A new approach for improving diagnostic accuracy in Alzheimer's disease and frontal lobe dementia utilising the intrinsic properties of the SPET dataset.

Alzheimer's disease (AD) and frontal lobe dementia (FLD) show characteristic patterns of regional cerebral blood flow (rCBF). However, these patterns may overlap with those observed in the aging brain in elderly normal individuals. The aim of this study was to develop a new method for better classification and recognition of AD and FLD cases as compared with normal controls. Forty-six patients with AD, 7 patients with FLD and 34 normal controls (CTR) were included in the study. rCBF was assessed by technetium-99m hexamethylpropylene amine oxime and a three-headed single-photon emission tomography (SPET) camera. A brain atlas was used to define volumes of interest (VOIs) corresponding to the brain lobes. In addition to conventional image processing methods, based on count density/voxel, the new approach also analysed other intrinsic properties of the data by means of gradient computation steps. Hereby, five factors were assessed and tested separately: the mean count density/voxel and its histogram, the mean gradient and its histogram, and the gradient angle co-occurrence matrix. A feature vector concatenating single features was also created and tested. Preliminary feature discrimination was performed using a two-sided t-test and a K-means clustering was then used to classify the image sets into categories. Finally, five-dimensional co-occurrence matrices combining the different intrinsic properties were computed for each VOI, and their ability to recognise the group to which each individual scan belonged was investigated. For correct classification of the AD-CTR groups, the gradient histogram in the parieto-temporal lobes was the most useful single feature (accuracy 91%). FLD and CTR were better classified by the count density/voxel histogram (frontal and occipital lobes) and by the mean gradient (frontal, temporal and parietal lobes, accuracy 98%). For AD and FLD the count density/voxel histogram in the frontal, parietal and occipital lobes classified the groups with the highest accuracy (85%). The concatenated joint feature correctly classified 96% of the AD-CTR, 98% of the FLD-CTR and 94% of the AD-FLD cases. 5D co-occurrence matrices correctly recognised 98% of the AD-CTR cases, 100% of the FLD-CTR cases and 98% of the AD-FLD cases. The proposed approach classified and diagnosed AD and FLD patients with higher accuracy than conventional analytical methods used for rCBF-SPET. This was achieved by extracting from the SPET data the intrinsic information content in each of the selected VOIs.