Abstract:

Existing multi-modal medical image fusion methods suffer from incomplete preservation of structural information and phase features. Therefore, this study proposes a medical image fusion method based on hybrid multi-scale decomposition and local structure similarity optimization. First, the proposed hybrid multi-scale decomposition method addresses limitations of a single filter in preserving the structure and details of an image. This method combines Multi-level Decomposition Latent Low-Rank Representation (MDLatLRR) and NonSubsampled Contourlet Transform (NSCT). MDLatLRR decomposes the source image into low-rank and salient layers, and NSCT further decomposes the low-rank layer. Second, a fusion rule based on the local Laplacian energy sum applied to the base layer improves the visual effects of the fused image. A Pulse-Coupled Neural Network (PCNN) calculates the global coupling to obtain the fusion weight to fuse the detail layer. Finally, considering spatial consistency, linearly adjusted images are obtained from the initial fusion image using the weighted local structure similarity index measure to determine the correction coefficient. The accuracy of the information in the fusion image improves by correcting the initial fusion. Experiments demonstrate that compared with nine other methods such as MSMG, EMFusion, and CFL, the proposed method has a higher evaluation performance in ten objective evaluation indexes such as the normalized mutual information and spatial frequency error ratio. The average improvements in phase consistency, cosine features mutual information, and the sum of the correlations of differences are 13.89%, 19.62%, and 35.8%, respectively, compared with those in the second method. The fused image has better visual effects and richer, more accurate, and more detailed information.

Key words: medical image fusion, multi-scale decomposition, latent low-rank representation, NonSubsampled Contourlet Transform(NSCT), Pulse Coupled Neural Network(PCNN)

摘要：

现有的多模态医学图像融合方法存在结构信息以及相位特征保存不完整的问题, 为此, 提出一种基于混合多尺度分解和结构相似度优化的医学图像融合方法。首先, 针对单一滤波器在保留图像结构和细节方面的局限性, 提出一种多尺度分解潜在低秩表示(MDLatLRR)和非下采样轮廓波变换(NSCT)相结合的混合多尺度分解方法, 利用MDLatLRR分解源图像获取低秩层和显著层, 使用NSCT对低秩层做进一步分解; 其次, 在基础层上使用基于局部拉普拉斯能量和的融合规则, 使融合图像具有更好的视觉效果, 对于细节层, 通过脉冲耦合神经网络(PCNN)计算全局耦合以获得融合权重, 从而融合细节层; 最后, 考虑到空间一致性, 由初始融合图像获取线性调整图像, 利用加权局部结构相似度进行测量从而得到修正系数, 并对初始融合图像进行修正, 提高融合图像中信息的准确性。实验结果表明, 相比于MSMG、EMFusion、CFL等9种方法, 该方法在归一化互信息、空间频率误差比等10个客观评价指标上评估性能更高, 特别在相位一致性、余弦特征互信息以及差异相关和指标上, 分别比次优方法平均提升了13.89%、19.62%和35.8%, 所提方法的融合图像具有更丰富、更准确的细节信息和良好的视觉效果。

关键词: 医学图像融合, 多尺度分解, 潜在低秩表示, 非下采样轮廓波变换, 脉冲耦合神经网络