Skip to main content

Magnetic Resonance Image Tissue Classification Using A Partial Volume Model

Citations: 938 | Paper Published: 01 May 2001 | Institutions: University of Southern California, TRW Inc., Veterans Health Administration, University of Minnesota | Authors: David W. Shattuck, Stephanie R. Sandor-Leahy, Kirt Schaper, David A. Rottenberg, David A. Rottenberg, Richard M. Leahy

Abstract in 2 Lines

The paper describes a method to isolate and classify brain tissue in T1-weighted MRI images. The proposed algorithm achieves high accuracy on both real and phantom data sets.

Contributions​

The main contribution of this paper is the development and validation of a three-stage method for isolating and classifying brain tissue in T1-weighted MRI images. The proposed algorithm achieves high accuracy on both real and phantom data sets, with average kappa indices ranging from 0.746 to 0.928 depending on the type of tissue being classified. Additionally, the authors review related literature on each stage of their approach to provide context for their work within existing research efforts in low-level tissue classification using MRI imaging techniques.

Practical Implications​

The practical implications of this paper are that the proposed method provides a reliable and accurate way to isolate and classify brain tissue in T1-weighted MRI images. This can be useful for various applications, such as studying changes in brain structure due to disease or injury, planning surgical interventions on the brain, or developing new treatments for neurological disorders. The high accuracy achieved by the algorithm also suggests that it may outperform other existing methods currently used in clinical practice.

Methods​

Methodology
  • Non-brain tissue removal: This stage uses anisotropic diffusion filtering, edge detection, and mathematical morphology to remove non-brain tissues from the image.
  • Image normalization: This stage compensates for magnetic field inhomogeneities by fitting a tricubic B-spline gain field to local estimates of image nonuniformity spaced throughout the MRI volume.
  • Tissue classification: In this final stage, voxels in the intensity-normalized image are classified into six different types of brain tissue using a maximum-a-posteriori classifier that combines partial volume measurement models with Gibbs priors modeling spatial properties of the brain.

Data​

The paper uses two types of data sets for validation testing: real MRI brain images and phantom (simulated) MRI brain images. The 20 normal MRI brain data sets used in the study were obtained from the Internet Brain Segmentation Repository, while a set of 12 volumes was taken from Montreal Neurological Institute's BrainWeb phantom dataset. Both datasets are widely recognized as standard benchmarks within the field of low-level tissue classification using magnetic resonance imaging techniques.

Results​

Performance Analysis
  • The paper presents a three-stage method for isolating and classifying brain tissue in T1-weighted MRI images. The proposed algorithm achieves high accuracy on both real and phantom data sets, with average kappa indices ranging from 0.746 to 0.928 depending on the type of tissue being classified.
  • Specifically, using data from the Internet Brain Segmentation Repository dataset, our method achieved an average kappa index of κ = 0.746 +/- 0.114 for gray matter (GM) classification and κ = 0.798 +/- .089 for white matter (WM) compared to expert-labeled ground truth labeling methods used as benchmarks during validation testing.
  • Using simulated volumes taken from Montreal Neurological Institute's BrainWeb phantom dataset we obtained even higher results: achieving an average kappa index of κ= .893+/- .041for GM classificationandκ=.928+/-.039for WMclassificationcomparedtothegroundtruthlabelingmethodsusedasbenchmarksduringvalidationtesting

Limitations​

  • Some technical details about the implementation are left out from the paper, making it difficult for others to reproduce exactly what was done without additional information.
  • The proposed method is designed specifically for T1-weighted MRI images and may not be applicable to other types of imaging modalities.

Table of Contents