2D3D Medical Image Registration Based on Fully Convolutional Network Approved by Attention Mechanism
Deqing Kong, Guogang Cao, Wenjv Li, Sicheng Li, Shu Zhang · 2021
2D3D medical image registration is very vital in the fields of surgical navigation and radiotherapy. In order to solve the problems of being time-consuming and trapping in local extreme value in previous methods, a 2D3D medical image registration method based on fully convolution network and attention mechanism was proposed. In this method, a large number of digital reconstructed radiological images were input as training data to convolutional neural network, and image features were automatically extracted, so as to predict the registration transformation parameters. The network structure was based on the modified VGGNet-16 approved by an attention mechanism, and a global average pooling layer was applied to replace the fully connected layer. The experimental results showed that the average offset error of the proposed method was 0.08 mm, and the average angular error was 0.05°. Its time consumption was approximately 42ms, which was much less than the traditional method's of 305.81s. The proposed method meets the real time and precision requirements of medical image registration, avoids the iterative process of traditional methods, and effectively improves the registration efficiency.