Study On DTI Image Denoising

Diffusion tensor magnetic resonance imaging (DTI) is a relatively new imaging modality. In DTI, the direction, intensity and anisotropy of the water diffusion can be measured by the water diffusion attenuated magnetic resonance images of biological tissues. DTI is the unique modality that can provide such information as white matter tractography and et al., which cannot be acquired from other imaging methods such as CT and tradition MRI. Furthermore, DTI is a non-invasive technique and needn't contrast medium. All the facts above have stirred great interest toward DTI in theory research and clinical application. What makes DTI special is that diffusion tensor fiber tractography is the only method to acquire white matter structure non-invasively and this uniqueness makes it be of great clinical importance.The SNR of DTI data is relatively low, which degrades the quality of visualized tensor field data. While tracking the fibers, the noisy tensor data is disorder and irregular visually, and the fibers tracked according to the polluted tensor data are not smooth or even different from the real one, thus brings great limit to adopting DTI clinically. To perform theoretically and practically study on DTI, it's not unnecessary to restore the DTI images or tensor fields by decreasing the effect of the noise.The DTI images are of vector-valued, which makes restoring DTI data a hot topic and a difficult task. So far, there are not golden standards in smoothing DTI images. To find a suitable smoothing method for DTI data, we did much work based on related project: analyzed the noise characteristics of DTI data; proposed several smoothing methods; designed many experiments based on synthetic and real data to test the good performance of the proposed methods, the main contributions are as following:Firstly, the noise characteristic of the DTI data is analyzed. To evaluate the noise performance of the DTI data quantitatively and qualitatively, we designed many experiments to analyze the noise characteristics. All these experiments bring us a new discovery: the direction of the maximum eigenvector of the unpolluted tensor can affect the possibility that the polluted tensor be non positive. That is, different directions of the maximum eigenvector may bring different possibilities of giving non positive tensor even if the anisotropy of the tensor is of the same degree.Secondly, the simulated tempering annealing based random field model of the DTI image is studied. Diffusion process is a Gaussian process, which makes it reasonable to model DTI image as Gaussian markov random field. To make the results optimized and escape from the trap of local minimum, the simulated tempering annealing (STA) strategy is adopted. Compared with the traditional simulated annealing method, the STA method can more efficiently decrease the noise effect and make the results far away from the local minimum.Thirdly, the complex diffusion filtering model is studied and a new model to restore DTI images is proposed. Compared with the real diffusion model, the complex one is of better performance and more suitable for smoothing heave noise polluted images, thus we restored the DTI images with complex diffusion filter. We designed many experiments and got the conclusion: for the scalar valued filter the complex one is better than the real one at lower SNR. When it comes to smoothing vector valued images, the new vector valued complex diffusion model is proposed and used. It's proved that the vector valued complex diffusion filter always has better performance than the real one or scalar valued one and is ideal for smoothing DTI images.Fourthly, an affine invariant gradient (AIG) based vector-valued partial differential equation (PDE) is presented. The AIG based PDE has the good characteristics of affine invariant in addition to the same performance of traditional Euclidean gradient invariant PDE. In this dissertation, we use AIG based PDE to restore DTI images and then developed the model to a vector valued diffusion model to restore vector-valued images. To test the good performance of the proposed model, we designed many experiments with the results that the proposed model is better than the scalar model when it comes to smoothing vector valued imaged.Fifthly, Vector wavelet smoothing model is proposed. To decrease the effects of the Rician noise, we propose to consider the multi-channel wavelet based method to denoise multi-channel typed diffusion weighted (DW) images. To smooth images, we squared the images and then made wavelet transform for the squared image. It is proved that the proposed method can efficiently decrease the impact of the noise.Sixthly, Hybrid diffusion approaches are studied. The wavelet filtering method and PDE filtering method are two key ways to denoise image, while they each has shortcomings. For example, wavelet method may produce Gibbs phenomena when thresholding and PDE method may brings staircase effect. To make the two methods compensate for each other, we proposed new combined models. All the experiments about the presented smoothing strategies lead to the same conclusion: the presented smoothing strategies, which utilize anisotropic nonlinear diffusion in wavelet domain, successfully remove noise with high speed.The research jobs of this dissertation are of some theoretical and realistic significance. The research results can not only enrich the content of medical image processing but also provide efficient references for image processing in other fields. What's more, if the research results can be successfully adopted clinically, they will be helpful to diagnose and preclude Alzheimers disease and other mental diseases.