Bayesian Markov model for cooperative clustering: application to robust MRI brain scan segmentation

Florence Forbes, Benoît Scherrer, Michel DojatHAL 2012

TL;DR

This paper introduces a novel Bayesian Markov model for cooperative clustering, specifically designed for robust MRI brain scan segmentation. By modeling the interdependence between tissue and structure segmentation, and incorporating prior knowledge, the method achieves more accurate and consistent results than existing approaches, particularly in the presence of noise and intensity variations.

What Problem Does This Paper Address?

Accurate segmentation of brain tissues and structures from MRI scans is crucial for various neuroscientific and clinical applications. However, this task is challenging due to several factors:

  • Interdependence: Tissue segmentation (e.g., white matter, grey matter, CSF) and structure segmentation (e.g., ventricles, thalamus) are inherently linked. A structure is composed of specific tissues, and knowing the location of structures can inform tissue classification. Existing methods often treat these as separate, sequential tasks, missing out on valuable complementary information.
  • Noise and Intensity Variations: MRI scans can suffer from noise and intensity inhomogeneities across the volume, making it difficult to consistently classify tissues and structures.
  • Atlas Registration: Automatic segmentation often relies on registering an anatomical atlas to the MRI scan. This registration process itself is complex and can introduce biases if not performed accurately or if treated independently from segmentation.

The authors aim to address these challenges by developing a unified framework that simultaneously segments tissues and structures, leverages their interdependence, and integrates prior knowledge, including atlas information, in a principled Bayesian manner.

Methods

The core of the proposed method is a Bayesian cooperative clustering framework built upon Markov Random Fields (MRFs). Key methodological aspects include:

  • Cooperative Clustering: Instead of producing a single partition, the model aims to generate multiple, related partitions (one for tissues, one for structures) by explicitly modeling their interdependence. This is achieved through a joint probabilistic model.
  • Conditional Models and Generalized Alternating Minimization (GAM): The authors propose a framework based on a system of coherently linked conditional models. This allows for inference using Generalized Alternating Minimization (GAM) procedures, which generalize the standard Expectation-Maximization (EM) algorithm. This approach breaks down the complex joint optimization into a sequence of more interpretable steps.
  • Hierarchical Modeling: The model incorporates dependencies at multiple levels:
  • Spatial Dependencies: Within each label set (tissues or structures), MRFs are used to model spatial interactions between neighboring voxels, ensuring local consistency.
  • Cooperative Dependencies: Interactions between tissue and structure labels are explicitly modeled, allowing information from one to inform the other.
  • Prior Knowledge: A probabilistic atlas is used to incorporate prior anatomical knowledge, and this knowledge is integrated into the model through registration parameters.
  • Joint Registration and Segmentation: The model simultaneously estimates segmentation labels and registration parameters. This allows for adaptive correction of the initial atlas registration, preventing commitment to potentially suboptimal initial alignments.
  • Handling Intensity Nonuniformity: A specific parametrization of tissue intensity distributions is used, which allows for local estimation of parameters and improves robustness to intensity variations.

Datasets

  • BrainWeb Phantoms: The authors use the normal 1 mm³ BrainWeb phantoms database. These are simulated MRI scans generated from a realistic brain anatomical model, allowing for quantitative evaluation against ground truth. They are generated with varying levels of noise (3%, 5%, 7%, 9%) and nonuniformity (20%, 40%).
  • Real 3T Brain Scans: The method is also evaluated on real 3T MRI brain scans from the Grenoble Institut of Neuroscience (GIN).

The paper does not introduce new benchmark datasets but rather uses existing simulated and real data for evaluation.

Key Insights and Ideas

  • Cooperation is Key: The most significant insight is that explicitly modeling the interdependence between tissue and structure segmentation, and between segmentation and registration, leads to substantial improvements in accuracy and robustness. This "cooperative" approach allows the model to leverage complementary information across different tasks.
  • Unified Framework: By formulating the problem within a single Bayesian probabilistic model, the authors achieve a principled and consistent treatment of segmentation, registration, and handling of intensity variations. This avoids the limitations of sequential or independent processing.
  • Adaptive Registration: Simultaneously estimating registration parameters alongside segmentation labels allows the model to adapt and refine the atlas alignment locally, overcoming the limitations of fixed, global registration.
  • GAM for Tractability: The use of Generalized Alternating Minimization procedures makes the complex joint model computationally tractable, enabling practical application.

Key Results

  • Tissue Segmentation: For tissue segmentation alone, the proposed method (LOCUSB-T) achieved a mean Dice metric of 89% on BrainWeb phantoms, comparable to or slightly better than state-of-the-art methods like LOCUS-T (89%) and FAST (88%), while being significantly faster than SPM5 (86%).
  • Joint Tissue and Structure Segmentation:
    • When registration was fixed (LOCUSB-TS), performance was good, but iterative estimation of registration parameters (LOCUSB-TSR) led to significant improvements.
  • For the Left Caudate Nucleus, LOCUSB-TSR achieved a 23% relative improvement in Dice coefficient compared to LOCUSB-TS.
  • Overall, LOCUSB-TSR showed stable gains across different noise and inhomogeneity levels for the three structures evaluated.
  • Robustness: The method demonstrated satisfying robustness to noise and intensity nonuniformity.
  • Computational Efficiency: The method achieved competitive segmentation times, with joint tissue and structure segmentation taking approximately 10 minutes for three structures.

Limitations and Future Work

  • Interpolation Step: The authors acknowledge that the interpolation step used to handle intensity nonuniformity at a voxel level could be further improved. They suggest generalizing this stage by considering various degrees of locality, mimicking a multiresolution approach.
  • Prior Specification: The choice of prior for intensity distribution parameters was guided by the need for a valid Markov model. While flexible, exploring other conditional specifications and their implications for inference is suggested.
  • Spatial Component in Cooperation: The current model incorporates spatial interactions within label sets and cooperative interactions between label sets. The authors suggest that adding a spatial component to the cooperation mechanisms themselves could be a valuable extension.
  • Hyperparameter Tuning: While the paper mentions fixing some hyperparameters (ητ, ης), further investigation into their estimation or more principled selection could be beneficial.
  • Generalization: The framework is presented as general and adaptable to other applications involving complex joint processes. Future work could explore its application beyond MRI segmentation.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{forbes2012bayesian,
  author = {Florence Forbes and Benoît Scherrer and Michel Dojat},
  title = {{Bayesian Markov model for cooperative clustering: application to robust MRI brain scan segmentation}},
  booktitle = {HAL},
  year = {2012},
}
Figure