European Radiology (2020) 30:1062–1074 https://doi.org/10.1007/s00330-019-06459-3 MAGNETIC RESONANCE Facing privacy in neuroimaging: removing facial features degrades performance of image analysis methods A. de Sitter 1 & M. Visser 1 & I. Brouwer 1 & K. S. Cover 1 & R. A. van Schijndel 1 & R. S. Eijgelaar 2 & D. M. J. Müller 3 & S. Ropele 4 & L. Kappos 5 & Á. Rovira 6 & M. Filippi 7 & C. Enzinger 8 & J. Frederiksen 9 & O. Ciccarelli 10 & C. R. G. Guttmann 11 & M. P. Wattjes 1,12 & M. G. Witte 2 & P. C. de Witt Hamer 3 & F. Barkhof 1,13 & H. Vrenken 1 & on behalf of the MAGNIMS Study Group and Alzheimer’s Disease Neuroimaging Initiative Received: 18 April 2019 / Revised: 16 August 2019 / Accepted: 13 September 2019 / Published online: 5 November 2019 # The Author(s) 2019 Abstract 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 Electronic supplementary material The online version of this article (https://doi.org/10.1007/s00330-019-06459-3) contains supplementary material, which is available to authorized users. * A. de Sitter 7 Neuroimaging Research Unit, Institute of Experimental Neurology, A.deSitter@vumc.nl Division of Neuroscience, San Raffaele Scientific Institute, UniSR, Milan, Italy 1 8 Department of Radiology and Nuclear Medicine, Amsterdam Division of Neuroradiology, Vascular and Interventional Radiology, Neuroscience Amsterdam UMC, location VUmc, Amsterdam, the Department of Radiology, Medical University of Graz, Graz, Austria Netherlands 9 Department of Neurology, Glostrup University Hospital, 2 Department of Radiotherapy, The Netherlands Cancer Institute, Copenhagen, Denmark Amsterdam, the Netherlands 10 3 UK/NIHR UCL-UCLH Biomedical Research Centre, Institute of Department of Neurosurgery, Amsterdam UMC, location VUmc, Neurology, UCL, London, UK Amsterdam, the Netherlands 11 4 Center for Neurological Imaging, Department of Radiology, Brigham Department of Neurology, Medical University of Graz, Graz, Austria and Women’s Hospital, Harvard Medical School, Boston, MA, USA 5 Department of Neurology, University Hospital, Kantonsspital, 12 Basel, Switzerland Department of Diagnostic and Interventional Neuroradiology, 6 Hannover Medical School, Hannover, Germany Unitat de Ressonància Magnètica (Servei de Radiologia), Hospital 13 universitari Vall d’Hebron, Universitat Autònoma de Barcelona, Institutes of Neurology & Healthcare Engineering, UCL, Barcelona, Spain London, UK Eur Radiol (2020) 30:1062–1074 1063 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 Brandom^ 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. Keywords Magnetic resonance imaging . Ethics . Database . Neuroimaging . Privacy Abbreviations enough to alleviate privacy concerns, since typical structural AD Alzheimer’s disease magnetic resonance imaging (MRI) provides good enough BV Unnormalized brain volume skin to air contrast and spatial resolution to perform facial EDSS Expanded Disability Status Scale recognition from a 3D-rendered version of the image, whether FFR Facial features removal by the human eye or using facial recognition software [1–5]. FLAIR Fluid-attenuated inversion recovery Therefore, in addition to identifying metadata, it has been sug- FN False negative gested that facial features should also be removed, and this has FP False positive been widely embraced [6–9]. However, it is not yet clear GBV Tumor volume of glioblastoma patients whether the removal of the facial features affects subsequent ICC Intra-class correlation coefficient measurement of quantitative indices of brain pathology. LST-LPA Lesion prediction algorithm in lesion segmenta- Therefore, the current study assessed the impact of facial tion toolbox features removal (FFR) on clinically relevant outcome mea- MCI Mild cognitive impairment surements. We selected three FFR methods that are publicly MPRAGE Magnetization-prepared rapid acquisition gradi- available, well documented, and have been used in data shar- ent echo ing initiatives [10, 11]: QuickShear [12], FaceMasking [13], MR Magnetic resonance and Defacing [14]. We assessed their effects on clinically rel- MRI Magnetic resonance imaging evant outcome measures in three different diseases: normal- MS Multiple sclerosis ized brain volumes (NBV) in Alzheimer’s disease (AD), white NBV Normalized brain volume matter lesion volumes (WMLV) in multiple sclerosis (MS), SI Dice’s similarity index and tumor volumes (GBV) in glioblastoma patients. T Tesla WML White matter lesion WMLV White matter lesion volume Materials and methods Subject Introduction Subjects in this study were obtained from three different Sharing participant image data can offer many benefits to dataset: for AD, a dataset from the ADNI study (http://adni. neuroradiological research: a better understanding of diseases loni.usc.edu/) [15]; for MS, a multicenter dataset from the can be achieved by access to larger participant populations in MAGNIMS Study Group (https://www.magnims.eu/) [16]; combined multicenter datasets; researchers without access to and for treatment-naïve glioblastoma patients, a clinical their own data on a specific disease can still contribute to its dataset from the PICTURE project collected in the understanding by using shared datasets; and methodological Amsterdam UMC, location VUmc, in Amsterdam, the improvements can be stimulated by publicly shared bench- Netherlands. Primary studies were approved by the respective mark datasets. local ethics committee for all three datasets. A summary of the However, for shared data, it is crucial to protect participants’ demographics is given in Supplementary Table 1. privacy. Image files should not contain identifying information such as name, date of birth, or any national or hospital-based Alzheimer’s disease registration numbers. Such data are often saved in metadata or even filenames of magnetic resonance (MR) images and Data used in the preparation of this article were obtained from should be removed before sharing. Unfortunately, this is not the ADNI database. The ADNI was launched in 2003 as a 1064 Eur Radiol (2020) 30:1062–1074 public-private partnership, led by Principal Investigator Facial features removal methods Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography, Three publicly available methods were selected: QuickShear other biological markers, and clinical and neuropsychological [12], FaceMasking [13], and Defacing [14] (Fig. 1). For all assessment can be combined to measure the progression of three methods, default settings were used in this study. mild cognitive impairment (MCI) and early Alzheimer’s dis- FaceMasking was applied on all MR modalities separately. ease (AD). QuickShear and Defacing can only remove facial features From the ADNI1 dataset, we selected the subset of subjects from 3D T1 images. To remove the facial features from the who had a 3-Tesla (T) magnetization-prepared rapid acquisi- other images, the full 3D T1 image of each subject was regis- tion gradient echo (MPRAGE) baseline MRI, which is a sub- tered to the other full images of the same subject, using FSL- set of the 562 subjects that are in the ADNI1 dataset [15, 17]. FLIRT (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FLIRT [20]), This subset included in total 110 (23% female) subjects with with 12 degrees of freedom before applying the FFR an average age of 75 (range 60–87) years. This dataset includ- methods. Using the resulting transformation matrices, the 3D ed 39 healthy elderly controls, 52 patients with mild cognitive T1 image without facial features was transformed to each of impairment, and 19 patients with AD. the other image spaces, and subsequently binarized and used as a mask to remove the face from the other images. Multiple sclerosis QuickShear For MS, a multicenter dataset of the MAGNIMS Study Group was previously used to study iron accumulation in deep gray Starting from a user-supplied brain mask, QuickShear [12] matter [18] and lesion segmentation software performance uses two algorithms [21, 22] to create a plane that divides [16]. The dataset consisted of 70 patients (67% female), the MRI into two parts. One part contains the facial features, scanned in six different MAGNIMS centers. On average, the and the other part contains the remainder of the head, includ- age was 34.9 (range 17–52) years. The mean disease duration ing the brain. After finding this plane, the intensity of all from onset was 7.6 (range 1–28) years and the disease severity voxels on the Bfacial features^ side of the plane is set to zero. was measured using the Expanded Disability Status Scale In this study, the brain mask was made with FSL-BET (EDSS) on the day of scanning; patients had a median (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/BET) using previously EDSS score of 2 (range 0.0–6.5) [19]. determined optimal settings [23, 24]. FaceMasking Glioblastoma FaceMasking [13] deforms the surface of the face with a filter. For glioblastoma, a total of 84 (38% female) patients were In this study, the normalized filtering method was used, which selected from a cohort treated at the Neurosurgical Center of is the recommended filter. The method first selects the bound- the Amsterdam UMC, location VUmc, Amsterdam, the ary of the skull and the face, and registers the image volume to Netherlands, in 2012 and 2013. On average, the age was an atlas with annotated face coordinates; then, the identified 61.4 (range 21–84) years. All patients had histopathologically face region of interest layer is normalized and filtered and, confirmed WHO grade IV glioblastoma. The preoperative finally, transformed back to the original image. MRI was made on average within 1 week before resection. Defacing MRI procedure Defacing [14] uses an algorithm that calculates the probability In the MS and AD datasets, all imaging was performed on 3-T of voxels being brain tissue or part of the face, based on 10 whole-body MR systems, and for imaging of the glioblastoma annotated atlases of healthy controls. Voxels that are labeled dataset on 1.5- and 3-T MR systems. The protocol for the AD as part of the face and have zero probability of being brain dataset included a 3D T1-weighted sequence, while the pro- tissue are considered to contain facial features, and their signal tocol for MS included a 3D T1-weighted sequence, as well as intensities are set to zero in order to remove the facial features. a 2D fluid-attenuated inversion recovery (FLAIR) sequence. The protocol for glioblastoma contained a 3D T1-weighted Clinical research outcome measurements post contrast–enhanced scan, 3D FLAIR, and 2D T2- weighted and non-enhanced 2D T1-weighted sequence. In For all three datasets, commonly used, previously validated Table 1, more details are listed on data acquisition of the automated methods were used to obtain clinically relevant datasets. outcome measures on the full images (i.e., images without Eur Radiol (2020) 30:1062–1074 1065 Table 1 Details on the data acquisition of the AD, MS, and glioblastoma datasets Sequence parameters Dataset Scanner brands Scanner types Field strength Sequence TR (ms) TE (ms) TI (ms) Slice (Tesla) thickness (mm) AD Siemens Not known 3 3D T1 2300–3000 2.98 853–900 1.2 GE Medical Systems Philips MS Siemens Trio 3 2D FLAIR 8000–11,000 69–136 2400–2800 3.0 Philips Achieva 3D T1 6.9–2300 2.8–298 815–900 1.0 Glioblastoma Siemens Sonata or Avanto 1.5 and 3 2D FLAIR 6500 355 2200 1.3 GE Medical Systems Signa HDxt or 3D T1* 2300–2700 4.5–5.0 950 1.0–1.5 Toshiba DISCOVERY 2D T1 520–600 8.0–12.0 5 Philips MR750 2D T2 5190–8670 93–101 5 Titan3T Panorama HFO or Achieva *Post contrast (0.2 mmol/kg) AD, Alzheimer’s disease; MS, multiple sclerosis; FLAIR, fluid-attenuated inversion recovery; TR, repetition time; TE, echo time; TI, inversion time FFR processing) as well as on all images after FFR. In the AD successfully process the data after FFR. A method was con- dataset, NBVand unnormalized brain volume (BV) were mea- sidered to have failed on a particular input image if the method sured with SIENAX [23]. In the MS dataset, WMLV was gave an error or no output. Images were not excluded if the measured by segmenting the lesions on the FLAIR images output quality was considered bad by human observers. The with the lesion prediction algorithm in the lesion segmentation percentages of images for which the FFR methods produced toolbox (LST-LPA) software [25]. In the glioblastoma dataset, output and the percentages of images for which FFR- the GBV was measured by taking the union of the segmenta- processed images could be analyzed by the automated tion of the glioblastoma necrotic core and enhancing tumor methods were calculated. generated by BraTumIA [26]. A short description of the Next, the impact of the FFR on the outcome measures was methods is provided in the supplementary data. evaluated. In order to allow a direct and fair comparison of To provide context to any observed differences between metrics between FFR methods, only the subjects for whom all results from full images and images after FFR, reproducibility three FFR methods produced output and for which both the of SIENAX, LST-LPA, and BraTumIA was assessed. This full images and all FFR-processed images could be analyzed was done by repeating the analysis on 10 native images per by the subsequent image analysis method were included. dataset, selected based on the results from the analyses of images after FFR to include in each case 5 images with large Volumetric analyses effects of FFR and 5 images with small effects of FFR. The effect of FFR on volumes was evaluated by assessing Statistical analyses changes in NBV and BV (AD dataset), WMLV (MS dataset), and GBV (glioblastoma dataset) in three different ways: in First, we investigated whether the FFR methods would suc- data distribution, variability, and systematic differences. cessfully process the data, and if the automated methods could First, to assess data distribution, histogram characteristics Fig. 1 Example 3D-rendered MRI: full (a) and after removal of facial features with QuickShear (b), FaceMasking (c), and Defacing (d). The subject gave written informed consent for using data and for displaying rendering in this figure 1066 Eur Radiol (2020) 30:1062–1074 (median, first and third quartiles, means, and standard devia- FFR method failed on this subject. This resulted in 110, 55, tion) were calculated for four images (1 full; 3 FFR-processed) and 66 subjects in the AD, MS, and GB datasets, respectively. and difference characteristics (mL and percentage difference) were calculated for 3 FFR-processed images compared with Volumetric analysis the full image, and scatter plots and Bland-Altman plots were made. Second, to assess variability in the data, whether ran- Full image results dom we analyzed intra-class correlation coefficient (ICC) for absolute agreement between volumes obtained from full and In all datasets, outcome measures obtained from the full im- FFR-processed images [27, 28] with the lower and upper ages were in the expected range and showed expected distri- bounds of 95% confidence interval [CI] in parentheses. butions. The methods showed good reproducibility, as shown Third, to assess systematic differences, two-tailed paired t tests in Table 3, on 10 subjects per dataset, ICC (lower-upper band were performed between volumes measured in full images of 95% CI) of 0.988 (0.973–0.992), 0.998 (0.992–1.000), and those obtained from each of the FFR-processed images, 0.996 (0.971–0.999), and 0.998 (0.998–1.000) for, respective- using a Bonferroni-corrected p = 0.05 as threshold for statisti- ly, the AD (NBVand BV), MS, and GB datasets. The top rows cal significance. of Tables 4 and 5 provide the measured values of NBV, BV, WMLV, and GBV for the full images. Overlap analysis (MS and glioblastoma datasets) AD dataset In MS and glioblastoma datasets, we also compared voxelwise differences between the segmentations obtained Both NBV and BV were affected by FFR processing, in terms with and without FFR, because the image analysis methods of both variability and systematic differences. In Fig. 3, an used in these datasets produce location-sensitive segmenta- example of effected SIENAX by FFR processing is given. tions of the structures of interest. The full dataset is used as Figure 4 a and b show scatter plots of NBV and BV for Bgold standard^ and is compared with each of the three FFR- FFR-processed images versus full images in the AD dataset; processed datasets separately, quantifying spatial agreement corresponding Bland-Altman plots are provided in the supple- using Dice’s similarity index (SI) [29]: mentary section. These results suggest that FFR affected NBV variability more than BV variability, which is confirmed by SI ¼ ð2  TPÞ=ð2  TP þ FP þ FNÞ the ICCs (Table 4): absolute agreement of NBV between FFR- processed images and full images ranged from 0.715 where TP, FP, and FN are, respectively, true positive, false (Defacing) to 0.896 (FaceMasking), while for BV, absolute positive, and false negative volumes. SI can range from 0 to agreement ranged from 0.933 (Defacing) to 0.982 1 and SI = 0 means no overlap and SI = 1, a perfect overlap. (QuickShear). Pairwise comparisons showed that NBV was We calculated the median and first and third quartiles of SI, typically overestimated after processing data with QuickShear FP, and FN. (median [1st and 3rd quartiles] 1.26 [− 4.40; 8.62] mL) and FaceMasking (5.91 [− 1.57; 16.38] mL), and underestimated after Defacing (− 2.17 [− 10.80; 7.81] mL). BV was typically Results underestimated after FFR, with median [1st and 3rd quartiles] volume differences of − 2.44 [− 5.11; − 1.22] mL Failure of pipelines (QuickShear), − 0.52 [− 2.47; 1.10] mL (FaceMasking), and − 7.60 [− 12.13; − 4.71] mL (Defacing). The supplementary A simplified flowchart summarizing the study steps is shown section provides means and standard deviations. in Fig. 2. An overview of the percentages of images for which the FFR methods and automated methods did not fail, i.e., MS dataset executed without error and with output, is shown for each dataset in Table 2. FFR failed only in the glioblastoma dataset, For WMLV, absolute agreement between FFR-processed im- specifically in 2% of cases for QuickShear and 1% of cases for ages and full images was high, but there were small but sig- FaceMasking. Automated method failures varied: while nificant systematic differences. In Fig. 5, an example of affect- SIENAX completed successfully on all FFR-processed im- ed lesion segmentation by FFR processing is given. Figure 4c ages (AD dataset), LST-LPA produced errors in 4% of cases shows the WMLV scatter plot; corresponding Bland-Altman for QuickShear and 19% of cases for Defacing (MS dataset); plots are provided in the supplementary section. The corre- and BraTumIA in 17% of cases with QuickShear, 2% with sponding ICCs in Table 5 are all ≥ 0.988. The median [1st and FaceMasking, and 1% with Defacing (glioblastoma dataset). 3rd quartiles] WMLV values after FFR were 2.94 [1.35; 8.12] We excluded a subject from further analyses if at least one mL (QuickShear), 3.15 [1.60; 8.15] mL (FaceMasking), and Eur Radiol (2020) 30:1062–1074 1067 Fig. 2 A flowchart summarizing the study steps. Starting with 3 datasets which are FFR-processed, followed with automated (segmentation) methods, selection subjects, and comparing outcome measurements of the FFR-processed images with native images 2.77 [1.21; 7.77] mL (Defacing). For the full image, the Fig. 6, an example of affected glioblastoma segmentation by WMLV value was 2.71 [1.32; 7.76] mL. In the case of FFR processing is given. The scatter plot in Fig. 4d, the cor- FaceMasking, the WMLV is significantly higher than the responding Bland-Altman plots in the supplementary section, WMLV of full images, Bonferroni-corrected p < 0.001. The and the ICCs in Table 5 show that FFR was affected, irrespec- supplementary section provides means and standard tive of tumor size. ICCs of 0.843, 0.312, and 0.810 between deviations. the full and FFR-processed images for QuickShear, FaceMasking, and Defacing indicate substantial effects of Glioblastoma dataset FFR. GBV were lower after FFR: differences with full image values were − 2.46 [− 7.08; − 0.54] mL (QuickShear), − 1.31 GBV appeared to be the outcome measure that is most strong- [− 7.74; 0.57] mL (FaceMasking), and Defacing − 3.28 [− ly affected by FFR, with low to poor agreement and system- 8.16; − 0.72] mL (Defacing). The supplementary section pro- atic volume underestimation after FFR by FaceMasking. In vides means and standard deviations. Table 2 Amount and percentage of images for which FFR methods completed successfully (left half of table) and for which the automated methods SIENAX, LST-LPA, and BraTumIA completed successfully (right half of table) Facial features removal Measurement AD (n = 110) MS (n = 70) Glioblastoma (n = 84) AD (n = 110) MS (n = 70) Glioblastoma (n = 84) SIENAX LST-LPA BraTumIA Full 110/110 70/70 83/84 100% 100% 99% QuickShear 110/110 70/70 82/84 110/110 67/70 68/82 100% 100% 98% 100% 96% 83% FaceMasking 110/110 70/70 83/84 110/110 70/70 81/83 100% 100% 99% 100% 100% 98% Defacing 110/100 70/70 84/84 110/110 57/70 83/84 100% 100% 100% 100% 81% 99% AD, Alzheimer’s disease; MS, multiple sclerosis 1068 Eur Radiol (2020) 30:1062–1074 Table 3 Reproducibility of automated methods on 10 subjects per images; ICC (absolute agreement (lower-upper band of 95% CI)) dataset. From left to right, the table lists median [1st and 3rd quartiles] between volumes from full and FFR-processed images; and Dice’s sim- for volumes; mean ± std for volumes; p values for the pairwise ilarity index between segmentation from first and second time processed comparison of volumes from first and second time processed full full images N = 10 Volume p value ICC SI AD NBV (L) First 1.34 [1.29; 1.38] 1.32 ± 0.10 Second 1.34 [1.30; 1.40] 1.34 ± 0.07 0.30 0.988 (0.973–0.992) AD BV (L) First 1.09 [1.00; 1.13] 1.06 ± 0.13 Second 1.08 [1.30; 1.40] 1.07 ± 0.14 0.35 0.998 (0.992–1.000) MS LV (mL) First 3.00 [1.64; 8.08] 6.39 ± 6.87 Second 3.82 [2.11; 8.35] 6.68 ± 6.74 0.16 0.996 (0.971–0.999) 0.95 [0.84–0.98] GB GBV (mL) First 39.89 [25.11; 85.92] 68.50 ± 69.66 Second 40.65 [25.28; 86.61] 69.38 ± 70.16 0.06 0.998 (0.998–1.000) 0.94 [0.89–0.96] ICC, intra-class correlation coefficient; N = amount of subjects; std, standard deviation; AD, Alzheimer’s disease; MS, multiple sclerosis; SI, Dice’s similarity index; FN, false negative; FP, false positive Overlap analysis Discussion MS dataset When sharing MRI data between research institutions, it is crucial to protect the privacy of participants. In addition to Table 5 also presents the spatial agreement of the WM removing identifying metadata from MRI, facial features lesion (WML) segmentations. The median [1st and 3rd should also be removed. The current study evaluated how quartiles] SI values of the WML segmentations from three publicly available FFR methods affect clinically relevant FFR-processed images with the corresponding segmenta- imaging outcome measures in AD, MS, and glioblastoma as tion s from ful l ima ges w ere 0 .93 [ 0.8 6; 0.9 5] derived using commonly used automated methods. Our re- (QuickShear), 0.90 [0.72; 0.94] (FaceMasking), and sults showed that the commonly used FFR methods can lead 0.86 [0.39; 0.94] (Defacing). The FN volumes of FFR- to subsequent failures of automated volumetric pipelines. p r o c e s s e d i m a g e s w e r e 0 . 2 7 [ 0 . 11 ; 0 . 7 6 ] m L Moreover, FFR can lead to substantial changes—both random (QuickShear), 0.24 [0.09; 0.66] mL (FaceMasking), and (low ICC) and systematic (significant differences)—in vol- 0.59 [0.17; 1.35] mL (Defacing). The FP volumes were umes obtained by automated methods. The observed differ- 0.24 [0.11; 0.74] mL (QuickShear), 0.58 [0.12; 2.01] mL ences in outcome measures between full images and images (FaceMasking), and 0.48 [0.20; 1.10] mL (Defacing). after FFR cannot be attributed to random variation of The supplementary section provides means and standard SIENAX, LST-LPA, or BraTumIA, because the reproducibil- deviations. ity of those methods was high. The automated methods LST-LPA for WMLV and BraTumIA for GBV failed to successfully execute on multiple Glioblastoma dataset FFR-processed images. It should be mentioned that we ap- plied the automated methods with their default settings and The SI, FN, and FP of the glioblastoma segmentation did not attempt to remedy the errors. We did, however, assess are shown in Table 5. The median [1st and 3rd quar- the failures and we suspect that the failures were related to tiles] SI values of the FFR-processed glioblastoma seg- image registration steps, because registration methods can be mentations with full image segmentations were 0.87 susceptible to (disease related) artifacts as recently highlighted [0.75; 0.92] (QuickShear), 0.86 [0.74; 0.92] by Dadar et al [30]. This recent study showed that registration (FaceMasking), and 0.86 [0.74; 0.92] (Defacing). The used in the automated methods could have problems with FN volumes were 5.33 [2.70; 9.74] mL (QuickShear), higher levels of noise and non-uniformity in images and that 4.81 [2.00; 11.75] mL (FaceMasking), and 5.94 [2.79; head size could have an effect on registration methods. It is 9.75] mL (Defacing). The FP volumes were 1.89 [0.79; conceivable that if the face is removed or deformed, the level 3.80] mL (QuickShea r), 2.91 [1.79; 4.77] mL of noise and non-uniformity could change and lead to failures. (FaceMasking), and 1.85 [0.79; 3.21] mL (Defacing). The possible importance of image registration in causing The supplementary section provides means and standard changes after FFR is further suggested by the higher variabil- deviations. ity of NBV compared with BV after FFR. To compute the Table 4 Normalized brain volume and brain volume calculated with SIENAX in the AD dataset. From left to right, the table lists median [1st and 3rd quartiles] for volumes; Bonferroni-corrected p values for the pairwise comparison of volumes from FFR-processed and full images; volume differences between the full and FFR-processed images as mL and % differences (median [1st and 3rd quartiles]); and ICC (absolute agreement (lower-upper band of 95% CI)) between volumes from full and FFR-processed images AD n = (110) Normalized brain volume Brain volume Volume (L) p value Difference ICC (lower-upper) Volume (L) p value Difference ICC (lower-upper) mL % Ml % Eur Radiol (2020) 30:1062–1074 Full 1.39 1.09 [1.34; 1.44] [1.01; 1.16] QuickShear 1.39 0.266 1.26 0.09 0.835 1.09 0.001 − 2.44 − 025 0.982 [1.35; 1.44] [− 4.40; 8.62] [− 0.32; 0.64] (0.767–0.884) [1.01; 1.16] [− 5.11; − 1.22] [− 0.48; − 0.11] (0.982–0.993) FaceMasking 1.40 0.003 5.91 0.44 0.896 1.09 0.392 − 0.52 − 0.05 0.973 [1.35; 1.46] [− 1.57; 16.38] [− 0.12; 1.19] (0.842–0.931) [1.01; 1.16] [− 2.47; 1.10] [− 0.2; 0.10] (0.960–0.981) Defacing 1.39 1.000 − 2.17 − 0.16 0.715 1.07 < 0.001 − 7.60 − 0.69 0.933 [1.35; 1.44] [− 10.80; 7.81] [− 0.80; 0.55] (0.610–0.795) [0.99; 1.14] [− 12.13; − 4.71] [− 1.14; 0.43] (0.8.74–0.961) Only the images for which FFR was successful and for which the segmentation was successful are included ICC, intra-class correlation coefficient; n, number of subjects; AD, Alzheimer’s disease Table 5 Lesion volume in the MS dataset and tumor volume in the glioblastoma dataset. From left to right, the table lists median [1st and 3rd quartiles] for volumes; Bonferroni-corrected p values for the pairwise comparison of volumes from FFR-processed and full images; volume differences between the full and FFR-processed images as mL and % differences (median [1st and 3rd quartiles]); ICC (absolute agreement (lower-upper band of 95% CI)) between volumes from full and FFR-processed images; Dice’s similarity index between segmentation from full and FFR-processed images; false negative between segmentations from full and FFR-processed images; and false positive between segmentations from full FFR-processed images MS; lesion (n = 55) Volume (mL) p value Difference ICC (lower-upper) SI FN (mL) FP (mL) mL % Full 2.71 [1.32; 7.76] QuickShear 2.94 [1.35; 8.12] 1.000 0.00 [− 0.08; 0.04] 0.01 [− 2.13; 1.08] 0.992 (0.986–0.995) 0.93 [0.86; 0.95] 0.27 [0.11; 0.76] 0.24 [0.11; 0.74] FaceMasking 3.15 [1.60; 8.15] < 0.001 0.23 [− 0.01; 1.26] 4.21 [− 0.64; 25.58] 0.988 (0.972–0.994) 0.90 [0.72; 0.94] 0.24 [0.09; 0.66] 0.58 [0.12; 2.01] Defacing 2.77 [1.21; 7.77] 0.197 0.00 [− 0.13; 0.02] − 0.12 [− 3.79; 0.83] 0.998 (0.997–0.999) 0.86 [0.39; 0.94] 0.59 [0.17; 1.35] 0.48 [0.20; 1.10] Glioblastoma; tumor (n = 66) Volume (mL) p value Difference ICC SI FN (mL) FP (mL) mL % Full 34.77 [19.54; 53.77] QuickShear 29.22 [16.74; 50.06] < 0.001 − 2.46 [− 7.08; − 0.54] − 8.11 [− 23.43; − 1.27] 0.843 (0.704–0.912) 0.87 [0.75; 0.92] 5.33 [2.70; 9.74] 1.89 [0.79; 3.80] FaceMasking 31.65 [14.80; 51.13] 1.000 − 1.31 [− 7.74; 0.57] − 5.2 [− 16.69; − 2.21] 0.312 (0.074–0.515) 0.86 [0.74; 0.92] 4.81 [2.00; 11.75] 2.19 [1.79; 4.77] Defacing 28.47 [13.41; 49.21] < 0.001 − 3.28 [− 8.16; − 0.72] − 0.70 [− 3.16; − 0.14] 0.810 (0.560–0.901) 0.86 [0.74; 0.92] 5.94 [2.79; 9.75] 1.85 [0.79; 3.21] Only the images for which FFR was successful and for which the segmentation was successful are included 1069 ICC, intra-class correlation; n, amount of subjects; MS, multiple sclerosis; SI, Dice’s similarity index; FN, false negative; FP, false positive 1070 Eur Radiol (2020) 30:1062–1074 Fig. 3 An example of SIENAX affecting the BV by FFR processing showing 3D T1 images (a), 3D T1 images with the brain tissue segmentation shown in red (1.04 L) on the full image (b), and 3D T1 images with the brain tissue segmentation shown in red (0.87 L) on the FFR- processed image with Defacing (c) NBV, SIENAX multiplies the BV (calculated in native subject thereby cause a different registration result, culminating in space) by a volumetric scaling factor obtained from a linear altered NBV values. Differences in shapes of the head and registration of the brain image to a standard brain image, ad- face between people (related to, e.g., sex or ethnicity) may ditionally using a derived skull image. FFR could affect the affect performance of standard FFR algorithms which may removal of non-brain tissue and identification of the skull, and have ramifications especially for subsequent analysis methods Fig. 4 Scatter plots of the normalized brain volume (a), brain volume (b), scatter plots have an identity line indicating perfect agreement. NBV, white matter lesion volume (c), and glioblastoma volume (d). The facial normalized brain volume; BV, brain volume; FFR, facial features removal datasets are plotted against the full scan; QuickShear, blue removal; WMLV, white matter lesion volume; GBV, tumor volume; diamond; FaceMasking, red cross; and Defacing, green plus sign. All mL, milliliter; L, liter Eur Radiol (2020) 30:1062–1074 1071 Fig. 5 An example of lesion segmentation affected by FFR processing showing 2D FLAIR image (a), 2D FLAIR image with the lesion segmentation shown in red (8.30 mL) on full image (b), and 2D FLAIR image with the lesion segmentation shown in red (15.91 mL) on FFR-processed image with Defacing (c). Dice’s similarity index between the complete 3D segmentations ob- tained from the full image and FFR-processed image was 0.48 that use the skull such as SIENAX. However, there are also processing pipelines. Moreover, it would also be helpful to cases in which the NBV was not affected by FFR, so maybe study if changing the default settings of the automated there is a cutoff on how much of the head can be removed methods improves the segmentation. without affecting the NBV measurement. In a further study, an The measured volumes of the automated methods are optimum could be determined between the amount of facial affected—both random (low ICC) and systematic (significant features that should be removed for de-identification and the differences). The random effects are mostly visible in the vol- amount that should remain for correct analyses with automat- ume change of the NBV and GBV after FFR processing and ed methods. the significant differences are visible in the volume change of Therefore, it would be interesting to study in more detail BV and GBV. The Bland-Altman plots show that the volume the effect of FFR methods on registration and other processing changes are not dependent on the measured volumes. FFR steps in a systematic way. Milchenko and Marcus [13] and affects not only the measured volumes of the automated Bischoff-Grethe et al [14] both addressed the effects of their methods but also the extent and precise spatial location of FFR method on skull stripping; however, it would be interest- the WML and the glioblastomas, as demonstrated by the over- ing to analyze the effects of these methods on other processing lap analyses. For the WML in MS, median FP and FN frac- steps as well as multiple skull stripping methods, all in the tions ranged between about 10 and 25% of the median total same dataset for an objective comparison. Next, to remedy WMLV. Similar effects were observed for glioblastoma, with those errors, analysis methods and processing steps should median FN fractions between about 15 and 20%, and median be made robust against the absence or distortion of facial FP fractions between about 6 and 9%. Both the volumetric and features. As an example, facial features could be removed spatial results indicate that the differences between full image from fixed reference images in registration steps or in refer- segmentations and FFR-processed segmentations are substan- ence templates (e.g., of tissue probabilities) in image tial. This is unexpected, especially for LV and GBV, as given Fig. 6 An example of glioblastoma segmentation affected by FFR glioblastoma segmentation shown in red (22.87 mL) on FFR-processed processing showing 3D T1 post contrast images (a), 3D T1 post image with QuickShear (c). Dice’s similarity index between the complete contrast images with the glioblastoma segmentation shown in red 3D segmentations obtained from the full image and FFR-processed image (69.99 mL) on full image (b), and 3D T1 post contrast images with the was 0.48 1072 Eur Radiol (2020) 30:1062–1074 that both the MS lesions and the glioblastoma are located database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided within the region occupied by brain tissue that should not data but did not participate in analysis or writing of this report. A complete be, and judging from our visual inspections indeed was not, listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp- affected by the FFR methods. Both in MS and glioblastoma, content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf the exact location and extent of pathological changes are of Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National importance; therefore, these artifactual post FFR segmentation Institutes of Health Grant U01 AG024904) and DOD ADNI changes should be investigated in more detail and methods (Department of Defense award number W81XWH-12-2-0012). ADNI should be devised and tested to mitigate these effects. is funded by the National Institute on Aging, the National Institute of Our results showed that the effects of FFR on current Biomedical Imaging and Bioengineering, and through generous contri- butions from the following: AbbVie, Alzheimer’s Association; methods are a common problem across domains: brain vol- Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, umes, MS lesion volumes, and glioblastoma volumes were all Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; to some degree affected by FFR. The next step would be to Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; investigate how to overcome such issues for SIENAX, LST- EuroImmun; F. Hoffmann-La Roche Ltd. and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen LPA, and BraTumIA, or in a broader sense, to study and Alzheimer Immunotherapy Research & Development, LLC.; Johnson mitigate sources of error after FFR for multiple methods & Johnson Pharmaceutical Research & Development LLC.; Lumosity; aimed at brain volume, MS lesion, or glioblastoma segmenta- Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx tion. Another option would be to consider removing the facial Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda features from fixed reference images in registration steps or in Pharmaceutical Company; and Transition Therapeutics. The Canadian reference templates (e.g., of tissue probabilities) in image pro- Institutes of Health Research is providing funds to support ADNI clinical cessing pipelines. sites in Canada. Private sector contributions are facilitated by the It should be noted that in this study, we did not test if the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research FFR methods indeed made the participant unrecognizable. and Education, and the study is coordinated by the Alzheimer’s However, we observed that the FFR methods in some cases Therapeutic Research Institute at the University of Southern California. seemed to leave parts of the face intact. We did not assess ADNI data are disseminated by the Laboratory for Neuro Imaging at the whether this made the person recognizable, because this University of Southern California. MAGNIMS Study Group; MAGNIMS (Magnetic Resonance would require a more rigorous setup outside the scope of this Imaging in MS) Study Group is a European study group of academics study. However, it would be important to establish guidelines that share the interest in MS and imaging. on how to make participants unrecognizable, specifically We want to thank Bob van Hoek that we could use the MR image of which parts of the face should be removed or otherwise proc- his head for the 3D render figure (Fig. 1). essed to ensure participants’ privacy. Moreover, for protecting Funding information The PICTURE project is supported by the program participants’ privacy, it may be important to take into account Innovative Medical Devices Initiative (IMDI) with project number 10- that reconstruction of removed or deformed facial features 10400-96-14003, which is financed by the Netherlands Organization for may be possible [31]. Scientific Research (NWO). This research is also supported by a research In conclusion, this study highlighted a new challenge to the grant from the Dutch Cancer Society (VU2014-7113). The MS Center Amsterdam is supported by the Dutch MS Research neuroimaging research community, which is to ensure high- Foundation through program grants (09-358d and 14-358e). quality neuroimaging research while protecting participants’ pri- Frederik Barkhof and Olga Ciccarelli were supported by the NIHR vacy. Our results demonstrate that facial features removal of Biomedical Research Centre at UCLH. brain MRI can lead both to failure of automated analysis methods (mostly by LST-LPA and BraTumIA) and to changes Compliance with ethical standards in volumes obtained by the analysis methods, including both Brandom^ variation (mostly by NBV and GBV) and systematic Guarantor The scientific guarantor of this publication is Hugo Vrenken, h.vrenken@vumc.nl differences (mostly by BV and GBV). Therefore, volumetric image analysis methods need to be carefully assessed and opti- Conflict of interest The authors of this manuscript declare relationships mized with regard to FFR methods, in order to ensure the reli- with the following companies: ability of clinical research outcomes while protecting partici- A. de Sitter is employed on a project sponsored by a research grant pants’ privacy in multicentric, collaborative studies. This could from Teva Pharmaceuticals (grant to H. Vrenken and F. Barkhof). M. Visser has nothing to disclose. be done by improving image registration accuracy after FFR, I. Brouwer is partly employed on a project sponsored by a research addressing in more detail the effect of FFR methods on other grant from Teva Pharmaceuticals (grant to H. Vrenken and F. Barkhof) processing steps or by developing methods that are tailored to and is partly employed on a project sponsored by a research grant from images from which facial features have been removed. Novartis (grant to H. Vrenken and F. Barkhof). K.S. Cover has nothing to disclose. R.A. van Schijndel has nothing to disclose. Acknowledgments Part of the data used in preparation of this article were R.S. Eijgelaar has nothing to disclose. obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) Eur Radiol (2020) 30:1062–1074 1073 D.M.J. Müller has nothing to disclose. Study subjects or cohorts overlap Some study subjects or cohorts have S. Ropele has nothing to disclose. been previously reported in Wyman et al (2013) and de Sitter et al (2017). L. Kappos has served in the last 24 months as international or local All patients of the paper from de Sitter et al (2017) were used and a subset principal investigator for the following drug studies: BOLD EXT., of the data published in Wyman et al (2013). The difference in this study EXPAND (Siponimod, Novartis), DECIDE, DECIDE EXT. is that we use the dataset to study the effect of facial features removal, (Daclizumab HYP, Biogen), ENDORSE (DMF, Biogen), FINGORETT, what is not done with this data yet. FTY-UMBRELLA, INFORMS, INFORMS EXT LONGTERMS. (Fingolimod, Novartis), MOMENTUM (Amiselimod, Mitsubishi) Methodology OCRELIZUMAB PHASE II EXT., OPERA, ORATORIO and exten- • prospective sions (Ocrelizumab, Roche), REFLEXION (IFN β-1a, Merck), • observational STRATA EXT, TOP (Natalizumab, Biogen), TERIFLUNOMIDE EXT, • multicenter study TERRIKIDS (Teriflunomide, Sanofi-Aventis), and ASCLEPIOS I/II (Ofatumumab, Novartis). L. Kappos is a member in the Editorial Open Access This article is distributed under the terms of the Creative Boards of the following journals: BJournal of Neurology^, BMultiple Commons Attribution 4.0 International License (http:// Sclerosis Journal^, BNeurology and Clinical NeuroscienceB, BMultiple creativecommons.org/licenses/by/4.0/), which permits unrestricted use, Sclerosis and Related Disorders^, BClinical and Translational distribution, and reproduction in any medium, provided you give Neuroscience^. appropriate credit to the original author(s) and the source, provide a link Á. Rovira serves on scientific advisory boards for Novartis, Sanofi- to the Creative Commons license, and indicate if changes were made. Genzyme, Icometrix, SyntheticMR, Bayer, Biogen, and OLEA Medical, and has received speaker honoraria from Bayer, Sanofi-Genzyme, Bracco, Merck Serono, Teva Pharmaceutical Industries Ltd., Novartis, Roche, and Biogen. M. Filippi received speaker honoraria from Biogen Idec, Novartis, References Genzyme, Sanofi-Aventis, Teva, Merck Serono, and Roche and receives research support from the Italian Ministry of Health and Fondazione 1. Gkoulalas-Divanis A, Grigorios L (2015) Medical data privacy Italiana Sclerosi Multipla. handbook. Springer C. Enzinger declares no conflict of interest. 2. Budin F, Zeng D, Ghosh A, Bullitt E (2008) Preventing facial J. Frederiksen has received no funding to support the presented work. recognition when rendering MR images of the head in three dimen- She has served on scientific advisory boards for and received funding for sions. Med Image Anal 12:229–239 travel related to these activities as well as honoraria from Biogen Idec, 3. Prior FW, Brunsden B, Hildebolt C et al (2009) Facial recognition Merck Serono, Sanofi-Aventis, Teva, Novartis, and Almirall. She has from volume-rendered magnetic resonance imaging data. IEEE received speaker honoraria from Biogen Idec, Teva, and Novartis. She Trans Inf Technol Biomed 13:5–9 has served as advisor on preclinical development for Takeda. O. Ciccarelli has received research grants from the MS Society of Great 4. Parks CL, Monson KL (2017) Automated facial recognition of Britain & Northern Ireland, National Institute for Health Research (NIHR) computed tomography-derived facial images: patient privacy impli- University College London Hospitals Biomedical Research Centre, cations. J Digit Imaging 30:204–214 EUH2020, Spinal Cord Research Foundation, and Rosetrees Trust. 5. Song X, Wang J, Wang A et al (2015) DeID - a data sharing tool for Professor Ciccarelli serves as a consultant for Novartis, Teva, and Roche neuroimaging studies. Front Neurosci 9:325 and has received an honorarium from the AAN as Associate Editor of 6. Langer SG, Shih G, Nagy P, Landman BA (2018) Collaborative Neurology and serves on the Editorial Board of Multiple Sclerosis Journal. and reproducible research: goals, challenges, and strategies. J Digit C.R.G. Guttmann has received research grants form Sanofi and the Imaging 31:275–282 National Multiple Sclerosis Society. 7. Holmes AJ, Hollinshead MO, O’Keefe TM et al (2015) Brain ge- M.P. Wattjes has received speaker and/or consultancy fees from nomics superstruct project initial data release with structural, func- Biogen, Biologix, Bayer Healthcare, Celgene, Genilac, IXICO, tional, and behavioral measures. Sci Data 2:150031 Novartis, Merck Serono, Sanofi-Genzyme, Springer Healthcare, and 8. Liew SL, Anglin JM, Banks NW et al (2018) A large, open source Roche. dataset of stroke anatomical brain images and manual lesion seg- M.G. Witte has nothing to disclose. mentations. Sci Data 5:180011 P.C. de Witt Hamer has nothing to disclose. 9. Van Essen DC, Smith SM, Barch DM et al (2013) The WU-Minn F. Barkhof has received compensation for consulting services and/or Human Connectome Project: an overview. Neuroimage 80:62–79 speaking activities from Bayer Schering Pharma, Biogen Idec, Merck 10. Kushida CA, Nichols DA, Jadrnicek R, Miller R, Walsh JK, Griffin Serono, Novartis, Genzyme, Synthon BV, Roche, Teva, Jansen research K (2012) Strategies for de-identification and anonymization of elec- and IXICO and is supported by the NIHR Biomedical Research Centre at tronic health record data for use in multicenter research studies. UCLH. Med Care 50(Suppl):S82–S101 H. Vrenken has received research grants from Novartis, Teva, Merck 11. Marcus DS, Wang TH, Parker J, Csernansky JG, Morris JC, Serono, and Pfizer; consulting fees from Merck Serono; and speaker Buckner RL (2007) Open Access Series of Imaging Studies honoraria from Novartis; all funds were paid directly to his institution. (OASIS): cross-sectional MRI data in young, middle aged, nondemented, and demented older adults. J Cogn Neurosci 19: Statistics and biometry No complex statistical methods were necessary 1498–1507 for this paper. 12. Schimke N, Hale J (2011) Quickshear defacing for neuroimages. Proceedings of the 2nd USENIX conference on Health security and Informed consent Written informed consent was obtained for primary privacy USENIX Association studies from all subjects (patients) in this study. 13. Milchenko M, Marcus D (2013) Obscuring surface anatomy in volumetric imaging data. Neuroinformatics 11:65–75 14. Bischoff-Grethe A, Ozyurt IB, Busa E et al (2007) A technique for Ethical approval Institutional Review Board approval was obtained for the deidentification of structural brain MR images. Hum Brain primary studies. Mapp 28:892–903 1074 Eur Radiol (2020) 30:1062–1074 15. Wyman BT, Harvey DJ, Crawford K et al (2013) Standardization of 25. Schmidt P (2017) Bayesian inference for structured additive regres- analysis sets for reporting results from ADNI MRI data. Alzheimers sion models for large-scale problems with applications to medical Dement 9:332–337 imaging. PhD thesis, LudwigMaximilians-Universität München. 16. de Sitter A, Steenwijk MD, Ruet A et al (2017) Performance of five Available via http://nbn-resolvingde/urn:nbn:de:bvb:19–203731 research-domain automated WM lesion segmentation methods in a 26. Bauer S, Fejes T, Slotboom J, Wiest R, Nolte LP, Reyes M (2012) multi-center MS study. Neuroimage 163:106–114 Segmentation of brain tumor images based on integrated hierarchi- 17. Cover KS, van Schijndel RA, Versteeg A et al (2016) cal classification and regularization. Proceedings of MICCAI Reproducibility of hippocampal atrophy rates measured with man- BraTS Work, pp 10–13 ual, FreeSurfer, AdaBoost, FSL/FIRST and the MAPS-HBSI 27. Weir JP (2005) Quantifying test-retest reliability using the intraclass methods in Alzheimer’s disease. Psychiatry Res Neuroimaging correlation coefficient and the SEM. J Strength Cond Res 19:231– 252:26–35 240 18. Ropele S, Kilsdonk ID, Wattjes MP et al (2014) Determinants of 28. Cicchetti DV (1994) Guidelines, criteria, and rules of thumb for iron accumulation in deep grey matter of multiple sclerosis patients. evaluating normed and standardized assessment instruments in psy- Mult Scler 20:1692–1698 chology. Psychol Assess 6:284–290 19. Kurtzke JF (1983) Rating neurologic impairment in multiple scle- 29. Dice LR (1954) Measures of the amount of ecologic association rosis: an Expanded Disability Status Scale (EDSS). Neurology 33: between species. Ecology 26:297–302 1444–1452 30. Dadar M, Fonov VS, Collins DL, Alzheimer’s Disease 20. Jenkinson M, Bannister P, Brady M, Smith S (2002) Improved Neuroimaging Initiative (2018) A comparison of publicly available optimization for the robust and accurate linear registration and mo- linear MRI stereotaxic registration techniques. Neuroimage 174: tion correction of brain images. Neuroimage 17:825–841 191–200 21. Cormen TH, Leiserson CE, Rivest RL, Stein C (2007) Introduction to algorithms, second edition. The MIT Press, Cambridge, London 31. Abramian D, Eklund A (2018) Refacing: reconstructing 22. Andrew AM (1979) Another efficient algorithm for conex hills in anonymized facial features using GANs. arXiv preprint arXiv: two dimensions. Inf Process Lett 9:216–219 1810.06455 23. Smith SM (2002) Fast robust automated brain extraction. Hum Brain Mapp 17:143–155 Publisher’s note Springer Nature remains neutral with regard to 24. Popescu V, Battaglini M, Hoogstrate WS et al (2012) Optimizing jurisdictional claims in published maps and institutional affiliations. parameter choice for FSL-Brain Extraction Tool (BET) on 3D T1 images in multiple sclerosis. Neuroimage 61:1484–1494