Instance-Specific Brain Image Registration: Learning Deformable Fields from a Single Image Pair without External Datasets

Abstract Objective: To eliminate the reliance on large-scale external training datasets and achieve high-precision alignment for complex brain structures by overcoming the prevalent challenges of data scarcity and domain shifts in current deep learning-based registration methods. Approach: We propose a novel Instance-Specific Deep Brain Registration Framework that independently trains a dedicated model for each specific image pair. By reconceptualizing a single image pair as an internal dataset generator, we introduce a Multi-Scale Random Patch Learning strategy to construct a diverse local-to-global sample space. Architecturally, we design a Context-Aware Flow Predictor integrated with a Recurrent Context Transfer mechanism within a coarse-to-fine paradigm. Furthermore, a Topology-Aware Adaptive Optimization strategy is incorporated during the inference phase. By dynamically tracking the local spatial folding status, it adaptively adjusts spatial regularization to effectively suppress topological distortions while maximizing voxel-level accuracy. Main results: Extensive experiments on mono-modal, multi-modal, and pathological brain MRI datasets demonstrate that our method can handle large anatomical deformations without relying on external training datasets. It achieves highly competitive registration performance compared to mainstream population-based and implicit neural representation methods. Crucially, it balances this alignment precision with effective numerical topology preservation. Significance: This framework challenges the traditional population-based learning paradigm, offering a highly flexible, robust, and promising solution for practical neuroimage analysis, particularly in clinical scenarios characterized by extreme data scarcity and complex cross-modality variations.