Studies On Brain Medical Image Registration Based On Deep Learning

As the main direction of medical image analysis,medical image registration has important research significance and clinical application value,and is widely used in disease diagnosis,surgical guidance and radiation therapy.Although the method based on deep learning has effectively promoted the rapid development of current medical image registration research,there is still much space for improvement in its registration accuracy.This paper starts with the selection and optimization of image features in registration.Based on brain medical image registration,this paper proposes a contour weakly supervised brain registration model based on attention mechanism and a brain registration model based on contrastive learning and attention mechanism.Its main innovations are as follows:（1）The existing registration networks are mostly based on U-Net encoding and decoding structures.The encoder and decoder perform skip connections without optimizing the weight of the input decoder’s features,resulting in insufficient attention to local information and incomplete details of the generated deformation field,resulting in the loss of some local information in the image.By introducing a channel attention mechanism based on skip connections in the registration network,feature weight filtering is achieved,which compensates for the shortcomings of existing registration methods.The introduction of this attention mechanism can more accurately capture the deformation information of details in the feature fusion process,improving the accuracy of registration results.（2）By adding the contour feature loss function7)₄8)0)（92）)to the similarity loss function,uses contour feature information as additional geometric constraints.Adding this loss function can reduce the impact caused by the difference of gray values between the registering image pairs,and can more accurately measure the similarity between the generated registered image and the reference image.（3）The traditional contrastive learning method usually calculates the contrast loss in the last stage of feature extraction,ignoring the differences in feature information between different stages.This article conducts multi-stage comparison during the downsampling process of the registration network,calculates the comparison loss of features in each stage,compensates for the shortcomings of traditional comparative learning methods,optimizes feature extraction methods,and improves the robustness and stability of the registration network.Combining the above three points,this article proposes a contour weakly supervised brain registration method based on attention mechanism（CACF）and a brain registration method based on contrastive learning and attention mechanism（CACC）.The experimental results on LPBA40 dataset show that these two methods can effectively improve the accuracy of brain medical image registration compared with the current mainstream methods.