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.
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:
- 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.
- Maximization Step (M-Step): Estimates the "bias field" (intensity corruption caused by the MRI scanner) based on current tissue classifications.
- 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.
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