MIDL Vision 2021

“ We had a new image registration pipeline to register, ” Henrik explains. “ For an individual, the acquisition order is the 2p microscopy image, then the backlit image, then the Nissl image. Each time you move to a different microscope, or you alter the image a bit, the tissue is slightly deformed. At the end we must undo all these deformations, and because we have so many images, we need to do that automatically. ” The challenge is the amount of data. The model must work reliably because if something goes wrong, it is not feasible to correct it manually. The team uses the ANTS Image Registration Toolkit , a popular and widely used image registration tool. Initially, they solely used a multi- modal image registration to register the images of different modalities, and it worked well for most of the images. However, sometimes the image registration had a bad initial alignment for scale and orientation, and then it failed. The team figured out they could attribute that problem to the different imaging modalities. “ This work has been going on for several years and image translation models and artificial generation of new images have become very popular, ” Henrik tells us. “ We thought, why not adapt the image modality so that the image registration itself has an easier job 15 Henrik Skibbe VISION MIDL Registration pipeline (outline)