Abstract
Multi-atlas segmentation is a cornerstone of medical image analysis, yet its performance is often hampered by the assumption of intensity consistency between atlases and target images. This paper introduces a flexible, generative probabilistic framework for label fusion that overcomes this limitation. By employing a non-parametric model to learn the conditional distribution of intensities across modalities and incorporating a Markov Random Field for spatial smoothness, the approach enables accurate segmentation in inter-modality scenarios. Evaluated on T1-weighted atlases and Proton Density (PD) weighted targets, the non-parametric model achieves superior performance, particularly in subcortical regions, outperforming standard voting and Gaussian-based methods.
Introduction
Multi-atlas segmentation involves warping several expert-labeled images (atlases) to a target scan and fusing their labels. A significant challenge in this process is the "intensity consistency" assumption: the expectation that corresponding tissues have similar intensities across scans. This assumption fails in inter-modality applications, such as using T1-weighted atlases to segment Proton Density (PD) scans. While previous methods used Gaussian models to bridge this gap, they often struggle with complex, multi-modal intensity relationships found in structures like the thalamus.
Methods
The proposed framework treats label fusion as a generative inference problem. The core of the model is a latent membership field ($M$) that determines which atlas is responsible for a voxel's label and intensity.
Key components include:
- Latent Membership Field ($M$): Smoothed via a Markov Random Field (MRF) prior to ensure spatial consistency.
- Non-parametric Intensity Mapping: Instead of rigid Gaussian assumptions, the model learns a conditional distribution $p(I_{target} | I_{atlas})$ using 1D histograms. This allows the capture of complex, non-linear relationships between modalities.
- Bias Field Correction: The model explicitly accounts for low-frequency intensity distortions common in MRI by incorporating a multiplicative bias field into the generative process.
- Variational Inference: Parameters are estimated using Variational Expectation Maximization (VEM), iteratively updating the membership field, intensity mapping, and bias field.
Experimental Results
The framework was evaluated using 39 T1-weighted MRI atlases to segment 8 PD-weighted target scans across 36 brain structures. The non-parametric (NP) model was compared against Majority Voting and a state-of-the-art semi-parametric (SP) Gaussian-based method.
The NP model achieved the highest mean Dice score of 82.5%, a significant improvement over the Gaussian model (80.8%) and Majority Voting (77.2%). Notably, the NP approach showed marked superiority in subcortical structures like the Thalamus (TH) and Putamen (PT), where intensity relationships are particularly complex and non-Gaussian.
Discussion and Conclusion
The results demonstrate that atlas intensities provide crucial context for label fusion, especially when the relationship between modalities is non-linear. By moving beyond simple intensity matching, this framework provides a robust solution for inter-modality segmentation. Future work will focus on integrating the registration step directly into the probabilistic framework to mitigate errors from initial misalignments and exploring more sophisticated intensity quantization methods.
Summary updated