The brain cortical surface method described in this work for FDDNP demonstrated, in the living brain of human subjects, the gradual progression of cortical pathology deposition with AD progression. The FDDNP results are entirely consistent with known pathology of aggregate deposition obtained earlier with brain specimens (Braak and Braak, 1991
). Equally important, the predictable progression of FDDNP provides a ‘fingerprint’ of progressive pathology deposition and provides an opportunity to classify patients for clinical diagnosis. Thus, the FDDNP cortical binding status may help determine whether a given FDDNP brain pattern is compatible with possible AD. There are other PET probes besides FDDNP that have been used to image neuropathology deposition in AD, but they present alternative characteristics. For example, PIB does not have the same progression pattern, and various explanations have been offered in the literature. PIB’s binding pattern is different from that of FDDNP and does not follow the progressive nature demonstrated by neuropathological evaluation of autopsy specimens(Braak and Braak, 1991
). PIB accumulation as AD progresses (e.g., from controls, MCI to AD) follow a pattern that has been described as an ‘on and off’ pattern that is typically not found in pathology specimens (Kemppainen et al., 2007
; Mintun et al., 2006
). Moreover, PIB has signals in AD in most brain regions except medial temporal compared with control patients (Shin et al., 2008
), but a significant number of controls do present positive PIB binding (Mintun et al., 2006
) and also some AD subjects present negative PIB binding (Leinonen et al., 2008
). One of the explanations for the difference may be attributed to the fact that PIB does not bind to NFTs while FDDNP does (Shin et al., 2008
; Tolboom et al., 2009
The methodological analysis presented in this work demonstrated the importance of movement correction, optimization of kernel size, and partial volume correction in data quantification. In a previous work, we only used a kernel of 7 mm without optimization, did not use partial volume correction or movement correction (Braskie et al., 2008
Movement correction eliminated artifacts usually affecting regional DVR values. After movement correction of dynamic FDDNP determinations, DVR values appeared in general more left-right symmetrical on the cortical surface. Discriminant analysis based on the movement corrected images had better classification of AD subjects from the normal control and was more robust. However the movement correction method used in this study also has some limitations, such as inability to correct for intra-frame head movements. Work is on-going in our laboratory to examine these limitations to achieve further improvements, and will be addressed separately.
A parameter that is important to optimize for the cortical surface method is the kernel size of the spherical ROI used to calculate the FDDNP DVR values for each point on the cortical surface. With smaller size kernel, noise due to movement, for example, would be higher. However, a larger radius has a stronger smoothing effect that reduces variations between adjacent pixel values and thus inter-subject variability on each cortical surface point in each group. With a kernel size of 17 mm, a significant difference was observed between control and AD throughout most regions of the cortex including the motor strip, which is supposed to have low Aβ and NFT. However, this difference assessment has not considered the correlation of the neighboring cortical surface points. To choose the appropriate kernel size in this study, the following variables were analyzed: (1) comparison of the values obtained by a traditional VOI method against those values obtained by the cortical surface ROIs, and (2) Discriminant analysis on FDDNP DVR cortical surface ROIs for separation of AD group from normal controls. For the first step, the kernel size that gave ROI values closest to those of the VOI method was deemed more appropriate. With a kernel size of 9 mm (up to 11 mm), the quantitative ROI values from the two methods gave mean values that were closest for all ROIs. Using discriminant analysis, it was found that 9 mm kernel size had the best classification between AD and control groups. Combining the results from both tests, the appropriate kernel size should thus be 9 mm.
In addition, the use of MR cortical surface to map FDDNP DVR allowed correction for Partial Volume Effect (PVE), which is a common concern due to the limited spatial resolution of PET (Rousset et al., 1998
). Co-registration and co-mapping of PET images (FDG and/or FDDNP) with MRI provided the opportunity to investigate and correct for PVE. PVE and correction methods have been studied in many biomedical imaging applications (Meltzer et al., 1990
; Muller-Gartner et al., 1992
; Rousset et al., 1998
; Yang et al., 1996
). Even though there are limitations to common PVC methods due to their inability to account for true image variations within each regional mask and their sensitivity to exact boundaries between regions, application of reliable PVC is important to reveal the underlying biological changes in tissue. In this study, correction was made directly on the 3D cortical surface by creating a simulated FDDNP PET image from the MR derived cortical surface without segmenting out separate regions for gray and white matter regions. Alternatively, PVC can also be performed voxel-by-voxel on the 3D PET image first and the PVC results mapped to the cortical surface. Though the results are not expected to be much different between the two alternatives, performing PVC directly on the cortical surface is computationally less intensive due to the fact that there are less cortical points than voxels, and is less noise sensitive since the averaging by the spherical kernel is done first before multiplying with the PVCF. PVC FDDNP DVR was found to be important as it increased the separation of the DVR values between control and AD groups in the medial cortical regions, particularly the posterior cingulate and anterior cingulate regions. Thus, PVC PET increased the signal to noise ratio in the medial region. In addition, the PVC FDDNP discriminant function was more robust than that without PVC. PVC PET had 20 models that had classification/cross-validation accuracy of 94.1%/82.4% or higher, while, without PVC, there were only 14 models that had that percentage or better.
The standardization of the cortical surface maps not only facilitates examination of FDDNP PET cortical surfaces among different subject groups, but helps to evaluate the correlation of the surface maps with other behavior variables(Braskie et al., 2008
). Defining the ‘fingerprint’ pattern of FDDNP cortical binding provides a powerful tool to delineate pathology progression that appears in consistent agreement with clinical diagnosis and disease progression. Therefore, individual subjects with unknown diagnosis could be classified based on the fingerprint, similarly to what is currently done with clinical diagnosis of dementia using FDG (Silverman, 2004
; Silverman et al., 2002
). By applying regression analysis in the present work, we observed significant progression of FDDNP binding starting from the temporal and propagating to the frontal cortex as MMSE score decreased (). The regions of highest slope of the regression analysis match well those of significant Aβ and neurofibrillary tangles deposition as determined by previous autopsy studies of AD (Braak and Braak, 1991
) that include the lateral temporal, lateral frontal, anterior cingulate, medial temporal and posterior cingulate regions. Efforts to simplify and to streamline the cortical surface mapping procedure to make it more practical for routine analysis are ongoing. The model developed in this study thus can potentially be used routinely to diagnose AD suspected subjects based on their FDDNP patterns with a greater specificity than the global ROI value (sensitivity=86%, specificity=80%).