Convolutive Attention for Image Registration

Tim J. Parbs, Philipp Koch, Alfred Mertins

In: Proc. European Signal Processing Conference. European Signal Processing Conference (EUSIPCO-2022) August 28-September 2 Belgrade Serbia Seiten 1348-1352 IEEE 2022.


Elastic registration of deformed images is a vitalcomponent of many computer vision tasks, especially whenconsidering medical image data. Deep learning techniques, particularlyU-Nets, offer state-of-the-art performance, but do notyet use the rich spatial information context available in naturalimages. We propose an augmentation based on the recentlyintroduced attention mechanism to allow a U-Net to use spatialimage context. A dedicated convolutive attention scheme hasbeen developed by calculating local similarity scores of themultidimensional inputs. Additionally, a dedicated compositeerror function based on common image similarity measures isintroduced in order to further improve the registration results.To evaluate our approach, we conducted several experiments onan augmented real-world dataset containing cardiac cine MRIscans. The comparison with state-of-the-art registration schemeshighlights the potential of our approach.

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence