| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 91 === Tc-99m HMPAO is a typical tracer used in the analysis of cerebral blood flow. Single photon emission computed tomography (SPECT) brain imaging utilizing Tc-99m HMPAO is thus a popular method to assess brain function.
In this thesis, we present an image analysis system for SPECT images that performs a series of image processing procedures including 3D image registration, brain extraction, and gray-level normalization, which map all the 3D brain data to the same space for further analysis. Afterward, the aligned images undertake a standard statistical analysis, the paired t test, to detect the areas that have significant deviations between the two images. The statistical t map is then represented with a color plot for each brain slice to indicate the activation foci. The resulting maps are found very helpful to doctors for the lesion evaluation and localization in the clinical diagnosis of schizophrenic patients.
Besides, a study on image registration by maximization of mutual information is also given. We address the concept of joint histogram, which is used to derive the probability distribution and thus compute the mutual information (MI) value. A well-known interpolation artifact problem caused from the partial volume (PV) interpolation is discussed and reviewed here. Hence, we propose a weighting method, which adds the local joint probability term to the original joint probability, to eliminate the artifacts effectively. Under the consideration to computational cost, we also construct a multiresolution hierarchy to reduce the search time.
The registration results are evaluated using the designed phantom and the acquired clinical SPECT images. We show that the subvoxel accuracy is achieved and the consistency is also maintained. It is also proved to be superior to the PV interpolation when the proposed weighting method is used.
|
|---|
