An image registration method and apparatus based on depth learning

Abstract

The invention provides an image registration method and a device based on depth learning. The image registration model is constructed by utilizing a capsule network. The scalar representation and pooling mechanism in the traditional depth learning convolution network are replaced by a vector feature representation and routing mechanism, and capsules of different levels are connected step by step to carry out feature combination. An image fusion network based on capsule vector is constructed, and the fusion image with the same dimension size as the reference image is output as the registrationoutput. The loss function based on similarity measure between images is constructed to optimize the unsupervised learning registration network by feedback training network parameters. The image registration operation with high accuracy and high robustness is realized.

Description

technical field [0001] The present invention relates to the field of image processing, in particular to an image registration method and device based on deep learning. Background technique [0002] Medical image registration refers to the process of matching and superimposing two or more medical images acquired at different times, different imaging devices or under different conditions. At present, information technology represented by deep learning and high-end medical imaging technology have continuously made major breakthroughs, and registration using deep learning technology has become a new hot spot in the field of medical image registration. [0003] An existing medical image registration method based on convolutional neural network (Chinese patent application CN201711017916.8). The method introduces a tensor column into a weight matrix of a fully connected layer of a convolutional neural network to obtain a tensor convolutional neural network, thereby obtaining at le...

AI Technical Summary

Problems solved by technology

However, when processing the same instanced feature, the spatial invariance is guaranteed by using the downsampling operation, ignoring the spatial hierarchy and orientation information of the image, making it insensitive to the rotation and angle information of the image, most of the existing registration models use The convolution operation is easy to lose the direction feature information, thus affecting the accuracy of the registration model

[0004] The following problems exist in the commonly used methods at present: 1. The existing methods of deep learning technology to deal with medical image registration problems usually use convolution operations and pooling operations to obtain image features, and when dealing with the invariance of image features Due to the difference in position and angle transformation between the features of the registration image and the reference image, the traditional convolutional neural network cannot effectively detect its specific direction information. This lack of spatial transformation information will eventually affect the performance of the registration model.

But in fact, the pooling operation makes it lose spatial information. This situation has little effect on the classification task, but has a great impact on the registration task.

When using the convolutional neural network for registration, only by using methods such as data enhancement, the convolutional network can better identify the position change information such as rotation, but in essence, it is easy to lose the spatial level information and direction information of the image. , making it insensitive to the recognition of spatial transformation information of the same semantic content of the registered image, thus affecting the registration accuracy

2. The above-mentioned use of deep learning technology for medical image registration adopts a supervised learning method. Supervised image registration needs to manually mark the registration image area of ​​​​the training samples. During the training process, the prediction information is compared with the label information for feedback. Network loss, the process of manually labeling samples is time-consuming and labor-intensive, and the labeling process for large-sample datasets takes too long to achieve

In addition, manual sample labeling is highly dependent on the professional knowledge of professional doctors. Large quantities of manual sample labeling are prone to labeling errors, and registration results are easily affected by labeling quality.

3. The pooling operation adopted by the traditional convolutional neural network is easy to lose the spatial position information, and to achieve high-precision results often requires a large number of training samples in the training set, so that the neural network can fully learn the shapes of various targets. Pose features, so traditional deep learning registration models often require a large number of training samples

For medical image processing tasks, it is difficult to obtain large batches of training samples with relatively uniform specifications due to reasons such as patient privacy and different diagnostic instruments. Therefore, traditional deep learning registration tasks require a large training set due to the network structure, but it is difficult to fully obtain it. Training samples, which affect the registration accuracy

Images