Incorporating Non-rigid Registration into Expectation Maximization Algorithm to Segment MR Images

Kilian M. Pohl, William M. Wells Iii, Alexandre Guimond, Kiyoto Kasai, Martha Elizabeth Shenton, Ron Kikinis, W. Eric L. Grimson, Simon K. Warfield

Abstract

Automated segmentation of complex brain structures like the superior temporal gyrus (STG) remains a challenge due to high anatomical variability and image artifacts. This paper presents the EM-MF-LP algorithm, which integrates non-rigid registration into an Expectation-Maximization (EM) framework. By utilizing deformable registration to align local prior probability maps (PPMs) to subject-specific anatomy, the method provides necessary spatial context while simultaneously correcting for intensity inhomogeneities and noise via a Mean-Field (MF) approximation. Validation against manual segmentations demonstrates that the algorithm achieves human-level performance with a Dice Similarity Coefficient (DSC) exceeding 0.70.

Introduction

Quantitative neuroanatomy is crucial for diagnosing and monitoring neurological disorders. However, manual segmentation is labor-intensive and subject to inter-rater variability. Automated methods typically fall into two categories: intensity-based and registration-based. Intensity-based methods, such as standard EM, effectively handle bias fields but lack spatial context, often confusing different structures with similar intensities. Registration-based methods provide spatial context but often lack the flexibility to account for unique subject anatomy. This study addresses these limitations by proposing a hybrid approach, specifically targeting the difficult-to-segment Superior Temporal Gyrus (STG).

Methodology

The proposed Expectation Maximization - Mean Field - Local Prior (EM-MF-LP) algorithm operates through a synergistic combination of registration and statistical estimation.

Workflow of the EM-MF-LP algorithm integrating non-rigid registration and the EM framework.

The process begins with Non-rigid Registration, where a statistical atlas derived from 82 subjects is warped to fit the specific subject's brain. This transformation projects Local Prior Probability Maps (PPMs) into the subject's coordinate space, providing voxel-specific anatomical priors ($P(T|x,y,z)$).

The core of the segmentation is the EM-MF Framework:

  1. Expectation Step (E-Step): Calculates tissue class likelihoods for each voxel. Unlike standard EM, this step uses spatially dependent Local Priors. A Mean-Field (MF) approximation is incorporated to model spatial dependencies between neighboring voxels, acting as a powerful noise reduction filter.
  2. Maximization Step (M-Step): Estimates the "bias field" (intensity corruption caused by the MRI scanner) based on current tissue classifications.
  3. Iteration: The algorithm iterates between these steps until convergence, progressively refining both the segmentation and the bias field correction.

Results and Discussion

The algorithm was validated using multi-channel MRI data (SPGR and T2-weighted) from four subjects, compared against ground-truth manual segmentations from three experts. The study successfully segmented the brain into 11 different tissue classes, including the amygdala and hippocampus.

Representative segmentation result showing the parcellation of various brain structures.

The EM-MF-LP algorithm achieved a DSC of over 0.70, indicating excellent agreement with manual raters and performing within the range of inter-human variability. The use of non-rigid registration was found to be critical; simpler affine registrations often led to "bleeding" into adjacent structures like the middle temporal gyrus. By estimating the bias field during the segmentation process, the model remained robust to signal variations.

Future work aims to address minor boundary spreading by including more neighboring structures as distinct tissue classes and exploring iterative refinement where the initial segmentation guides a second round of non-rigid registration to achieve even tighter alignment.

Summary updated

Abstract

The paper introduces an algorithm which allows the automatic segmentation of multi channel magnetic resonance images. We extended the Expectation Maximization-Mean Field Approximation Segmenter, to include Local Prior Probability Maps. Thereby our algorithm estimates the bias field in the image while simultaneously assigning voxels to different tissue classes under prior probability maps. The probability maps were aligned to the subject using non-rigid registration. This allowed the parcellation of cortical sub-structures including the superior temporal gyrus. To our knowledge this is the first description of an algorithm capable of automatic cortical parcellation incorporating strong noise reduction and image intensity correction.
BibTeX
@inproceedings{pohl2002incorporating,
  author = {Kilian M. Pohl and William M. Wells III and Alexandre Guimond and Kiyoto Kasai and Martha Elizabeth Shenton and Ron Kikinis and W. Eric L. Grimson and Simon K. Warfield},
  title = {{Incorporating Non-rigid Registration into Expectation Maximization Algorithm to Segment MR Images}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2002},
  doi = {10.1007/3-540-45786-0\_70},
}
Figure