During recent years, there has been an increase in the number of large-scale projects that entail sharing of data across sites. NIH has initiated a data sharing policy, which requires researchers with NIH-funded grants above a certain monetary threshold to make their final research data available to other investigators. These data include human subject data acquired for basic or clinical research. With the recent enactment of HIPAA, researchers in the neuroimaging field have the added complication of removing identifying facial features from morphometric scans, in order to make the images unlike a facial photograph, without the removal or distortion of brain tissue. Most university institutional review boards require HIPAA compliance; therefore, in order to share data it must be deidentified as described by the Privacy Rule. These rules include the omission of “full facial photographic images and any comparable images,” unless informed consent is obtained from the subject to share facial images. One solution has been to apply skull-stripping to the data, as is suggested by the fMRI Data Center, a neuroimaging data repository at Dartmouth College (http://www.fmridc.org
). However, our experience has shown that automated skull-stripping algorithms are far from perfect and might remove brain tissue due to a variety of issues, including the subject population and scanner performance during data acquisition [Fennema-Notestine et al., 2006
; Smith, 2002
]. Human intervention is often required to minimize brain tissue loss, a time consuming process that is untenable when working with large datasets. Additionally, the variation in the performance of different automated skull-stripping algorithms further brings into question whether potentially vital information may be retained with one algorithm but removed by another. Therefore, as part of the BIRN initiative, we explored a possible solution to automate the deidentification of morphometric T1-weighted images that would not remove brain tissue or extracranial CSF. The defacing algorithm has been approved by the Institutional Review Boards within the BIRN consortium as sufficient for deidentification of anatomical MRI images, thus allowing for the sharing of neuroimaging data across research sites associated with the project. Our algorithm protects against casual identification of subjects. While skull-stripping takes the anonymization one step further than defacing, it may not be useful under all conditions. The loss of cranial features interferes with research combining MRI and EEG/MEG, and the technique may remove certain tissues and fluids, such as extracranial CSF, that are of interest for some fields of research.
The defacing algorithm employed herein has been conclusively shown to remove identifying facial features without disturbing brain tissue, and provides a reliable method that can be applied automatically with little human intervention required to review the outcome. The algorithm is very robust; our visual inspection of 342 datasets (some of them of poor quality) failed to find datasets in which brain tissue was removed. While the processing time is greater than that of the more widely used skull-stripping algorithms [25 min, compared with 15 s to 8 min as reported by Fennema-Notestine et al., 2006
], our experience has been that it often takes far longer to skull-strip images due to manual tuning of the parameters. The algorithm can handle a variety of data formats (DICOM, AFNI, ANALYZE, etc.), and optional parameters allow users to, for example, adjust the defacing radius (i.e., distance from the brain that is stripped), as well as the intensity values of the removed voxels.
While our primary goal was to determine that the defacing algorithm did not remove brain tissue, it is worthwhile noting that defacing did not have a detrimental effect on subsequent data processing. Overall, defacing prior to automated skull-stripping did not interfere with the chosen skull-stripping techniques. In some cases, defacing prior to skull stripping improved the quality of automated skull-stripping, such that more nonbrain tissue was removed. In one case, defacing prior to skull-stripping achieved poor results; this is not a limitation of defacing per se, but does clearly suggest one be mindful of the skull-stripping methodologies applied following defacing. Defacing likely influenced BSE’s edge-detection algorithm; selecting a different set of parameters may have improved the outcome. Because the purpose of this experiment was to determine if defacing removed brain tissue by using automated skull-stripping as a metric for analysis, manual intervention to improve results was not pursued.
It should also be made clear that the two skull-stripping algorithms used, HWA and BSE, use different methodologies to remove skull and nonbrain tissue, and hence have different outcomes whether or not defacing was applied before automated skull-stripping. These differences would influence the whole-brain analyses; HWA showed a main effect of slice when comparing automated skull stripping with and without prior defacing. Because HWA tends to be conservative to the point of leaving nonbrain tissue, including CSF, behind, whereas defacing operates primarily on the slices in which prominent facial features are present (e.g., eyes vs. cheek), the effect of slice is no doubt related to the voxels that defacing removed which HWA might retain. This was supported by the set-difference comparison. The voxels retained by skull-stripping with HWA that were removed by defacing were generally located in the regions surrounding the eye. These differences may be reduced had we chosen to manually select parameters that would give the best skull-stripping performance; however, our goal was not to review the merits of skull-stripping algorithms, nor examine their capabilities with and without human intervention.
One limitation of the proposed algorithm is that it can only be applied to T1-weighted datasets since the face atlas was constructed with T1-weighted images. However, if a T1-weighted image is acquired in addition to other image types (e.g., proton density or T2-weighted images), the mask generated during the defacing process of the T1-weighted image may be used to deface these other co-registered image types as well. Our preliminary exploration with defacing non-T1-weighted data has shown that a minimal amount of effort on the part of the researcher is required, and that visual inspection of these non-T1-weighted images indicated that brain tissue was untouched. An improvement to this algorithm would be the creation of T2-weighted and proton density atlases to enable it to function on differently weighted acquisitions.
Overall, we determined that the defacing algorithm does an effective job of removing facial features without sacrificing brain tissue. The results of defacing do not interfere with subsequent data processing, and in fact in some cases appears to make subsequent skull stripping more robust. The algorithm is fully automated and can be scripted to process large quantities of data, making it easy to deidentify data for subsequent sharing in multisite projects.