基于简化自适应脉冲耦合神经网络模型和非下采样剪切波变换的多模态图像融合算法研究

Study on the multimodal image fusion algorithm based on simplified self-adaptive PCNN model and NSST

目的:提出一种简化自适应脉冲耦合神经网络(PCNN)模型和非下采样剪切波变换(NSST)的CT和MR图像融合算法,用于PCNN模型构建融合规则。方法:采用NSST变换将已配准的CT和MR图像分解成高频和低频子图像,分别采用基于平方差之和与方向梯度特征之和的PCNN模型融合低频子图像;而高频子图像采用绝对值取大与方向梯度特征之和进行融合。采用NSST逆变换重建融合图像,选取哈佛大学脑图库8组临床实例的CT和MR图像进行仿真试验,仿真算法为本研究算法和5种基于NSST变换的融合算法,采用互信息、空间频率、标准差、边缘强度和相似度对融合图像质量进行评估。结果:定性分析与其他5种基于NSST变换的算法相比,基于图像融合算法获得的融合图像对比度、清晰度和边缘强度均最佳;本研究算法所得定量结果中互信息为2.71%~13.63%,空间频率为12.25%~73.69%,标准差为1.86%~26.33%,边缘强度为0.37%~29.90%,相似度为0.37%~67.07%,均比5种基于NSST变换的算法有不同程度提升。结论:基于简化自适应PCNN模型和NSST变换的融合算法性能优越,是一种可行的CT和MR图像融合算法。

Objective:To proposed a simplified self-adaptive image fusion algorithm of computed tomography(CT)and magnetic resonance(MR)with pulse coupled neural network(PCNN)and non-subsampled shearlet transformation(NSST),which can be used in the fusion rule of PCNN model construction.Methods:Firstly,the preregistered CT and MR images were decomposed into high frequency(HF)and low frequency(LF)sub-bands by using NSST.Then,the sum of variation and sum of directional gradients feature at each location were respectively fed to PCNN model for fusing LF sub-bands,and maximum coefficient absolute value and the sum of directional gradients feature at each location were respectively fed to PCNN model for fusing HF sub-bands.Finally,the inverse NSST was applied to reconstruct the fused image.Eight groups of clinical cases of CT and MR images were selected from Harvard brain image database were used in simulation experiments.The simulated algorithms were the proposed algorithm and other five fusion algorithms based on NSST,and the performances of fused images were assessed by mutual information(MI),spatial frequency(SF),standard deviation(STD),edge information-based index(Qab/f),and similarity(Q).Results:Comparing with other five NSST-based algorithms,qualitative analysis showed that the algorithm based on images fusion were able to attain best contrast,definition and edge intensity.In the quantitative results of this proposed algorithm,the MI was 2.71%-13.63%,and SF was 12.25%-73.69%,and STD was 1.86%-26.33%,and Qab/f was 0.37%-29.90%,and the Q was 0.37%-67.07%.And compared with the 5 algorithms based on NSST,all of them were increased in different degree.Conclusion:The performance of fused algorithm based on simplified selfadaptive PCNN model and NSST is superior,which is a feasible image fusion algorithm of CT and MR.