Two-Scale Multimodal Medical Image Fusion Based on Guided Filtering and Sparse Representation
Two-Scale Multimodal Medical Image Fusion Based on Guided Filtering and Sparse Representation
复制标题
基于引导滤波和稀疏表示的两尺度多模态医学图像融合
DOI：
10.1109/access.2020.3013027
复制
发表时间：
2020
期刊：
影响因子：
3.9
通讯作者：
Wenshuai Wang
中科院分区：
文献类型：
--
作者：
Chunyang Pei;Kuangang Fan;Wenshuai Wang
医学图像融合技术主要是将不同医学图像的互补特征整合在一起，获得质量较好的单一合成图像，减少病变分析的不确定性。然而，利用多尺度变换方法从医学图像中同时提取更多的显著特征和较少的无意义细节是一项具有挑战性的任务。为了克服上述局限性，本研究提出了一种多模态医学图像的双尺度融合框架。在该框架中，使用引导滤波器将源图像分解为基层和细节层，大致分离源图像的结构信息和纹理细节两个特征。为了有效地保留大部分结构信息，采用拉普拉斯金字塔和稀疏表示相结合的规则融合基层，其中引入了基于图像补丁选择的字典构建方案，排除了源图像中无意义的补丁，增强了金字塔分解低频层的稀疏表示能力。随后使用基于引导滤波的方法合并细节层，该方法通过尽可能多的噪声滤波来增强对比度水平。对融合后的基层和细节层进行重构，生成融合图像。采用两种基本融合方案，并对9对不同模态的医学图像进行对比实验，验证了该方法的优越性。将融合后的视觉效果与客观评价结果进行比较，结果表明，该方法有效地保留了有意义的显著特征而不产生异常细节，在改进客观测量的基础上提供了更好的视觉效果。
Medical image fusion techniques primarily integrate the complementary features of different medical images to acquire a single composite image with superior quality, reducing the uncertainty of lesion analysis. However, the simultaneous extraction of more salient features and less meaningless details from medical images by using multi-scale transform methods is a challenging task. This study presents a two-scale fusion framework for multimodal medical images to overcome the aforementioned limitation. In this framework, a guided filter is used to decompose source images into the base and detail layers to roughly separate the two characteristics of source images, namely, structural information and texture details. To effectively preserve most of the structural information, the base layers are fused using the combined Laplacian pyramid and sparse representation rule, in which an image patch selection-based dictionary construction scheme is introduced to exclude the meaningless patches from the source images and enhance the sparse representation capability of the pyramid-decomposed low-frequency layer. The detail layers are subsequently merged using a guided filtering-based approach, which enhances contrast level via noise filtering as much as possible. The fused base and detail layers are reconstructed to generate the fused image. We experimentally verify the superiority of the proposed method by using two basic fusion schemes and conducting comparison experiments on nine pairs of medical images from diverse modalities. The comparison of the fused results in terms of visual effect and objective assessment demonstrates that the proposed method provides better visual effect with an improved objective measurement because it effectively preserves meaningful salient features without producing abnormal details.