Newborn Brain Magnetic Resonance Based On Deep Learning Image Registration Research

In the process of modern clinical diagnosis,doctors often need to analyze the lesions through a series of medical imaging techniques to finally confirm the diagnosis.This shows that medical imaging technology plays a vital role in the development of modern medicine.Medical image registration,as an important and difficult point in the preprocessing of medical images,has also received extensive attention.Traditional image registration algorithms,such as B-spline and SYN,have achieved good results in medical registration,but they have certain limitations.For example,poor adaptability,a method or a set of parameters is only applicable to a specific modal or even a specific data set;the processing speed is slow,and most of the traditional registration algorithms adopt iterative optimization methods,which leads to their processing speed is quite slow and unable to real-time registration.In recent years,the rapid development of deep learning has provided us with new ideas for solving medical image registration problems.This paper proposes an unsupervised medical image registration model based on convolutional neural networks,and improves the accuracy of its registration results through a series of improvements.The work of this article mainly focuses on the following aspects:(1)Preprocess the magnetic resonance image of the newborn’s brain.Since medical images have more complex details than natural images,the preprocessing process of medical images is more complicated.In this paper,a series of pre-processing is performed on the collected original images to improve the quality of the data set and thus the accuracy of model prediction.(2)A network model of unsupervised image registration—RegNet is proposed.The model introduces the spatial transformation network(STN)so that the network only needs to input moving images and fixed images during the training process to realize the training process.Experiments show that the registration speed of the network model is very fast,and the registration can be completed in less than one second through GPU acceleration technology,but the accuracy of the registration result is low.(3)In order to improve the accuracy of the RegNet registration result,this paper introduces the spatial activation function and the amSE attention module.The spatial activation function enables the network to have a better fitting ability to the data.The amSE attention module introduces spatial information and channel information at the same time.Experimental results show that these two methods can significantly improve the accuracy of RegNet registration results.