Convolutive Attention for Image Registration
Tim J. Parbs, Philipp Koch, Alfred Mertins · 2022 30th European Signal Processing Conference (EUSIPCO) · 2022
Elastic registration of deformed images is a vital component of many computer vision tasks, especially when considering medical image data. Deep learning techniques, par-ticularly U-Nets, offer state-of-the-art performance, but do not yet use the rich spatial information context available in natural images. We propose an augmentation based on the recently introduced attention mechanism to allow a U - Net to use spatial image context. A dedicated convolutive attention scheme has been developed by calculating local similarity scores of the multidimensional inputs. Additionally, a dedicated composite error function based on common image similarity measures is introduced in order to further improve the registration results. To evaluate our approach, we conducted several experiments on an augmented real-world dataset containing cardiac cine MRI scans. The comparison with state-of-the-art registration schemes highlights the potential of our approach.