Difference between revisions of "Automatic de-Skulling"

Revision as of 19:16, 14 January 2014

Automatic de-Skulling provides tools that help anonymizing the 3D head image, so that the person cannot be recognized when displayed in a 2D-slice or 3D-volume view. This can be done as a progression from basic face-anonymizing, to de-skulling, to complete brain segmentation and extraction. Fully automatic brain segmentation and extraction is an area of on-going research.

Introduction

Full-scale de-skulling requires removal of all skin, muscle, skull, eyes, exterior blood vessels and nerve tissue. Brain segmentation goes further to define the exact components (white matter, gray matter, cerebellum, brain stem) to include in the final image. Although brain segmentation is not necessary to de-identify a face, several components of brain segmentation algorithms can be used as strategies for removing the identifying features of a face.

Background

In this algorithm we decided to err on the side of leaving too much data rather than removing data that was part of the brain. One of the key steps in the algorithm is segmenting the white-matter of the brain,using a technique described in the Skull-Striping algorithm [Sadananthan]. Given a good white-matter segmentation, the algorithm can reliability find a good skull segmentation without removing portions of the surface of the brain. A modified version of the Brain Extraction Tool (AlgorithmBrainExtractor) in MIPAV is then applied to further refine the brain surface.

The algorithm was tested on nearly two dozen images and on several examples of the different MRI modalities. The algorithm produced good segmentation for all the images except CT images. For the CT images the algorithm provides the option of using an atlas, or reference, image to generate the segmentation. The atlas image is first registered to the target image, then the algorithm is run on the atlas image and the result is applied to the target image. This worked well for the CT images tested, using a T1-weighted MRI image as the atlas.

The goal for this algorithm is to produce fully anonymous results for

• a variety of image types and modalities,
• all sage ranges from children to elderly,
• both healthy and non-healthy patient images.

Input Image Challenges

A significant challenge in creating the algorithm is the variety of input images. Much of the reference work in brain segmentation is done on T1 or T2 images only. Examples of different image types are listed and shown below.

• Computed Tomography (CT)
• Magnetic Resonance Images (MRI) including but not limited to:
• Diffusion-weighted (DWI), Diffusion-tensor (DTI), and High-angular-resolution (HARDI)
• Fluid attenuated inversion recovery MRI (FLAIR)
• Functional MRI (fMRI)
• Proton density MRI (PD)
• T1-weighted MRI
• T2-weighted MRI

Algorithm Implementation

The algorithm performs the following steps:

1. It calculates the input image histogram.
2. It finds all pixel values that represent 2% and 98% of the histogram values from Step 1, and then clamps the input image data to that range.
3. It sets i) the cerebrospinal fluid (CSF) threshold minimum value to 30 % of the clamped image range and ii) the CSF threshold maximum value to 70% of the clamped image range from Step 2.
4. It calculates the brain center of gravity (COG) as the weighted sum of all voxels between the CSF minimum and maximum range using the voxel position scaled by intensity value.
5. It estimates the white matter intensity threshold by i) segmenting the image into 5x5x5 voxel cubes and ii) finding the cube with the highest mean intensity and lowest variance. The mean intensity of the selected cube is used to set the white matter threshold value.
6. It performs a white-matter segmentation of the brain by using a flood-fill algorithm with the white matter threshold seed point from Step 5 as the starting point. The flood-fill algorithm is limited by the CSF threshold values as well as a radius centered around the initial seed point.
7. It takes the convex-hull of the white matter segmentation and passes this mesh into a modified version of the AlgorithmBrainExtractor.
8. The output of the AlgorithmBrainExtractor is a segmented mesh representing the surface of the brain. To err on the side of not removing too much data, the convex-hull of the brain surface is generated.
9. The final mask representing the data outside the brain is generated using all voxels outside the convex-hull of the brain surface.
10. If the user chooses to segment only the face region instead of the entire skull region the final step produces the face segmentation from the convex-hull of the brain surface.

Steps 1-4

Are derived from the image pre-processing steps described in the [Extract Brain: Extract Brain Surface (BSE)|Brain Extraction Tool] (BET and BET2) [Smith]. The preprocessed steps generate the center of gravity (COG) and cerebrospinal fluid (CSF) thresholds, which are used to find the initial white-matter seed point. Potential white matter seed points are limited to a radius around the COG. The minimum variance is selected from potential seed points with a maximum mean-intensity greater than the CSF threshold. These steps prevent the algorithm from finding a seed point in the background, where the variance may be zero but where the intensity is lower than the CSF threshold.