Impact of image acquisition systems on Alzheimer's disease-related atrophy detection
Pavel Falkovskiy1,2,3, Bénédicte Maréchal1,2,3, Tobias Kober1,2,3, Philippe Maeder1, Reto Meuli1, Jean-Philippe Thiran3, and Alexis Roche1,2,3

1Department of Radiology, University Hospital (CHUV), Lausanne, Switzerland, 2Advanced Clinical Imaging Technology (HC CMEA SUI DI BM PI), Siemens Healthcare AG, Lausanne, Switzerland, 3LTS5, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland


We investigate the potentially confounding effect of using different image acquisition systems (field strength, manufacturers) on automated Alzheimer's disease detection using standardized Alzheimer's Disease Neuroimaging Initiative (ADNI) data. Disease classifiers based on brain volumetric markers computed by FreeSurfer and the MorphoBox prototype were evaluated with and without correcting for variations in acquisition systems. While the correction overall had limited impact on Alzheimer's disease detection, it enabled significant error reduction for the classification of mildly cognitively impaired patients versus both healthy controls and Alzheimer's patients.


To study the potentially confounding effect of using different image acquisition systems on volumetric assessments in the context of Alzheimer's disease (AD), and evaluate methods to correct for such effects.


Automated brain morphometry based on T1-weighted images is increasingly used as a quantitative tool to assist brain disease diagnostics in clinical settings. Among others, its potential to detect AD-related brain atrophy at early disease stage is now well established1. There are, however, recent indications that brain volumetric measures are affected by changes in image acquisition protocols and acquisition hardware2. Our goal here is to investigate to which extent acquisition-related volume variations may hamper atrophy detection, and how they may be accounted for in practice.

Material and methods

Experiments were conducted on a standardized Alzheimer's Disease Neuroimaging Initiative (ADNI) analysis set3 of 1860 T1-weighted MR screening scans from 784 distinct subjects (age range: 54-91 years), including 220 healthy subjects, 386 patients diagnosed with mild cognitive impairment (MCI), and 178 diagnosed with AD. Images were acquired using a common protocol4 on different sites and 18 different acquisition systems from GE, Philips and Siemens. Each subject was scanned twice without repositioning at 1.5T with a raw voxel size of 1.25×1.25×1.2 mm3. Scans acquired on GE and Philips systems were subjected to in-plane sinc interpolation (zero-filled reconstruction), resulting in 0.94×0.94 mm2 pixel spacing. In addition, 148 subjects (~19%) were also scanned twice at 3T using a protocol roughly SNR-matched to the 1.5T data with voxel size 1×1×1.2 mm3. No interpolation was applied to the 3T data.

All scans were processed by the MorphoBox prototype5 and the widely used FreeSurfer package6 (version 5.3.0) to estimate the volumes normalized by total intra-cranial volume of ten brain regions known to be affected by AD: total gray matter (GM), left and right temporal lobe GM, left and right hippocampus, total cerebrospinal fluid, lateral, third and fourth ventricles. Volumes were submitted to logistic regression in order to blindly predict clinical diagnostic. Three particular strategies were investigated : one in which the logistic regressors were the ten normalized volumes plus age and gender ("basic classification"); one in which field strength, pixel spacing and their interaction were considered as additional regressors ("protocol-corrected classification"); and, finally, one in which additional regressors consisted of offsets specific to the different acquisition systems ("system-corrected classification"). For each morphometry method and classification strategy, accuracy was evaluated using leave-one-out cross-validation, and McNemar's chi square tests were performed to determine whether classifiers were significantly different.

Results and discussion

Cross-validated balanced accuracy values in three distinct binary classification scenarios (AD vs. Normal, MCI vs. Normal, MCI vs. AD) are reported in Figures 1 and 2 for MorphoBox and FreeSurfer, respectively. In the case of MorphoBox, both protocol-corrected and system-corrected classifications significantly improved the basic classification by about 2% for both MCI vs. Normal and MCI vs. AD. The effect of correction on AD vs. Normal classification was however insignificant. Also, there were no significant differences between protocol-corrected and system-corrected classifications according to McNemar's tests. FreeSurfer-based classification results showed even smaller differences between classifiers, none of which was found to be statistically significant.

These results suggest a limited impact of image acquisition systems on Alzheimer's disease detection using brain volumetry, hence justifying to some extent training disease classifiers with datasets acquired from different manufacturers and field strengths as long as they conform to a standard imaging protocol, as is the case in ADNI. The small or negligible reductions in classification errors achieved by correcting for system heterogeneity indicate that the actual effects of Alzheimer's disease on brain morphometry dominate acquisition-related variations. We note, however, that corrections led to small but significant improvements using MorphoBox in both MCI vs. Normal and AD vs. MCI classifications. Therefore, correcting for acquisition system heterogeneity may become more important as the morphometric changes to be detected are smaller.


Data collection and sharing for this study was funded by the ADNI4 (PI: Michael Weiner; NIH grant U01 AG024904).


1. Cuingnet, Rémi, et al. "Automatic classification of patients with Alzheimer's disease from structural MRI: a comparison of ten methods using the ADNI database." neuroimage 56.2 (2011): 766-781.

2. Bernal-Rusiel, Jorge L., et al. "Statistical analysis of longitudinal neuroimage data with linear mixed effects models." Neuroimage 66 (2013): 249-260.

3. Wyman, Bradley T., et al. "Standardization of analysis sets for reporting results from ADNI MRI data." Alzheimer's & Dementia 9.3 (2013): 332-337.


5. Schmitter, Daniel, et al. "An evaluation of volume-based morphometry for prediction of mild cognitive impairment and Alzheimer's disease." NeuroImage: Clinical 7 (2015): 7-17.

6. Fischl, Bruce. "FreeSurfer." Neuroimage 62.2 (2012): 774-781.


Cross-validated balanced accuracy levels for MorphoBox-based classification. Stars indicate significant differences with the basic classifier (*: p<0.05, **: p<0.01).

Cross-validated balanced accuracy levels for FreeSurfer-based classification. Effects of protocol/system correction were found non-significant.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)