Brain atrophy and patch-based grading in individuals from the CIMA-Q study: a progressive continuum from subjective cognitive decline to AD

It has been proposed that individuals developing Alzheimer’s disease (AD) first experience a phase expressing subjective complaints of cognitive decline (SCD) without objective cognitive impairment. Using magnetic resonance imaging (MRI), our objective was to verify whether SNIPE probability grading, a new MRI analysis technique, would distinguish between clinical dementia stage of AD: Cognitively healthy controls without complaint (CH), SCD, mild cognitive impairment, and AD. SNIPE score in the hippocampus and entorhinal cortex was applied to anatomical T1-weighted MRI of 143 participants from the Consortium pour l’identification précoce de la maladie Alzheimer - Québec (CIMA-Q) study and compared to standard atrophy measures (volumes and cortical thicknesses). Compared to standard atrophy measures, SNIPE score appeared more sensitive to differentiate clinical AD since differences between groups reached a higher level of significance and larger effect sizes. However, no significant difference was observed between SCD and CH groups. Combining both types of measures did not improve between-group differences. Further studies using a combination of biomarkers beyond anatomical MRI might be needed to identify individuals with SCD who are on the beginning of the clinical continuum of AD.

pathophysiological biomarkers in cerebrospinal fluid (CSF) and neuroimaging, with the caveat that neuroimaging is still at an immature stage in terms of sufficiently sensitive and specific techniques.
Indeed, while comparative atrophy of brain structures (e.g. volumes, surfaces and thicknesses) as measured from magnetic resonance imaging (MRI) appears to be present in SCD 11 , differences between SCD and cognitively healthy controls without complaint (CH) are subtle 12 . Other, newer MRI analysis techniques drawing upon machine learning principles, such as probability grading via a nonlocal image patch estimator (SNIPE) 13 , might be more sensitive to these differences. In a 2015 study, Coupe et al. 14 demonstrated that hippocampal SNIPE probability grading in CH yielded a prediction accuracy up to 72.5%, seven years before conversion to probable AD; significantly higher than hippocampal volumes (58.1%).
Our objective was to compare individuals on the clinical progression continuum of AD (CH, SCD, MCI, and AD) in terms of brain morphometry using both standard atrophy measures (cortical thicknesses, volumes, and surfaces) as well as SNIPE probability grading. For the latter, we specifically focused on the hippocampal (HPC) and entorhinal cortices (EC) since AD neuropathology is initially expected in those regions 15 . We hypothesized that for both EC and HPC there would be (1) gradients of increasing biomarker severity from CH to AD, and (2) differences between SCD and CH, with SNIPE scores showing larger differences than atrophy measures. We further hypothesized that other limbic structures (e.g. medial and lateral temporal lobe) would gradually become affected as individuals experience further objective cognitive decline.

Methods
participants. The data used in this article were obtained from the Consortium pour l'identification précoce de la maladie Alzheimer -Québec (CIMA-Q) 16 , founded in 2013 with a $2,500,000 grant from the Fonds d'Innovation Pfizer -Fond de Recherche Québec -Santé sur la maladie d'Alzheimer et les maladies apparentées. The main objective is to build a cohort of participants characterized in terms of cognition, neuroimaging and clinical outcomes in order to acquire biological samples allowing (1) to establish early diagnoses of Alzheimer's disease, (2) to provide a well characterized cohort and (3) to identify new therapeutic targets. The principal investigator and director of CIMA-Q is Dr Sylvie Belleville from the Centre de recherche de l'Institut universitaire de gériatrie de Montréal, CIUSSS Centre-sud-de-l'Île-de-Montréal. CIMA-Q represent a common effort of several researchers from Québec affiliated to Université Laval, Université McGill, Université de Montréal, et Université de Sherbrooke. CIMA-Q recruited 350 cognitively healthy participants, with subjective cognitive impairment, mild cognitive impairment, or Alzheimer's disease, between 2013-2016. Volunteers were recruited from memory clinics, through advertisements posted in the community and amongst participants from the NuAge population study 17 .
Inclusion criteria were: 1. Being 65 years old and over 2. Living in the community or residence for an independent person 3. Have a score on the telephone-mini mental state examination 18 (T-MMSE) of 17 or higher 4. Understand, read and write French or English 5. Have sufficient visual and auditory acuity to be able to go through the neuropsychology tests visit. 6. For participants with mild AD, be accompanied during clinical visits. For other participants, have an informant to answer questions (in person or by phone or in writing). 7. Be willing to: answer questionnaires about one's state of health; have a physical and neuropsychological evaluation; submit to a blood test. Have an active addiction to alcohol, drugs or narcotics. 7. Have regular consumption of benzodiazepines greater than the equivalent of 1 mg per day lorazepam taken orally. 8. Have an illness/condition that is related to cognitive impairment or one that could interfere with the subject's participation in the project.
We based our analyses on the 143 participants from CIMA-Q study who underwent MRI. Ethics approval was obtained from the Institutional Review Board of the Institut universitaire de gériatrie de Montréal. All research was performed in accordance with relevant guidelines/regulations and informed consent was obtained from all participants at entry in the study. Participants were classified into four groups according to criteria by the National  Table 1.
Image acquisition and processing. T1-weighted MRI scanning was performed at five different sites (see Table 1). Scans were acquired from either Siemens Healthcare (TrioTim and Prisma Fit) or Philips Medical Systems (Achieva and Ingenia) scanners with a magnetic field strength of 3 Tesla. All scans were first visually inspected by a rater blinded to the diagnosis. Six participants were excluded due to poor quality scans (CH = 1, SCD = 1, AD = 4). The final sample included 137 participants (CH = 29, SCD = 66, eMCI = 22, lMCI = 8, AD = 12).

Atrophy state measurements.
To derive morphometry measurements for the purpose of computing atrophy states, we performed preprocessing and segmentation with the FreeSurfer (http://freesurfer.net) image analysis suite (version 5.3), using recon-all with default parameters. The technical details of these procedures are described in prior publications 23,24 . We used cortical morphometric measures generated from the Desikan-Killiany-Tourville atlas (aparc.DKTatlas40 files) and volumes from subcortical regions (aseg.stats file). The atrophy measure was obtained by adjusting morphometric measures for age, sex, estimated intracranial volume, scanner manufacturer and magnetic field strength according to normative values, expressing the final results as adjusted Z scores [25][26][27] , with negative values indicating smaller than expected values for healthy controls. Segmentation accuracy was visually examined using 19 coronal slices and 5 axial slices. Scans with apparent segmentation errors were then inspected slice by slice using FreeView (http://freesurfer.net). Cortical regions with inadequate inclusion (e.g. dura mater) or omission of approximately 100 voxels or more were excluded from statistical analyses. The HPC and EC had no segmentation error. On a total of 78 regions, the mean (±SD) number of excluded regions per participant was 1.8 (CH: 1.6 ± 2.3; SCD: 1.3 ± 2.3; eMCI: 1.6 ± 1.8; lMCI: 4.6 ± 2.9; AD: 3.5 ± 3.2). probability grading. SNIPE probability grading scores were provided under collaboration with True Positive Medical Device Inc. (TPMD, AlzMETRIX ™ ). In this patented technique, images first undergo preprocessing, which included denoising based on an optimized nonlocal means filter 28 ; correction of inhomogeneities using N3 29 ; registration to stereotaxic space using ICBM152 template linear transform (1 × 1 × 1 mm 3 voxel size) 30 using a template derived from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study database 31 ; linear intensity normalization of each subject on template intensity; and brain extraction using BEaST 32 , based on patch-based segmentation 33,34 . Then HPC and EC SNIPE grading scores 35,36 are obtained by estimating the nonlocal similarity of the subject's image to training samples of healthy aging subjects and participants with AD MRI. For each voxel in a new subject to be analyzed, the method defines a 7 × 7 × 7 voxel patch centered on the voxel. The procedure then searches the template library for similar patches. Template region labels are weighted by patch similarity, and the region label with the maximum weight is then associated with the voxel. At the same time, the template group (1 for healthy controls and 2 for AD subjects) is also weighted by the patch similarity. The resulting average weight is used as a grading value to indicate how similar this voxel is to the training AD group (for details see 36 ). The SNIPE grading score is the average of the voxels within a region (e.g., HPC or EC). HPC and EC atrophy state. Figure 1 shows atrophy states for the HPC and EC for each group. Qualitatively, atrophy results displayed a gradient of changes from CH to AD in the EC and HPC, especially for the left EC and HPC. However, only the AD group (left and right p < 0.001) had significantly lower EC and HPC volumes and lower EC thickness than CH. Furthermore, only the left EC volume and thickness were higher in CH than in SCD, with very weak effect sizes (d: 0.10 for both).
HPC and EC probability grading. SNIPE score results (Fig. 2)  composite scores. Figure 3 presents the results for the combined HPC-EC scores. AD (p < 0.001) and eMCI (p = 0.031) had significantly lowered composite atrophy score from the CH group, but not SCD (p = 0.900) and lMCI (p = 0.316). All groups had lower composite SNIPE score than the CH group (AD p < 0.001, eMCI p = 0.006, lMCI p = 0.008), except the SCD group (p = 0.900). Combining atrophy and SNIPE composite score did not improve group differences and did not yield a difference between CH and SCD groups (p = 0.900   www.nature.com/scientificreports www.nature.com/scientificreports/ (transverse temporal left thickness and right surface, left lingual thickness, and left thalamus volume) in SCD than in CH. However, none of those significant differences remained after FDR correction for multiple comparisons.
impact of co-morbidities. Anxiety and depressive symptoms had no significant influence on the relationships between cognitive complaints and SNIPE or atrophy scores (p > 0.05).

Discussion
The objective of this study was to compare individuals on the clinical progression continuum of AD on standard brain morphometry measures (volumes, surfaces and thicknesses) as well as a more recent, machine-learning derived technique for probability grading named SNIPE score.
Since hippocampal SNIPE score has been shown to predict dementia stage of AD in CH 7 years before conversion (72.5% accuracy), we expected SNIPE measures of the HPC and EC to highlight subtle differences between CH and SCD that could not be identified using usual morphometric measures. While the results confirmed our hypothesis of gradients of increasing biomarker severity from CH to clinical AD, they revealed no significant differences between CH and SCD. Notwithstanding, SNIPE measures appeared to be moderately more sensitive to dementia stage of AD since differences between groups reached a higher level of significance, and larger effect sizes between AD and CH were observed compared to standard atrophy measures of cortical thickness or structure volume.
Moreover, the combination of SNIPE score and standard atrophy measures did not highlight any group differences. Combining morphometry measures with other biomarkers should have improved highlighting very early differences on the clinical AD spectrum. However, the early preclinical nature of the disease makes the grey matter alteration very subtle and, consequently, very hard to detect.
In fact, as reported in previous studies 43,44 , cellular neurodegeneration and cognitive decline are more spatially and temporally related to neurofibrillary tangles (NFT), composed of tau proteins, than to amyloid plaques. While the choice of EC and HPC regions for SNIPE score calculation was justified as being early targets of tau accumulation, future studies should conduct probability grading in the transentorhinal cortex, given that the pattern of NFT propagation has been reported to initiate first in this region, then to the entorhinal cortex 15,45 . This may explain in part the lack of sensitivity of SNIPE score to differentiate between SCD and CH. However, at the moment no automated method seems available to segment the transentorhinal region, and segmentations thus www.nature.com/scientificreports www.nature.com/scientificreports/ have to be conducted manually. Alternatively, changes in the SCD group may be at an undetectable level, and dysregulation of long term potentiation or synaptic losses 46 could justify memory complaints, before neuronal death and quantifiable atrophy. This subtends the need and importance to follow this cohort longitudinally to assess the positive and negative predictive value of SNIPE score, as well as putative thresholds for baseline assessment that may be indicative of future decline.
While SCD, essentially defined by subjective memory decline, appears to be a promising feature to identify future cognitive decline related to AD, it is a relatively generic symptom encompassing a heterogeneous group of individuals since it is not specific to AD pathology, nor to AD symptomatology. First, a recent and large study 9 confirmed that SCD was associated with higher risks of incident of MCI (OR: 1.19) and AD (OR: 1.64), but also revealed that it increases the risk of AD pathological diagnostic (OR: 1.96). However, SCD is also associated with Lewy bodies (OR: 2.47), TDP-43 (OR: 1.33), hippocampal sclerosis (OR: 2.90), and amyloid angiopathy (OR: www.nature.com/scientificreports www.nature.com/scientificreports/ 1.46). Indeed, in addition to hippocampal sclerosis 47 , Lewy bodies 48 and TDP-43 49 are common phenomenon observed in the medial temporal lobe that have been shown to correlate with memory performance 47,48 . Secondly, SCD is unspecific to symptomatology, since it is related to multiple other symptoms, especially psychiatric symptoms. For example, it is well-documented that SCD is related to anxiety and depression 50,51 . However, in the present study, we showed that anxiety and depressive symptoms did not explain the lack of difference between CH and SCD.
Compared to standard atrophy measures, SNIPE score appeared slightly more sensitive to differentiate the dementia stage of AD, but not sufficiently to differentiate between SCD and CH.
Further studies using a combination of biomarkers beyond anatomical MRI might be needed to identify pathophysiological biomarkers of SCD.

Data Availability
The CIMA-Q dataset is publicly available at http://www.cima-q.ca.