Study On Distortion Correction For Diffusion Tensor Imaging Data And The Related Algorithms

Diffusion tensor imaging(DTI) is an innovate technique in magnetic resonance imaging (MRI), which enables visualization and quantitative characterization of structures in white matter, especially the fiber tracts. DTI measures the property of local water diffusion in different brain tissues being imaged. In the fibrous material, the water diffusion in directions perpendicular to the length of the fibers is hindered by the cell walls and the water molecules therefore diffuse more freely along axon bundles. Thus DTI can provide information on the orientation of the underlying fiber tracts. Diffusion tensors(DTs) can be reconstructed using the diffusion weighted imaging(DWI) data acquired along many gradient directions (at least6), and the dominant direction (principal direction) of the tensor is supposed to coincide with the orientation of the underlying fiber tracts. We can recover the trajectory of white matter by following fiber directions through the whole image. DTI is therefore widely used in the studies on neuroscience, psychiatric research and clinical brain diseases.DWI data are often corrupted by various kinds of artifacts due to the limitations inherited by the imaging method, which may result in inaccurate estimation of DTs. Geometric distortion and bulk motion are the most common artifacts in DWI data. The rapid switching between diffusion-weighted gradients inevitably creates residual gradients that produce eddy-current induced distortion. We can use image registration method to correct the distortion prior to the tensor reconstruction. Moreover, the eddy-current induced distortions are often accompanied by head motion, which can be corrected using an integrated image registration method.In addition to the two common artifacts, DWI data are also vulnerable to other artifacts such as noise-related artifacts, contrast artifacts or artifacts due to head motion within slices, which are termed as unstable artifacts throughout this thesis. When the unstable artifacts are serious, the image registration method cannot help to remove them and to improve the quality of the data. If we still want to save these corrupted DWI dataset and reconstruct the DTs using part of the data that are usable, identifying and eliminating the outliers in the dataset is necessary. Using an optimization method for tensor estimation that rejects outliers, the DTI data can be reconstructed reliably.As discussed, the correction for motion and distortion is based on certain proper image registration method. However, DWI data are different from scalar data, in that the DWI data contain not only magnitude information but also information of spatial orientation, which defines the orientation of the tensors calculated from the DWI data. Therefore, registration of DTI data not only needs warping the data to match the reference but also needs to properly adjust the orientation of the tensors to preserve the consistence between the tensors and the relevant anatomical structures.In this thesis, we focus on the distortion correction of DTI data and the related topics we described earlier. The main contributions of this thesis are:1. We proposed a novel correction method for eddy-current induced distortion based on one popularly used method, Iterative Cross-Correlation(ICC). Because the signal in the cerebral spinal fluid(CSF) region will hinder the estimation of ICC due to significant difference in contrast, we improved the ICC algorithm by excluding the CSF signals. We named this procedure as an "ICC-mask" algorithm. The ICC-mask method is able to provide a finer distortion estimation, which can be a good initialization for the subsequent procedure of distortion correction with high precision. We therefore used MI-based registration to correct the distortion in its following steps. Our experiments verified that the exclusion of CSF signals was also effective in the MI-based registration. This work will be presented in the first section in this thesis.2. We then proposed an integrated method for correcting eddy current induced distortion and bulk head motion effects by setting up a joint distortion model to satisfy their physical characteristics. Two different spatial transformations are combined for correcting the distortion induced by eddy current and for whole brain transformation caused by bulk head motion. The correction method is capable of dealing with both types of artifact simultaneously.3. To deal DWI datasets that contain uncorrectable, we proposed a tensor optimized method based on the automated artifact detection and removal. We firstly characterize the common artifacts of DWI images in an quantitatively way and draw the specified features to differ these artifacts from the normal data respectively. Experiments demonstrated that DTI images can be better reconstructed after the elimination of the corrupted DWI slices before the fitting procedure.4. Because tensors are associated to particular spatial orientations which are critically important information, tensors must be reoriented in a deformed space. We proposed a novel tensor reorientation method based on the reconstruction of diffusion gradient directions. Depending on the deformation field generated in the registration, we calculate new gradient directions of DWI data in a voxel-wise way. Then we estimate DTI images incorporating with these new directions. After that the reconstructed tensors are able to be reoriented. We assessed the performance of our method by applying it to correcting distortions of DWI data within the imaging space of one single subject, and also by applying it to intersubject DTI registration. The traditional strategies for tensor reorientation are usually performed in tensor space and may not applicable to cases using a different diffusion tensor model. Because our method works directly on DWI data, it is free of tensor model that the imaging procedure has been employed. For example, this method can be effectively used for registration of either conventional DTI or high angular resolution diffusion imaging(HARDI) data.