.On this basis,we designed a simulated surgery experiment in combination with the actual application scenario of clinical surgery,and verified the impact of the registration algorithm in this thesis on the overall performance improvement of the puncture surgery robot system.