SAME++: A Self-supervised Anatomical eMbeddings Enhanced medical image registration framework using stable sampling and regularized transformation

TL;DR

SAME++ addresses the challenge of semantically meaningful medical image registration by leveraging Self-supervised Anatomical eMbeddings (SAM) to guide a four-stage pipeline: SAM-affine, SAM-coarse, SAM-deform, and SAM Instance Optimization. The approach combines SAM-derived semantic correspondences with both linear and non-linear transforms, including a diffeomorphic SVF-based deformable stage, and introduces stable sampling via cycle-consistency to reduce false matches. Empirical results across head & neck, chest, and abdomen CT datasets (covering more than 50 labeled organs) show Dice-score improvements of 4.2%–8.2% over leading methods, with substantially faster inference than traditional optimization-based methods. The work demonstrates that SAM-based semantic guidance, coupled with instance-level refinement, yields accurate, robust registrations and provides a practical, plug-in framework for improving existing registration pipelines.

Abstract

Image registration is a fundamental medical image analysis task. Ideally, registration should focus on aligning semantically corresponding voxels, i.e., the same anatomical locations. However, existing methods often optimize similarity measures computed directly on intensities or on hand-crafted features, which lack anatomical semantic information. These similarity measures may lead to sub-optimal solutions where large deformations, complex anatomical differences, or cross-modality imagery exist. In this work, we introduce a fast and accurate method for unsupervised 3D medical image registration building on top of a Self-supervised Anatomical eMbedding (SAM) algorithm, which is capable of computing dense anatomical correspondences between two images at the voxel level. We name our approach SAM-Enhanced registration (SAME++), which decomposes image registration into four steps: affine transformation, coarse deformation, deep non-parametric transformation, and instance optimization. Using SAM embeddings, we enhance these steps by finding more coherent correspondence and providing features with better semantic guidance. We extensively evaluated SAME++ using more than 50 labeled organs on three challenging inter-subject registration tasks of different body parts. As a complete registration framework, SAME++ markedly outperforms leading methods by $4.2\%$ - $8.2\%$ in terms of Dice score while being orders of magnitude faster than numerical optimization-based methods. Code is available at \url{https://github.com/alibaba-damo-academy/same}.

Figures (5)

• Figure 1: The framework of SAME++. Based on the SAM space, we break down image registration into four steps: keypoint-based affine transformation, coarse deformation, deep deformable registration, and instance optimization. (a) Illustration of one incorrect match based on SAM. (b) Eliminating false correspondence via cycle consistency matching.
• Figure 2: Boxplot of the performance of different affine registration methods on 13 organs in the abdomen CT dataset.
• Figure 3: Comparison of deformable registration methods on all organ groups on the neck dataset using boxplot.
• Figure 4: Comparison of deformable registration methods on all organ groups on the chest dataset using boxplot.
• Figure 5: Qualitative comparisons between SAME++ and the second best methods Deeds, LapIRN and Deeds for Neck, Chest, and Abdomen registrations, respectively. Top row: visualization of coronal head neck and warped CT slices. Middle row: Overlay of coronal chest CT (gray) and warped segmentation (color) slices. Bottom row: Differences between warped and fixed coronal abdominal scans.