Facing privacy in neuroimaging: removing facial features degrades performance of image analysis methods.

BACKGROUND: Recent studies have created awareness that facial features can be reconstructed from high-resolution MRI. Therefore, data sharing in neuroimaging requires special attention to protect participants' privacy. Facial features removal (FFR) could alleviate these concerns. We assessed the impact of three FFR methods on subsequent automated image analysis to obtain clinically relevant outcome measurements in three clinical groups. METHODS: FFR was performed using QuickShear, FaceMasking, and Defacing. In 110 subjects of Alzheimer's Disease Neuroimaging Initiative, normalized brain volumes (NBV) were measured by SIENAX. In 70 multiple sclerosis patients of the MAGNIMS Study Group, lesion volumes (WMLV) were measured by lesion prediction algorithm in lesion segmentation toolbox. In 84 glioblastoma patients of the PICTURE Study Group, tumor volumes (GBV) were measured by BraTumIA. Failed analyses on FFR-processed images were recorded. Only cases in which all image analyses completed successfully were analyzed. Differences between outcomes obtained from FFR-processed and full images were assessed, by quantifying the intra-class correlation coefficient (ICC) for absolute agreement and by testing for systematic differences using paired t tests. RESULTS: Automated analysis methods failed in 0-19% of cases in FFR-processed images versus 0-2% of cases in full images. ICC for absolute agreement ranged from 0.312 (GBV after FaceMasking) to 0.998 (WMLV after Defacing). FaceMasking yielded higher NBV (p = 0.003) and WMLV (p ≤ 0.001). GBV was lower after QuickShear and Defacing (both p < 0.001). CONCLUSIONS: All three outcome measures were affected differently by FFR, including failure of analysis methods and both "random" variation and systematic differences. Further study is warranted to ensure high-quality neuroimaging research while protecting participants' privacy. KEY POINTS: • Protecting participants' privacy when sharing MRI data is important. • Impact of three facial features removal methods on subsequent analysis was assessed in three clinical groups. • Removing facial features degrades performance of image analysis methods.

Reproducibility of Deep Gray Matter Atrophy Rate Measurement in a Large Multicenter Dataset.

BACKGROUND AND PURPOSE: Precise in vivo measurement of deep GM volume change is a highly demanded prerequisite for an adequate evaluation of disease progression and new treatments. However, quantitative data on the reproducibility of deep GM structure volumetry are not yet available. In this paper we aim to investigate this reproducibility using a large multicenter dataset. MATERIALS AND METHODS: We have assessed the reproducibility of 2 automated segmentation software packages (FreeSurfer and the FMRIB Integrated Registration and Segmentation Tool) by quantifying the volume changes of deep GM structures by using back-to-back MR imaging scans from the Alzheimer Disease Neuroimaging Initiative's multicenter dataset. Five hundred sixty-two subjects with scans at baseline and 1 year were included. Reproducibility was investigated in the bilateral caudate nucleus, putamen, amygdala, globus pallidus, and thalamus by carrying out descriptives as well as multilevel and variance component analysis. RESULTS: Median absolute back-to-back differences varied between GM structures, ranging from 59.6-156.4 µL for volume change, and 1.26%-8.63% for percentage volume change. FreeSurfer had a better performance for the outcome of longitudinal volume change for the bilateral amygdala, putamen, left caudate nucleus (P < .005), and right thalamus (P < .001). For longitudinal percentage volume change, Freesurfer performed better for the left amygdala, bilateral caudate nucleus, and left putamen (P < .001). Smaller limits of agreement were found for FreeSurfer for both outcomes for all GM structures except the globus pallidus. Our results showed that back-to-back differences in 1-year percentage volume change were approximately 1.5-3.5 times larger than the mean measured 1-year volume change of those structures. CONCLUSIONS: Longitudinal deep GM atrophy measures should be interpreted with caution. Furthermore, deep GM atrophy measurement techniques require substantially improved reproducibility, specifically when aiming for personalized medicine.

Optimizing parameter choice for FSL-Brain Extraction Tool (BET) on 3D T1 images in multiple sclerosis.

BACKGROUND: Brain atrophy studies often use FSL-BET (Brain Extraction Tool) as the first step of image processing. Default BET does not always give satisfactory results on 3DT1 MR images, which negatively impacts atrophy measurements. Finding the right alternative BET settings can be a difficult and time-consuming task, which can introduce unwanted variability. AIM: To systematically analyze the performance of BET in images of MS patients by varying its parameters and options combinations, and quantitatively comparing its results to a manual gold standard. METHODS: Images from 159 MS patients were selected from different MAGNIMS consortium centers, and 16 different 3DT1 acquisition protocols at 1.5 T or 3T. Before running BET, one of three pre-processing pipelines was applied: (1) no pre-processing, (2) removal of neck slices, or (3) additional N3 inhomogeneity correction. Then BET was applied, systematically varying the fractional intensity threshold (the "f" parameter) and with either one of the main BET options ("B" - bias field correction and neck cleanup, "R" - robust brain center estimation, or "S" - eye and optic nerve cleanup) or none. For comparison, intracranial cavity masks were manually created for all image volumes. FSL-FAST (FMRIB's Automated Segmentation Tool) tissue-type segmentation was run on all BET output images and on the image volumes masked with the manual intracranial cavity masks (thus creating the gold-standard tissue masks). The resulting brain tissue masks were quantitatively compared to the gold standard using Dice overlap coefficient (DOC). Normalized brain volumes (NBV) were calculated with SIENAX. NBV values obtained using for SIENAX other BET settings than default were compared to gold standard NBV with the paired t-test. RESULTS: The parameter/preprocessing/options combinations resulted in 20,988 BET runs. The median DOC for default BET (f=0.5, g=0) was 0.913 (range 0.321-0.977) across all 159 native scans. For all acquisition protocols, brain extraction was substantially improved for lower values of "f" than the default value. Using native images, optimum BET performance was observed for f=0.2 with option "B", giving median DOC=0.979 (range 0.867-0.994). Using neck removal before BET, optimum BET performance was observed for f=0.1 with option "B", giving median DOC 0.983 (range 0.844-0.996). Using the above BET-options for SIENAX instead of default, the NBV values obtained from images after neck removal with f=0.1 and option "B" did not differ statistically from NBV values obtained with gold-standard. CONCLUSION: Although default BET performs reasonably well on most 3DT1 images of MS patients, the performance can be improved substantially. The removal of the neck slices, either externally or within BET, has a marked positive effect on the brain extraction quality. BET option "B" with f=0.1 after removal of the neck slices seems to work best for all acquisition protocols.

Regional DTI differences in multiple sclerosis patients.

Diffusion tensor imaging (DTI) measures have shown to be sensitive to white matter (WM) damage in multiple sclerosis (MS), not only inside focal lesions but also in user-defined regions in the so-called normal-appearing white matter (NAWM). New analysis techniques for DTI measures are now available that allow for hypothesis-free localization of damage. We performed DTI measurements of 30 MS patients selected for low focal lesion loads, and of 31 age-matched healthy controls and analyzed these using tract-based spatial statistics (TBSS). Patients were found to have a lower fractional anisotropy (FA) compared to controls in a number of brain regions, including the fornices, the left corona radiata, the inferior longitudinal fasciculus in both hemispheres, both optic radiations, and parts of the corpus callosum. In the regions of reduced FA, an increase in radial diffusivity and a less pronounced increase of axial diffusivity were found. Neurocognitive assessment showed that patients had normal visuospatial memory performance, just-normal attention, and impaired processing speed; the latter was associated with abnormal FA in the corpus callosum, an area which was relatively devoid of lesions visible on proton density-weighted images in our patients. TBSS can be useful in future studies with other MS patient samples to provide an unbiased localization of damage and generate location-specific hypotheses.

Detection of very high correlation in the alpha band between temporal regions of the human brain using MEG.

It is generally believed that alpha band (8-12 Hz) electric and magnetic activity in the area of the left and right temporal regions in the human brain are at best poorly correlated. There are no previous reports of very high alpha band correlation between left and right temporal regions by magnetoencephalography (MEG) or electroencephalography (EEG). We present whole head magnetoencephalography (MEG) results that demonstrate that, for temporal channels in the majority of healthy subjects tested, the alpha band signals are highly to very highly correlated and are antiparallel in direction. A correlation as high as -0.97 was found for a limited time in one subject. We suggest that the correlation found may be the consequence of strong direct or indirect coupling between homologue areas in left and right temporal regions rather than a common source. The correlation may provide a valuable index of loss of connectivity in the brain due to disease as well providing valuable insight to brain function and deserves further investigation.

A pathology-MRI study of the short-T2 component in formalin-fixed multiple sclerosis brain.

OBJECTIVE: To determine the pathologic basis of areas not exhibiting signal of the short-T2 component of the T2 relaxation distribution in MS, as studied in formalin-fixed brain. BACKGROUND: A myelin-specific MRI signal would be of great importance in assessing demyelination in patients with MS. Evidence indicates that the short-T2 (10 to 50 millisecond) component of the T2 relaxation distribution originates from water in myelin sheaths. The authors present two cases of MS in which the anatomic distribution of the short-T2 component was correlated with the pathologic findings in postmortem formalin-fixed brain. METHOD: One half of the formalin-fixed brain was suspended in a gelatin-albumin mixture cross-linked with glutaraldehyde, and scanned with a 32-echo MRI sequence. The brain was then cut along the center of the 5-mm slices scanned, photographed, dehydrated, and embedded in paraffin. Paraffin sections, stained with Luxol fast blue and immunocytochemically for 2',3'-cyclic nucleotide 3'-phosphohydrolase for myelin and by the Bielschowsky technique for axons, were compared with the distribution of the amplitude of the short-T2 component of the comparable image slices. RESULTS: The anatomic distribution of the short-T2 component signal corresponded to the myelin distribution. Chronic, silent MS plaques with myelin loss correlated with areas of absence of short-T2 signal. The numbers of axons within lesions were reduced, but many surviving axons were also seen in these areas of complete loss of myelin. CONCLUSION: In formalin-fixed MS brains the short-T2 component of the T2 relaxation distribution corresponds to the anatomic distribution of myelin. Chronic, silent demyelinated MS plaques show absence of the short-T2 component signal. These results support the hypothesis that the short-T2 component originates from water related to myelin.-1510