MR Brain Segmentation

brain

In order to analyze anatomical structures in 3d volume data, physicians need reliable segmentation methods. One of the most frequently addressed problems is the segmentation of MR images of the human brain into gray matter and white matter.

In recent years several methods for this have been made available to the public, e.g. Freesurfer, Slicer and SPM. All of those methods rely on a registration of the analyzed image with a pre-computed anatomical atlas providing a priori information about the spatial distribution of tissue types. The anatomical knowledge incorporated into the atlas allows high quality segmentation, even when structures are hard to distinguish visually. However, registration with an atlas is time consuming and might not find a globally optimal solution or even fail completely in the case of strong anatomical anomalies. In those cases physicians have to fall back to manual segmentation on a slice by slice basis which is extremely time consuming and prone to inter-rater discrepancies.

In 2005 and 2006, we have developed a new, atlas free brain segmentation method which was shown to be able to deliver results of similar quality to those produced by the established atlas based SPM method even in the case of images without anatomical anomalies. The method builds upon the 3D-CSC, a general segmentation method for voxel images which partitions a 3d image into gray value similar, spatially connected regions.

In 2007, we introduced several improvements to 3D-CSC based segmentation of T1 weighted MR images into white matter, gray matter and non-brain which further enhance both reliability and quality while pertaining the low computational complexity of the method. As no time consuming registration step is needed, the method is not only reliable, but also very fast. Because of the complete abdication of application specific a priori knowledge, the methods developed for brain segmentation can be easily adapted to other segmentation tasks in medical imaging. We have already developed a method for segmentation of CT data into soft tissue, bones and enclosed air and are currently investigating segmentation of aortic aneurysms in CT images using similar methods. 

Our method consists of a pipeline of several algorithms which can be classified into three stages:

  • During preprocessing, artifacts degrading the quality of MR images are reduced by invocation of the 3d-Kuwahara-Nagao noise filter and a variant of the bias field filter proposed by Vovk.
  • During 3D-CSC segmentation, the image is transformed into a set of gray value similar, spatially connected 3d regions.
  • During classification of 3D-CSC segments, the brain is extracted from the input image and all 3D-CSC segments are classified into the classes white matter, gray matter and non-brain tissue.

 

pipelinepipeline

 

MR brain Segmentation Video 1 (4.2 Mb)

MR brain Segmentation Video 2 (5.6 Mb)