Participants
Newly presenting patients visiting our outpatient memory clinic between January 2011 and September 2013, aged 45 to 70 years, and with a Mini Mental State Examination (MMSE) score ≥ 20 (indicating mild dementia) were prospectively considered for inclusion. All patients underwent neurological and neuropsychological examination as part of their routine diagnostic work up. We consecutively included patients with a diagnosis of possible or probable AD or FTD. In addition, patients were included with PPA in which the underlying aetiology can be either AD or FTD. The reference standard was a nosological diagnosis of AD or FTD by consensus according to the McKhann [13] and Rascovsky [14] criteria, or AD or FTD underlying PPA [4]. Diagnosis was established either at baseline (initial visit), or after follow-up when diagnosis at baseline was uncertain, and verified independently by two experienced neurologists. Conventional structural MRI was assessed as part of the diagnostic process and simultaneously assessed for exclusion criteria, ASL-MRI was not. Patients with psychiatric or neurological disorders other than dementia were excluded. Other exclusion criteria were normal pressure hydrocephalus, Huntington’s disease, cerebral vascular disease, alcohol abuse, brain tumour, epilepsy or encephalitis.
Healthy young (18 to 40 years) and older (45 to 70 years) controls were recruited through advertisement, and older controls also from their patient peers. Data from these young participants were previously reported in a reproducibility study of ASL [15]. Both control groups were matched for gender, and older controls for age with the patients. A researcher screened all participants, who were included only when there was no history of neurological or psychiatric disease, and no contraindications for MRI. Older controls were administered the MMSE to assess global cognitive functioning.
The study was approved by the local medical ethics committee. All participants gave written informed consent.
Image acquisition
All participants were examined at 3 T (Discovery MR750 system, GE Healthcare, USA). Perfusion was measured with state-of-the-art [16] whole brain 3D pseudo-continuous ASL (p-CASL) (background-suppressed, post-labeling delay 1525 ms, labeling duration 1450 ms, echo time (TE) 10.5 ms, repetition time (TR) 4632 ms, interleaved FSE stack-of-spiral readout of 512 sampling points on eight spirals, isotropic resolution 3.3 mm in a field of view (FOV) of 240 mm, 36 axial slices, number of excitations 3, acquisition time 4.29 min). The labeling plane was positioned 9 cm below the anterior commissure-posterior commissure line. A high resolution 3-D fast spoiled gradient-echo T1-weighted (T1w) image (FOV 240 mm, TR/TE/inversion time 7.9/3.06/450 ms, ASSET factor 2, matrix 240*240, and slice thickness 1 mm, acquisition time 4.41 min) was acquired for anatomical reference.
Image data processing
The data were processed according to methods described previously [17] to obtain partial volume effect corrected CBF values from gray matter (GM) only.
Tissue segmentation
Gray matter (GM), white matter, and cerebrospinal fluid maps were obtained from the T1w image using the unified tissue segmentation method [18] of SPM8 (Statistical Parametric Mapping, London, UK). GM volumes were computed from the GM map. CBF was analyzed in GM only.
ASL post-processing
The ASL imaging dataset consisted of two images, a perfusion-weighted image (PWI) and a proton density image (PD), that were required for CBF calculation [16]. CBF maps from representative patients are shown in Figure 1. The GM map derived from the T1w image was rigidly registered with the PD image for each participant (Elastix registration software [19]). Then GM maps were transformed to ASL image space to enable partial volume (PV) correction. PV effects were corrected in PWI and PD images using local linear regression within a 3D kernel based on tissue maps [20]. The PV-corrected ASL images were quantified as CBF maps using the single-compartment model [16] as implemented by the scanner manufacturer. Finally CBF maps were transformed to T1w image space for further analysis.
ROI labeling
For each participant, regions of interest (ROIs) were defined using a multi-atlas approach. This involved the registration of 30 labeled T1w images, each containing 83 ROIs [21, 22], to the participants’ T1w images. The labels of the 30 atlas images were fused using a majority voting algorithm to obtain a final ROI labeling [23]. Registration to the participants’ nonuniformity-corrected T1w images [24] were performed with a rigid, affine, and a non-rigid B-spline transformation model consecutively. For this registration, both the participants’ and the labeled T1w images were masked using the Brain Extraction Tool [25].
Region selection
CBF was assessed per participant globally in the entire supratentorial cortex, and regionally in ten predefined cortical regions relevant for dementia, based on previously reported PET-findings in AD and FTD [26–28] (Table 1). Mean GM CBF and volumes in these regions were extracted for the left and right hemisphere separately and subsequently reported as an average of the bilateral regions. GM volumes were reported as percentage of the total intracranial volume (% ICV).
Table 1 Selected regions of interest (ROIs)
Data analysis
Gender differences across patient and control groups were examined using chi-square tests (p < 0.05). One-way analysis of variance (ANOVA) with Bonferroni correction (p < 0.05) was used to examine age and MMSE differences across AD and FTD patients and older controls; and to compare global and regional GM CBF and volume across the patient and control groups. Variation within and between groups was visualized with a boxplot.
Sensitivity and specificity of regional CBF were evaluated for both patient groups using Receiver Operating Characteristic (ROC) analysis. We examined regions known to be affected in dementia that showed significant differences between FTD or AD patients and older controls. Regions significantly different between FTD and AD patients were selected to investigate their performance in differentiating the patient groups. Diagnostic performance was expressed by areas under the curve (AUC) with 95 % confidence intervals. For the regions with the highest AUCs, optimal cut-off points were determined to discriminate between the examined groups by locating the cut-off point where the distance from maximum sensitivity and specificity was minimal. Distance was calculated for each observed cut-off point using the equation: distance = √[(1 – sensitivity)2 + (1 – specificity)2]. Based on these cut-off points, false positives (FPs) and false negatives (FNs) were determined to explore whether age, gender, MMSE or PPA variant affected misclassification.
Statistical analyses were performed in IBM SPSS Statistics, version 20.0 (New York, USA).