Abstract: Deformable registration was regarded as a key task in medical image processing, whose performance directly affected the accuracy of subsequent tasks such as segmentation, classification, and prediction. However, due to the insensitivity of neural networks to high-frequency components, existing methods had difficulty capturing high-frequency information in images, which affected the fitting accuracy of the deformation field. To address these issues, a high-frequency-preserving medical image registration model based on frequency spectrum decomposition—DFRes (Decomposition in Frequency domain model for Registration), was proposed. A frequency decomposition strategy was introduced, and a dual-branch structure was adopted to process high-frequency and low-frequency information from the original image. Meanwhile, an invertible neural network structure with high-frequency preservation characteristics and a bridge-style feature fusion module with the ability to fuse high- and low-frequency information were designed. Alternating spatial-frequency information extraction modules were further incorporated to enhance the model’s ability to extract and fuse frequency-domain and spatial-domain information. Comparative experiments were conducted on the IXI, OSSAI, and Huaxi rectal cancer datasets against existing advanced models. DFRes achieved significant improvements on multiple metrics: compared to the suboptimal TransMorph model, Dice Similarity Coefficient (DSC) was increased by 2.5 percentage points, Average Surface Distance (ASD) was reduced by 1.2 percentage points, and SSIM was increased by 1.6 percentage points. Ablation experiments were also conducted to verify the effectiveness of the module design.

Key words: image registration, medical image analysis, spectrum analysis, Inverse Neural Network (INN), deep learning

摘要： 弹性配准是医学图像处理中的关键任务之一，其效果直接影响到了后续的分割、分类、预测等任务的准确性。然而由于神经网络的高频不敏感特性，现有的方法难以捕捉图像的高频信息，影响了配准场的拟合精度。为了解决这些问题，提出了基于频谱分解的高频保持医学图像配准模型——DFRes（Decomposition in Frequency domain model for Registration）。引入频谱分解的策略，采用双支结构处理原始图像中的高频信息和低频信息；同时设计了具有高频保持特性的可逆神经网络结构和具有高频低频融合能力的桥式特征融合模块，通过交替的空间频谱信息提取模块，进一步加强了模型对于频域和空域信息的提取和融合能力。在IXI、OSSAI、华西直肠数据集上与现有的先进模型进行了比较，DFRes在多个指标上取得了显著的提升，相较于次优的TransMorph模型，Dice相似系数(DSC)提高了2.5个百分点、平均表面距离（ASD）降低了1.2个百分点、SSIM↑提高了1.6个百分点。同时通过消融实验验证了模块设计的有效性。

关键词: 图像配准, 医学图像分析, 频谱分析, 可逆神经网络, 深度学习