【作者】 何长涛；

【导师】 李宏亮；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2013， 博士

【摘要】 随着医学影像技术的快速发展，临床可用的不同模态医学图像越来越多，为了克服单一模态医学图像在局部细节信息描述上的局限性，研究者提出了多模医学图像融合技术。该技术通过提取和综合不同模态的医学图像信息，获得对病灶部位更加清晰、全面、准确、可靠的图像描述，为医生对疾病的诊断和合理治疗方案的制定提供可靠依据。多模医学图像融合是多源图像融合在医学领域的一个重要分支，作为一门多学科交叉的新兴科研领域，它不仅有着重要的科研价值同时也与人们日常生活息息相关。经过近三十年的发展，多模医学图像融合取得了不少阶段性成果，也形成了一些成熟的理论方法，但到目前为止在医学图像融合的几个关键环节上仍有许多问题有待解决。为了更好地解决这些问题，本文着眼于融合过程的几个关键环节，从“MRI图像灰度信息校正”、“源图像配准”、“多光谱与全色医学图像融合”和“显著信息保存的医学图像融合”等方面开展了医学图像预处理和融合的研究工作，主要内容和贡献如下：在医学图像预处理研究中，本文提出了基于简化PCNN模型的MRI图像灰度不均匀性校正算法和基于级联PCNN模型的医学图像配准算法。前者利用PCNN的脉冲同步发放机制进行图像偏移场估计，在保证校正效果的同时有效提高了算法的实时性。后者利用级联PCNN模型提取图像目标区域的凹点，再结合FCM聚类和坐标系分块，完成医学图像配准。在多光谱与全色医学图像融合研究中，本文提出了基于IHS和PCA的图像融合算法，为了进一步改善融合图像的光谱特性，在原算法基础上引入了视网膜激励模型，改进算法的融合图像不仅提高了图像的空间分辨率，而且较好的保持了源图像的光谱信息，有效避免了光谱扭曲现象的发生。为了突出融合过程中源图像重要信息的转移，本文提出了基于显著信息保存的多模医学图像融合算法。该算法通过对源图像局部区域的显著性加权，使隐含在图像像素中的重要信息顺利从源图像转移到融合图像。为了突出图像中不同位置（纹理、强边缘、弱边缘、角点和平滑区域等）像素的不同特征，在原算法基础上又引入了区域内像素的特征加权，改进算法的融合图像在视觉效果和信息描述上都优于原算法。为了进一步提高融合图像质量，本文提出了两种基于初始融合图像的融合算法。第一种算法是在加权平均融合基础上，结合引导滤波和像素筛选策略得到最终融合图像，该算法的融合结果存留了加权平均融合的缺陷，融合图像对比度较低且纹理等细节信息较为模糊。第二种算法是通过图像块代替的方式得到初始融合图像，在此基础上进行边缘强化等处理得到最终融合图像。比较这两种融合算法，后者的初始融合图像在对比度和图像细节信息描述两方面都优于前者，因此最终的融合结果也要优于前者。更多还原

【Abstract】 Numerous different modality medical images are available with the fastdevelopment of medical imaging technology. In order to overcome the limitations thatthe single-modal medical images only describe the local detailed information,multimodal medical image fusion technology is proposed. Through extracting andcombining information from different modal medical images, the proposed multimodalfusion technology can obtain more clear, comprehensive, accurate and reliable imagedescription of the focal areas, thus providing a reliable basis for doctors to diagnosedisease and to establish reasonable treatment methods. Multimodal medical imagefusion is an important branch of multi-source image fusion in medical field. As amulti-disciplinary and emerging research field, it not only has important scientific value,but also is closely related to people’s everyday life. After developing for nearly30years,multimodality medical image fusion has made many achievements, and formed somemature theories and methods. However, there remains many problems to be solved onthe several key steps of the medical image fusion. In order to solve these problems, theauthor focuses on several key steps of the fusion process, including "MRI image grayinhomogeneity correction","source image registration","multispectral andpanchromatic medical image fusion" and "salient information preservation of medicalimage fusion" etc., and carries out research work on the medical image preprocessingand fusion. In this paper, the main contents and contributions are summarized asfollows:As to medical image preprocessing, proposing an MRI image gray inhomogeneitycorrection algorithm based on simplified PCNN model, and medical image registrationalgorithm based on cascaded PCNN model. The former uses pulse synchronizationmechanism of PCNN to estimate the image offset field, insuring the effect of correctionand meanwhile improving the real-time performance of the algorithm. The latter usesthe cascaded PCNN model to extract the foveations in targeted image area andcombines FCM clustering and blocked coordinate system to complete medical imageregistration. In the study of multispectral and panchromatic medical image fusion, proposingthe image fusion algorithm based on IHS and PCA. In order to further improve thespectral characteristics of the fused image, the retina inspired model is introduced intothe original algorithm. The improved algorithm not only improves the spatial resolutionof the image, but also maintains the spectrum information of the source image so thatspectral distortions are avoided substantially.To highlight important information transfer from source image in the process ofimage fusion, proposing the multimodal medical images fusion algorithm based on thesaliency preservation. Through the saliency weighted on pixels in local areas of thesource image, the algorithm transfers the important information contained in the imagepixels from source image to the fusion image. In order to highlight the differentcharacteristics of pixels at different positions (texture, strong edges, weak edges, cornersand smooth areas etc.), on the basis of original algorithm, the characteristics weightedon pixels in the area is introduced. The improved algorithm performs better than theoriginal algorithm in terms of both the visual effect of fusion image and the informationdescription.In order to further improve the quality of fused image, proposing two fusionalgorithms based on the initial fused image. Based on weighted average fusion image,the first algorithm combines the guided filter and pixel screening strategy to obtain thefinal fusion image. The results of the algorithm reserve the defects of the weightedaverage fusion image, that is, the contrast is low and the texture details are relativelyobscured. The other algorithm firstly obtains the initial fused image using image blockreplacement, and then acquires the final fusion image by reinforcing the edges based onthe initial fused image. The original fusion image of the second algorithm is better thanthe first one from the viewpoint of both image contrast and details description, and sothe final fusion result of the second algorithm outperforms the first one.更多还原