InterConv: Interacting Convolution Applied to End-to-End Unsupervised Image Registration

Our research has found that traditional convolutions are limited in their ability to process input data effectively. They only consider a single field of view and do not take into account the asymmetric relationships that exist between data points. This can lead to issues in tasks such as image registration, where a fixed image's target region may only correspond to a single point on a moved image. To address this issue, we propose a new type of convolution called InterConv. InterConv can extract the relationship between corresponding positions of data through an asymmetric field of view, reducing the redundancy of network parameters. We have deduced the theoretical basis for InterConv and have incorporated it into an image registration network based on the integration of Inception and U-Net. Our end-to-end unsupervised registration of image achieved excellent results. The use of InterConv ensures the registration network's effectiveness in difficult registration tasks while also reducing the network's parameters from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1,205,788$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{884, 764}$</tex>. This reduction in parameters has led to a decrease in the network's computational load and model file size, resulting in space resource savings.