Learning optimization updates for multimodal registration

Abstract

We address the problem of multimodal image registration using a supervised learning approach. We pose the problem as a regression task,whose goal is to estimate the unknown geometric transformation from the joint appearance of the fixed and moving images. Our method is based on (i) context-aware features,which allow us to guide the registration using not only local,but also global structural information,and (ii) regression forests to map the very large contextual feature space to transformation parameters. Our approach improves the capture range,as we demonstrate on the publicly available IXI dataset. Furthermore,it can also handle difficult settings where other similarity metrics tend to fail; for instance,we show results on the deformable registration of Intravascular Ultrasound (IVUS) and Histology images.