TGF-Net: Tri-Gated Fusion Network for Coarse-to-Fine Unsupervised Deformable Medical Image Registration

Medical image registration aligns anatomical structures by spatially transforming images within the same modality. However, existing deformable registration methods often struggle to accurately model large anatomical deformations and complex local structural variations, limiting registration accuracy in clinically challenging scenarios. To address these limitations, we propose TGF-Net, a Tri-Gated Fusion Network for unsupervised deformable medical image registration. TGF-Net employs dual encoders with non-shared weights to extract multiscale features from the moving and fixed images, followed by a coarse-to-fine decoder for progressive deformation estimation. The core of the proposed framework is a tri-gated fusion module that explicitly models complementary interactions among decoder, fixed-image, and warped-image features to progressively refine deformation fields. Furthermore, an adaptive multi-branch pooling strategy aggregates fine-scale structural responses from each feature stream and generates feature-specific gating weights, enabling more effective preservation of local anatomical details during refinement. The proposed framework was trained and evaluated on three publicly available datasets, including the Laboratory of Neuro Imaging Probabilistic Brain Atlas (LPBA), AbdomenCT, and Information eXtraction from Images (IXI). Experimental results demonstrate that TGF-Net achieves competitive performance in terms of Dice similarity coefficient, average symmetric surface distance, Hausdorff distance, and deformation regularity across multiple anatomical structures. These findings indicate that explicitly modeling feature interactions among three complementary sources through tri-gated fusion is an effective strategy for fine-grained deformation modeling in coarse-to-fine unsupervised deformable registration. The magnitude of the improvement varies across datasets and metrics, and TGF-Net is therefore best regarded as competitive with, rather than uniformly superior to, recent state-of-the-art methods.