EMSegment

• 

  Projects List

Key Investigators

• Sylvain Jaume, Polina Golland (MIT Computer Science and Artificial Intelligence Laboratory)
• Nicolas Rannou, Steve Pieper, Ron Kikinis (Harvard Medical School, Brigham and Women's Hospital)
• Koen Van Leemput (Harvard Medical School, Massachusetts General Hospital)

Collaborators

• Andriy Fedorov (Harvard Medical School, Brigham and Women's Hospital)
• Vidya Rajagopalan, Chris Wyatt (Virginia Tech, Wake Forest University)

Objective

The goal of the EMSegment project is to create a segmentation module in Slicer3 that offers a high productivity and reliability to the clinician. The targeted application is the segmentation of MRI images of the brain. The EMSegment module allows the segmentation of other anatomical regions as long as a statistical atlas is available.

Approach, Plan

Our algorithm builds upon the Expectation Maximization theory and is structured to let the clinician make the most efficient use of his/her anatomical knowledge. Because of our intuitive visualization of probabilities, the user can understand the probabilistic features of his/her data and can efficiently tune the parameters to obtain the most accurate segmentation. To meet the time constraints of the tasks in a hospital, a main focus has been devoted to the acceleration of the EM segmentation algorithm.

Progress

We developed 3 new features to improve the visualization of parameters in the EMSegment and to help the user select the optimum values:

• Intensity Normalization
• Intensity Distribution by 2D Gaussian
• Global Priors

Our development has been committed into the Slicer3 trunk and will be available in Slicer 3.5. As a side project, we developed a module for MRI Bias Field Correction: MRIBiasFieldCorrection We provided support to the NA-MIC group studying alcohol and stress interaction in primates, and we have tested the module on monkey datasets. Besides, a Command Line version of this module has been developed to enable automated Testing in Nightly builds.

New Features

Intensity Normalization

We compute the normalization parameter for the atlas based on the value selected by the user.

Intensity Distribution Visualization

We model the image intensities using a Gaussian distribution for every class.

Global Prior Visualization

We compute the Global Priors based on the selection in the intensity histogram.

Collaboration

• Bayesian Segmentation of MRI Images by Koen Van Leemput, Sylvain Jaume, Polina Golland, Steve Pieper and Ron Kikinis.
• NA-MIC NCBC Collaboration: Measuring Alcohol and Stress Interaction by Vidya Rajagopalan, Andriy Fedorov and Chris Wyatt.