The inexhaustible eye: Automated methods for the objective analysis of brain spect images

The analysis of nuclear medicine images of the brain may be improved by automated methods of image alignment and quantitation by comparison to normal databases. The focus of this work was the development and evaluation of algorithms for single-photon emission computed tomography (SPECT) of the brain to improve the reproducibility of interpretation and attain a high standard of quantitative accuracy.;The first section of this research describes algorithms developed for dopamine neuroreceptor images, which are utilized to distinguish idiopathic Parkinson's disease from other diseases that produce similar motor dysfunction. A novel algorithm was developed to align (i.e. register) the dopamine receptor images. This method was used to create a template that defines mean and standard deviation values for normal subject images. Patient images were then aligned with the template for testing the differential diagnostic efficacy. Both regional and voxel-based semi-quantitative techniques were examined, as well as an extensive set of parameters affecting quantitation. The optimal quantitation technique was determined by extensive automated tests, and receiver-operator characteristic curve analysis.;The second section of this research evaluated several linear registration techniques for the alignment of brain SPECT perfusion images utilized for the diagnosis of Alzheimer's disease. The early diagnosis of Alzheimer's disease is supported by the detection of hypoperfused lesions in the temporal-parietal regions of the patient's brain. This analysis may be improved by registering the brain image with a normal template image representing the mean of a database of normal subjects. The registration accuracy is limited, however, by the presence of hypoperfused lesions that act as structural noise. Three linear registration algorithms were tested with patient data, and with normal images that had simulated defects. The study determined that normalized mutual information was the least affected by the hypoperfused lesions.;Having developed a linear technique to determine the rigid-body transformation to align images, it is possible to calculate non-linear adjustments to the brain shape for improved registration. Thus, in the third section a non-linear, constrained registration technique was developed for brain SPECT. The accuracy was examined with patient data, and the results were compared with linear registration and a polynomial non-linear registration method. Furthermore, simulated lesions were created on normal images to examine whether these introduced significant errors in registration and quantitation. The novel warping algorithm was demonstrated to improve accuracy and possess the constraints necessary to minimize errors when there were severe lesions.;These studies constitute steps in the journey toward objective, reliable and accurate analysis of diagnostic brain images.