Perelman School of Medicine at the University of Pennsylvania

Section for Biomedical Image Analysis (SBIA)

Multi Atlas Skull Stripping - MASS

MASS [1] is a software package designed for robust and accurate brain extraction, applicable for both individual as well as large population studies. MASS is implemented as a Unix command-line tool. It is fully automatic and easy to use — users input an image, and MASS will output the extracted brain and the associated brain mask.

About the Algorithm

The MASS framework consists of 3 components: template selection, registration and label fusion. A general overview of the proposed method is given in the following figure.

Template Selection

The quality of a registration is directly related to the similarity between the template and the target images. Either due to differences between populations (e.g. age, disease, etc.) or changes in scanner type, technology and protocol (e.g. 1.5T to 3T), images from two different projects might be significantly different. In order to increase the template-subject similarity, and hence to improve the registration accuracy, we select a study-specific set of templates using a clustering-based approach. The same set of templates is used for processing all images in the study. In this way, we limit the work required for the preparation of the ground-truth brain masks, while using templates as similar as possible to the subjects in the study.


We have chosen a recently developed publicly available registration method DRAMMS because of its ability to meet two major challenges specific to registering raw brain MR images. The first major challenge is the large amount of intensity inhomogeneity and background noise in raw brain MR images. DRAMMS finds voxel-wise correspondences by looking at multi-scale and multi-orientation Gabor texture features around each voxel. Therefore, it is relatively robust to inhomogeneity and noise. The second major challenge in registering brain MR images with skull is the possible presence of outlier regions. Outlier regions, or missing correspondences, usually refer to regions that exist in one image but not in the other. For instance, the MR image of one subject may contain more neck regions, or may have part of superior skull missing due to different field-of-view (FOV) during MRI acquisition. DRAMMS meets this challenge using the mutual salience weighting, as it adaptively finds and relies on voxels/regions that are more likely to establish reliable correspondences across images. This way, it reduces the negative impact of outlier regions compared to other registration methods that forces matching for all voxels/regions.

Label Fusion

We adopt a spatially adaptive fusion strategy that takes into consideration the local similarities between the templates and the target image. At each voxel, a weight is assigned to each template such that a higher confidence is given to templates that are locally more similar, e.g. more easily mapped, to the target image. Our main premise here is that the Jacobian maps are good indicators of local similarities between source and target images. Large Jacobian values often correlate with large geometric differences between template and target images. It’s preferable to assign high weights to labels from masks that are locally similar to the subject image, as we have more confidence on the registration when the source and target images are more similar. Such a weighting mechanism is also efficient for making the method more robust. If the registration of one (or a few in the extreme case) template completely fails, the corresponding Jacobian map will have extreme values in most voxels. Thus the brain mask from this template will be ranked very low in general, and the template will not have any effect on the final extraction/segmentation.



Software License

The MASS software is freely available under a BSD-style open source license that is compatible with the Open Source Definition by The Open Source Initiative and contains no restrictions on use of the software. The full license text is included with the distribution package and available online.


MASS ChangeLog: Summary of changes, new features, and bug fixes.

MASS Manual: PDF version of software manual.

System Requirements

Operating System: Linux

To download visit our NITRC page for MASS



See the [BASIS_REF] on software installation for a complete list of build tools and detailed installation instructions.

Dependency Version* Description
[BASIS_REF] 2.1.2 A meta-project developed at SBIA to standardize the software development.
DRAMMS 1.4.1 A registration algorithm developed at SBIA to warp images.
AFNI   Using the version built on 2008_07_18_1710
FSL 4.1.5 A comprehensive library of analysis tools for brain imaging data
SCIKIT-LEARN 0.14.1 A python package providing several data mining and data analysis tools.
NIBABEL 1.2.0 A python package for read and write access to common medical file formats

* The versions listed are the minimum versions of the softwares for which the MASS package was tested.

Job Scheduler

If you have access to a computing cluster which has a job scheduler/queuing software (SGE, PBS etc) installed, it can be used to significantly reduce the (wall-clock) time it will take for the MASS software to produce the results. During the installation process, you can initialize the SCHEDULER variable with the particular version of your job scheduler. Currently, there are four options that are supported. You can select the one that best fits your system:

SGE - Sun Grid Engine 
PBS - Portable Batch System 
NONE - No queuing system (default) 
MISC - User defined setting

If you have a different queuing software and you select the “MISC” option, you need to modify the src/schedulerSettings/ file within the package with the appropriate options and arguments that are specific to your queuing system. You can refer to the corresponding files for SGE and PBS as examples.

  1. Extract source files:
tar -xzf mass-1.1.0-source.tar.gz
  1. Create build directory:
mkdir mass-1.1.0-build
  1. Change to build directory:
cd mass-1.1.0-build
  1. Run CMake to configure the build tree by using either one of the following commands:
cmake  -D CMAKE_INSTALL_PREFIX:STRING=/Full/path/to/install/mass/
       -D SCHEDULER:STRING=??? ../mass-1.1.0-source


ccmake ../mass-1.1.0-source
  • Press c to configure the build system and e to ignore warnings.
  • Set SCHEDULER variable with your job scheduler information.
  • Set CMAKE_INSTALL_PREFIX and other CMake variables and options.
  • Continue pressing c until the option g is available.
  • Then press g to generate the GNU Make configuration files.

After the configuration of the build tree, the software can be built using GNU Make:


After building the software, the software tests can be run using

make test

Allow 30-60 mins for the tests to finish. The last test, if the SCHEDULER variable is not set to NONE, is meant to check if submitting the jobs to the queuing system works. Please check your queue (for e.g. using qstat for SGE, PBS) to make sure that the jobs were submitted. If they are submitted, you can either delete them or wait for them to finish. As soon as these tests finish, you can proceed to the installation.


The final installation copies the built files and additional data and documentation files to the installation directory specified using the CMAKE_INSTALL_PREFIX option during the configuration of the build tree:

make install

After the successful installation, the build directory can be removed again.



  •  Christos Davatzikos
Software Development:
  •  Jimit Doshi
  • Guray Erus
  • Yangming Ou
  • Meng-Kang Hsieh
  • Bilwaj Gaonkar
  • Harsha Battapady
  • Xiao Da
  • Meng-Kang Hsieh
  • Guray Erus
  • Martin Rozycki




[1] Jimit Doshi, Guray Erus, Yangming Ou, Bilwaj Gaonkar, Christos Davatzikos, Multi-Atlas Skull-Stripping, Academic Radiology, Volume 20, Issue 12, December 2013, Pages 1566-1576, ISSN 1076-6332

[2] Yangming Ou, Aristeidis Sotiras, Nikos Paragios, Christos Davatzikos. DRAMMS: Deformable registration via attribute matching and mutual-saliency weighting. Medical Image Analysis 15(4): 622-639 (2011)

[3] Yangming Ou, Christos Davatzikos. DRAMMS: Deformable Registration via Attribute Matching and Mutual-Saliency Weighting. IPMI 2009: 50-62.


[4] Yangming Ou, Dong Hye Ye, Kilian M. Pohl, Christos Davatzikos. Validation of DRAMMS among 12 Popular Methods in Cross-Subject Cardiac MRI Registration. WBIR 2012: 209-219