Research On Medical Image Registration Based On Neural Architecture Search

Medical image registration technology is the basic task of medical image analysis and can provide key information for medical decision-making.In recent years,the rapid development of deep learning has driven the progress of medical image registration technology.Compared with traditional methods,the deep learning-based medical image registration method has higher accuracy and faster reasoning speed.However,it is costly to design a high performance and high flexibility medical image registration network model.In this thesis,we introduce the neural architecture search technology into the design of medical image registration model to automatically obtain the most suitable image registration model architecture for the current data.In addition,this thesis also summarizes the common characteristics of the search model,which can be used to guide the architectural design of medical image registration network in the future.In this thesis,we use the neural architecture search technique to search the medical image registration model at the level of cell architectures.Firstly,the search space for cell architectures is designed in this thesis.Then the model parameters and the architectural parameters were optimized alternately by using the gradient based a bilevel optimization algorithm.Finally,the search results are decoded and the optimal architecture of medical image registration network is constructed.In this thesis,we tested the network architecture using a dataset of brain MRI images.The results show that the performance of the searched network architecture is better than the manually designed network architecture.In addition to the above studies,the search space of the medical image registration model in the network level architecture is carried out in this thesis Different from the above research on the search of network cell architectures,a grid-like network search space is defined in this thesis,in which single-path and multi-path registration model are searched respectively.In this thesis,we tested the network architecture using several publicly available CT image datasets.The results show that the accuracy of the searched network architectures in this thesis is better than the manually designed network architectures,and the accuracy of the multi-path registration model is better than the single-path model.To sum up,this thesis introduces the neural architecture search into the design of medical image registration model,and studies cell architectures and network paths in the network architecture.This thesis proposes that the medical image registration model based on neural network architecture search has higher performance,stronger flexibility and wider application scenarios than the manually designed network model.