MG-UNet:无监督3D医学图像的多粒度配准框架OA

Medical image registration is crucial for clinical diagnosis and treatment.However,existing methods face challenges in handling complex anatomical structures and ensuring stability in the deformation field,while also struggling to balance multi-scale information modeling with computational efficiency.This paper proposes a novel unsupervised 3D medical image registra-tion model—MG-UNet(Multi-Granularity UNet),aiming to enhance registration accuracy and optimize the rationality and smoothness of the deformation field.MG-UNet adopts a multi-scale,multi-granularity feature modeling approach:it enhances fine-grained information representation in 3D space through a multi-directional convolutional initial layer,extracts local and global features under different receptive fields via a Multi-Granularity Dilated Convolution(MGDC)module,and incorporates a Multi-attention Fusion(MAF)module to reinforce feature responses in key regions.Experimental results demonstrate that MG-UNet achieves state-of-the-art performance on both the OASIS and IXI datasets.On the OASIS dataset,the Dice coefficient im-proved by 2.3%,HD95 decreased by 12.2%,and ASSD decreased by 16.3%;on the IXI dataset,the Dice coefficient reached 0.769,further surpassing the current best methods.Additionally,the MG-UNet-tiny variant achieved the lowest HD95 and the lowest percentage of negative Jacobian determinants,with a parameter count of only 1.597 M,significantly reducing computa-tional load compared to Transformer-based models.These results demonstrate the effectiveness of the MG-UNet model in reli-ably capturing the deformation relationships of anatomical structures in medical images,offering a new technical pathway for effi-cient medical image registration.