联合仿射与微分同胚图像配准的框架
A Framework for Joint Affine and Diffeomorphic Image Registration
浏览论文内容
中文总结 AI 辅助
该研究提出联合仿射-微分同胚图像配准框架，含FA和DA模型，优化策略结合渐进仿射丰富与变分加权，在Dice重叠度上优于CARL等基线模型。
中文摘要 AI 辅助
解剖图像配准通常依赖于串行流程：先估计仿射对齐，再固定该对齐并应用非刚性微分同胚变形。该两步流程常导致结果欠佳，因为初始阶段可能吸收局部变形，使传递给微分同胚配准的残差产生偏差。为解决此问题，我们引入基于大变形模型的联合仿射-微分同胚框架，在单次优化中同时估计全局仿射与局部微分同胚运动。我们提出两种模型：全仿射（FA）模型，结合仿射与微分同胚变形；分解仿射（DA）模型，将仿射部分限制为旋转、平移和各向异性缩放。为在图像配准任务中数值实现这些模型，我们开发了定制优化策略，结合渐进式仿射丰富（逐步增加仿射分量的复杂度）与变分加权方案（平滑处理仿射与微分同胚分量间的由粗到细过渡）。在2D合成数据集和IXI队列的3D脑部MRI上评估，我们的无监督方法避免了串行基线的病理性变形。我们证明，联合公式在Dice重叠度上优于两个最先进的深度学习基础模型CARL和uniGradI-CON，以及采用FLIRT（经典仿射配准）加LDDMM的串行基线。我们的实现公开可用。
英文摘要
Anatomical image registration commonly relies on a sequential pipeline where an affine alignment is estimated first and then held fixed while a non-rigid diffeomorphic deformation is applied. This two-step process often leads to suboptimal results, as the initial stage can absorb local deformations, biasing the residual passed to the diffeomorphic registration. To address this, we introduce a Joint Affine-Diffeomorphic framework, based on the large deformations model, that estimates both global affine and local diffeomorphic motions simultaneously within a single optimization. We propose two models: a Full Affine (FA) model that combines affine and diffeomorphic deformations, and a Decomposed Affine (DA) model that restricts the affine part from FA to rotations, translations, and anisotropic scalings. To numerically implement these models for image registration tasks, we develop a tailored optimization strategy that combines progressive affine enrichment, gradually increasing the complexity of the affine component, with a variational weighting scheme that smoothly manages the coarse-to-fine handover between the affine and diffeomorphic components. Evaluated on 2D synthetic datasets and 3D brain MRIs from the IXI cohort, our unsupervised approach avoids the pathological deformations of sequential baselines. We demonstrate that our joint formulation outperforms in Dice overlap two state-of-the-art deep learning foundation models, CARL and uniGradI-CON, as well as a sequential baseline using FLIRT for the classical affine registration followed by LDDMM. Our implementation is publicly available.
