Multi-Scale Constrained Lung Medical Image Registration Based on Feature Reweighting

Registration plays a key role in medical image analysis and is an important part of intelligent medicine. Traditional registration methods need a lot of iterative computation and take a long time. The registration methods based on deep learning can meet the requirements of rapidity and accuracy. To improve the registration effect of the convolution neural network on lung medical images, we present a multi-scale registration network based on feature reweighting. The mixed dilated convolution is introduced into the registration task and the receptive field of the convolution kernel is increased under the condition that the number of parameters and the size of the output characteristic graph is unchanged. An attention module is added in the skip connection between the decoder and the encoder to make the model pay more attention to the salient areas in the image. The weighted loss function at each stage trains the whole network and adopts the cumulative transformation mechanism. Experiments show that the proposed method can improve registration efficiency and meet the requirements of medical image registration speed and accuracy.