Accurate diagnosis in medical procedures has become widely attainable by the advent of the different medical imaging modalities. Among those, MRI is currently one of the most promising non-invasive diagnostic tools in medicine. The main problems in the present MRI technology are the bulkiness of the imaging apparatus, the long image acquisition period, and the sensitivity to magnetic field inhomogeneity and patient motion. The presence of these problems limits the use of MRI in a number of important applications such as heart imaging, stereotactic neurosurgery localization, and radiotherapy treatment planning. Several attempts to solve the problem of motion artifact in MRI have been reported in the literature. In spite of the success, these methods have met in their respective applications, they represent solution only to a restricted class of artifacts and cannot generally be applied to more complex types of motion such as rotational and deformable body motion. Moreover, the need for accurate a priori information about the motion may not be available in many cases. This proposal describes a novel approach called ITSI to detect and correct motion artifacts in MRI. In this approach, computer vision techniques are used to identify the prominent features in a phase-space decomposition of an acquired image which can then be used to automatically derive an accurate model for the motion. This approach is based on the observation that the edge information representing the signatures of different structures are distributed nearly uniformly in the k-space. Hence, by using this information to register the different parts of the k-space, it is possible to reconstruct artifact-free images by proper registration of the k-space segments. In the basic formulation of the technique, a windowed Fourier transform (WFT) is used to divide the acquired k-space in a number of regions, or sub-bands, that sum up the same k-space. Within each of these sub-band images, a set of features are identified and matched with the previous and following WFT images in the acquisition time line. Based on this, a set of estimates to the parameters of a deformable body motion model are derived using several independent techniques. These estimates are then fed into a trained functional link neural network in order to obtain the final estimates of the motion model parameters in an intelligent manner. Once the model has been completely determined, the sub-band images are correctly registered using a fast optimization technique based on genetic algorithms to render an artifact-free reconstruction. In this part, global features that describe the whole image are considered. That is, the model will be localized only in the k-space. Extensions of the model to include several sub-units describing local features is discussed. In this case, model localization in both the spatial domain and k-space is utilized. In principle, this should allow the applications of this technique to include local geometric distortion resulting from an unknown magnetic field inhomogeneity. Finally, the extension of the new approach in three-dimensional (3D) or multi-slice volumetric acquisitions is discussed and a number of possible clinical applications are described.
| Ahmed, Mohamed N; Yamany, Sameh M; Mohamed, Nevin et al. (2002) A modified fuzzy C-means algorithm for bias field estimation and segmentation of MRI data. IEEE Trans Med Imaging 21:193-9 |
