SAME: Deformable Image Registration Based on Self-supervised Anatomical Embeddings

In this work, we introduce a fast and accurate method for unsupervised 3D medical image registration. This work is built on top of recent algorithm self-supervised anatomical embedding (SAM), which capable computing dense anatomical/semantic correspondences between two images at the pixel level. Our named SAM-enhanced registration (SAME), breaks down into three steps: affine transformation, coarse deformation, deep deformable Using SAM embeddings, enhance these steps by finding more coherent correspondences, providing features loss function with better semantic guidance. We collect multi-phase chest computed tomography dataset 35 annotated organs each patient conduct inter-subject quantitative evaluation. Results show that SAME outperforms widely-used traditional techniques (Elastix FFD, ANTs SyN) learning based VoxelMorph least \(4.7\%\) \(2.7\%\) in Dice scores separate tasks within-contrast-phase across-contrast-phase registration, respectively. achieves comparable performance to best method, DEEDS (from our evaluation), while being orders magnitude faster 45 s 1.2 s).