Research On 3D Biomedical Brain Image Registration Algorithm Based On Deep Learning

Image registration plays an important role in medical image processing,remote sensing image processing and other fields.Through image registration,stitching and fusion between different data can be achieved,thereby providing richer and more intuitive image information.The traditional image registration algorithm uses image feature point information or image voxel gray value statistical information to iteratively optimize the similarity metric function to obtain the best transformation matrix when the similarity between the images to be registered is the highest,and then achieve registration.This kind of algorithm usually has a slow iteration speed and cannot meet the real-time requirements in scenarios such as intraoperative registration.With the development of deep learning algorithms,convolutional neural networks are widely used in image segmentation,target detection and other fields,and are also introduced into the field of image registration.Registration algorithms based on deep learning do not require iteration and are generally faster.This article focuses on deep learning based medical brain image registration algorithm,the main content is as follows:(1)This paper proposes an unsupervised image registration algorithm for 3D U-Net cascaded dilated convolution sub-network.Image registration requires fine spatial correspondence between pixels,but the existence of U-Net pooling down-sampling operation in the encoding stage leads to the loss of spatial detail information of the image(feature map),which is not conducive to fine image matching quasi.Therefore,the algorithm proposed in this paper is to cascade a sub-network module that does not contain downsampling operations and consists of continuous hole convolution behind the 3D U-Net network.This can achieve the original resolution of the image(feature map)without losing space In the case of detailed information,the image is finely registered.The unsupervised image registration network of the entire 3D U-Net cascade continuous hole convolution sub-network can achieve the effect of the image from coarse registration to fine registration as a whole.Compared with the 3D U-Net registration network,the image registration algorithm of the 3D U-Net cascaded hole convolution subnetwork proposed in this paper has improved theDice value of cerebrospinal fluid,gray matter,and white matter on the LPBA40 data set by12.44 %,3.31%,and 1.77%,respectively,increased by 14.82%,5.56%,and 5.99% on the IBSR18 data set.(2)This paper proposes a U-Net registration algorithm based on interval filling-attention mechanism.The gap filling mechanism can solve the problem of semantic gaps in the cross-layer connection of the feature maps in the U-Net codec stage.At the same time,for the similarity of anatomical structures existing between biomedical registration images,this paper introduces an attention mechanism to make the network focus Areas with large differences in the image,thereby improving the accuracy of registration.Compared with the 3D U-Net registration network,the Dice values of cerebrospinal fluid,gray matter,and white matter after registration in the U-Net network registration algorithm based on the gap-fill-attention mechanism proposed in this paper have increased by 4.74%,1.33%,and 1.85%,respectively on the LPBA40 data set,increased by 10.50%,4.97%,and4.65% on the IBSR18 data set.