Voxel-based statistical analysis and quantification of amyloid PET in the Japanese Alzheimer’s disease neuroimaging initiative (J-ADNI) multi-center study

We examined a total of 166 subjects [46 mild AD patients; 62 mild cognitive impairment (MCI) subjects; 58 normal control (NC) subjects] who underwent ¹¹C-PiB PET scans. These data were obtained from the Japanese Alzheimer’s Disease Neuroimaging Initiative (J-ADNI) database, which are available from the National Bioscience Database Center Human Database, Japan (Research ID: hum0043.v1, 2016). The J-ADNI was launched in 2007 as a public-private partnership, led by principal investigator Takeshi Iwatsubo, MD. The primary goal of J-ADNI was to test whether serial MRI, PET, other biological markers, and clinical and neuropsychological assessments can be combined to measure the progression of late MCI and mild AD in the Japanese population. The clinical inclusion criteria of neuropsychological tests are described in the article by Iwatsubo et al. [19]. This study was approved by the ethics committee of participating centers, and all subjects had signed an informed consent form for the retrospective data analysis of this kind. The present analysis was also approved by the ethics committee of IBRI, which downloaded the J-ADNI data.

Table 1 shows PET scanners and the reconstruction parameters used in the J-ADNI study. All PET sites, including their scanners, imaging protocols, and ¹¹C-PiB production, were certified by the J-ADNI PET QC core before scanning the first subject. Inter-scanner variability was minimized with the Hoffman three-dimensional (3D) phantom data by optimizing the reconstruction parameters so that the image quality and resolution was assured for each scanner [25].

All ¹¹C-PiB PET images were centrally and independently classified into positive, equivocal, and negative by three physicians specializing in nuclear neuroimaging more than 15 years. The visual interpretation criteria are described in the article by Yamane et al. [26]. For those cases where the three physicians’ interpretation did not match, they discussed these in a consensus reading session. A unified read at these sessions was determined as the official interpretation. These visual read results included additional comments by the physicians made during the consensus reading session.

Eighteen Aβ-negative PET images from 2 AD patients, 6 MCI subjects, and 10 NC subjects were used to create the normal database (NDB) of ¹¹C-PiB PET (Fig. 1). Table 2 shows mean cortical SUVR (mcSUVR) and maximum SUVR for each ¹¹C-PiB PET image, which were calculated by the method described below. The mcSUVRs in all subjects were smaller than 1.45. These ¹¹C-PiB PET images were acquired in other studies using the Discovery 690 PET/CT scanner in IBRI [27]. Imaging protocols such as the injection activity (555 MBq), uptake time (50 min), and scan duration (20 min) were the same as those used in the J-ADNI study [28]. The PET images were reconstructed using the 3D ordered subset expectation maximization (OSEM) algorithm (VUE point HD) with 4 iterations and 16 subsets. A Gaussian smoothing filter with 4 mm (n = 6) or 5 mm (n = 12) full width at half maximum (FWHM) was applied to the PET images. These parameters were determined by the phantom experiment mentioned earlier [29]. We used the PMOD ver. 3.7 and PNEURO tool for image analysis. In order to spatially normalize the PET images, we used the 3D T1-weighted MRI. The PET images were co-registered to the MRI with the normalized mutual information method implemented in the PMOD. The MRI was then spatially normalized to the Montreal Neurological Institute (MNI) T1 template [30, 31]. The co-registered PET images were also spatially normalized to the standard MNI space (matrix dimensions 91 × 109 × 91, voxel size 2 × 2 × 2 mm³) using the transformation data of the MRI spatial normalization. After that, the voxel values of the normalized PET images were divided by the mean value of the cerebellar cortex area, which was defined as the reference region in the empirical PiB-prone (EPP) ROI template (Fig. 2) [23]. The EPP-ROI template provided the target region, which characterized voxel-wise amyloid specific regions based on the actual human ¹¹C-PiB PET images. Then, average and standard deviation (SD) images were created from the NDB. We picked up some voxels as well the entire EPP-ROI and examined the normality using the Shapiro-Wilk test. As a result, we assumed that the distributions of values per voxel followed the Gaussian distribution. Figure 3 shows the average and SD images after the white matter regions were masked based on the EPP-ROI template.

Figure 4 shows the workflow of voxel-based statistical analysis for amyloid PET. Firstly, the PET images were spatially normalized to the MNI space using the PET-only adaptive template normalization method [23]. Then, voxel values were divided by the mean value of the cerebellar cortex area in the same way as the above NDB generation process. After that, the PET images were compared voxel-by-voxel against the NDB, and Z-score maps were generated within the EPP-ROI area (Figs. 2 and 3). The Z-score was calculated on each voxel using the following equation:

where x_i is the voxel value of the PET image, and \( \overline{x} \) and SD are the corresponding voxel value of the average and SD images, respectively, which are derived from the NDB (Fig. 2). After 3D Gaussian smoothing with 8 mm FWHM, Z-score values more than 2.0 were superimposed on the gray-scale PET images as a Z-score map. In addition to the Z-score map, we calculated a sum of the Z-score that were more than 2.0 within the EPP-ROI area (Z-sum).

In addition to the voxel-based approach, an ROI-based quantitative analysis was also applied to the normalized PET images to compare the results. We calculated a mean cortical standardized uptake value ratio (mcSUVR) within the EPP-ROI template [23]. This ROI template includes the frontal cortex, lateral temporal cortex, parietal cortex, posterior cingulate, precuneus, occipital cortex, and striatum as the target regions and the cerebellar cortex as the reference region. The reference region used was the same as in the voxel-based analysis.