跨尺度点匹配结合多尺度特征融合的图像配准

Image registration combining cross-scale point matching and multi-scale feature fusion

图像配准在脑部疾病的计算机辅助诊疗和远程手术等方面具有重要作用。U-Net及其变体网络广泛应用于医学图像配准领域,在配准精确度和配准时间上取得了较好效果。然而,现有的配准模型在处理复杂图像形变时,难以学习到图像中微小结构的边缘特征,且忽视了不同尺度上下文信息的关联。针对上述问题,本文提出了一种基于跨尺度点匹配结合多尺度特征融合的配准模型。首先,在模型的编码结构中引入跨尺度点匹配模块,增强对图像突出区域特征的表达以及对微小结构边缘细节特征的把握;然后,在解码结构中对多尺度特征进行融合,形成更全面的特征描述;最后,在多尺度特征融合模块中融入注意力模块,突出空间和通道的信息。在3个脑部核磁共振(Magnetic Resonance,MR)数据集上的实验结果表明,以OASIS-3数据集为例,本文方法的配准精确度相较于Affine、SyN、VoxelMorph以及CycleMorph等方法,本文方法分别提升了23.5%、12.4%、0.9%和2.1%;ASD值相较于各方法分别降低了1.074、0.434、0.043和0.076。本文提出的模型能更好地把握图像的特征信息,提升配准的精确度,对医学图像配准的发展具有重要意义。

Image registration plays an important role in computer-aided diagnosis of brain diseases and remote surgery.The U-Net and its variants have been widely used in the field of medical image registration,achieving good results in registration accuracy and time.However,existing registration models have difficulty in learning the edge features of small structures in complex image deformations and ignore the correlation of contextual information at different scales.To address these issues,a registration model is proposed based on cross-scale point matching combined with multi-scale feature fusion.Firstly,a cross-scale point matching module is introduced into encoding structure of the model to enhance the representation of prominent region features and grasp the edge details of small structure features.Then,multi-scale features are fused in the decoding structure to form a more comprehensive feature description.Finally,an attention module is integrated into the multi-scale feature fusion module to highlight spatial and channel information.The experimental results on three brain Magnetic Resonance(MR)datasets show that,taking the OASIS-3 dataset as an example,the registration accuracy has been improved by 23.5%,12.4%,0.9%,and 2.1%compared to methods such as Affine,SyN,VoxelMorph and CycleMorph,respectively.The corresponding ASD values for each method have decreased by 1.074,0.434,0.043,and 0.076.The proposed model can better grasp the feature information of images,which improves registration accuracy and has important implications for the development of medical image registration.