Supervoxel Clustering with a Novel 3D Descriptor for Brain Tissue Segmentation

Yongfan Liu, Sen Du, Youyong KongInternational Journal of Machine Learning and Computing 2020 · DOI

Abstract

Accurate segmentation of brain tissues from magnetic resonance imaging (MRI) is of significant importance for clinical application and scientific research. Traditional strategies to handle the 2D images have the limitation of 3D data. In this paper, to overcome these issues, a tissue segmentation approach with supervoxel clustering and the novel 3D texture extraction method are proposed. At first, the simple linear iterative clustering in three-dimension is applied, to reduce the number of calculation objects. Then, a novel local binary pattern in three-dimension is proposed for better discriminate the supervoxels with different tissues. A clustering approach is also developed to classify supervoxels with features into different types of tissues. The labels of supervoxel are finally mapped back to original data to have the tissue type of voxels. The performance of the proposed method is evaluated on the commonly utilized Internet Brain Segmentation Repository 18 dataset. The experiment showed promising results with insufficient trainset.

Introduction

Accurate segmentation of brain tissues—specifically White Matter (WM), Gray Matter (GM), and Cerebrospinal Fluid (CSF)—is a cornerstone of modern clinical diagnosis and neurological research. However, researchers face two major hurdles: the high dimensionality of 3D MRI volumes, which makes voxel-wise processing computationally exhaustive, and the chronic scarcity of labeled training data due to privacy and cost constraints.

This paper introduces a novel 3D texture descriptor based on Local Binary Patterns (LBP) specifically designed for brain tissue segmentation. By combining 3D supervoxel clustering with these specialized descriptors and a k-Nearest Neighbor (kNN) classifier, the authors achieve robust segmentation results even when working with very limited training datasets.

Methodology

The proposed pipeline consists of four main stages designed to capture 3D spatial relationships efficiently:

  1. Supervoxel Generation (SLIC_3D): The authors adapt the Simple Linear Iterative Clustering (SLIC) algorithm to three dimensions. This groups similar voxels into "supervoxels," drastically reducing the number of objects the computer needs to classify while maintaining strong adherence to anatomical boundaries.
  2. Novel 3D LBP Operators: Standard LBP is 2D. To avoid the exponential complexity of a naive 3D extension, the authors propose streamlined operators: LBPa (Adjoin), which considers 6 face-sharing neighbors; LBPd (Diagonal), which considers 8 diagonal neighbors; and MLBP (Multi-LBP), a combination of both.
  3. Feature Vector Construction: For each supervoxel, an eigenvector is created containing spatial coordinates $(x, y, z)$, a grayscale histogram, and a "Texture Descriptor" derived from the 3D LBP histograms.
  4. Classification: A kNN model ($k=5$) is used to evaluate the performance of the new descriptors clearly without the complexity of deep neural networks.
The proposed brain tissue segmentation pipeline, illustrating the flow from original data to final classification.

Experimental Results

The method was evaluated using the Internet Brain Segmentation Repository 18 (IBSR18), consisting of 18 real MRI volumes. The MLBP approach achieved the best overall results with a Dice Similarity Coefficient (DSC) of 68.38% and a Segmentation Accuracy (SA) of 79.94%.

Final segmentation result showing the delineated brain tissues (WM, GM, and CSF).

Beyond accuracy, the system is highly efficient; the entire pipeline—from processing to prediction—runs in less than five minutes on a standard computer, making it suitable for clinical environments without specialized hardware.

Key Insights and Discussion

  • Supervoxels as Data Augmentation: By treating each supervoxel as a training sample rather than each MRI volume, the authors effectively turn a small dataset of 18 images into a large training set of thousands of samples.
  • The "Clockwise" Problem: Unlike 2D LBP, 3D space has no natural clockwise order. The authors demonstrated that even with an arbitrary fixed order, the resulting texture histograms are highly discriminative.
  • Spatial Weighting: Spatial information $(x, y, z)$ was found to be "weak" compared to high-dimensional texture histograms. A weight parameter $w=2.5$ was introduced to balance spatial location against texture features.
  • Texture vs. Grayscale: Interestingly, combining LBP features with raw grayscale histograms actually decreased performance, suggesting that raw grayscale introduces noise that interferes with the more robust texture patterns captured by LBP_3D.

Conclusion

While the system excels with small data, challenges remain in segmenting the complex, liquid patterns of Cerebrospinal Fluid (CSF). Future work will explore defining coding orders based on the inherent symmetry of brain organs to further refine feature extraction. Despite these challenges, the approach offers a powerful, computationally efficient tool for 3D medical image analysis.

Summary updated

Abstract

Accurate segmentation of brain tissues from magnetic resonance imaging (MRI) is of significant importance for clinical application and scientific research. Traditional strategies to handle the 2D images have the limitation of 3D data. In this paper, to overcome these issues, a tissue segmentation approach with supervoxel clustering and the novel 3D texture extraction method are proposed. At first, the simple linear iterative clustering in three-dimension is applied, to reduce the number of calculation objects. Then, a novel local binary pattern in three-dimension is proposed for better discriminate the supervoxels with different tissues. A clustering approach is also developed to classify supervoxels with features into different types of tissues. The labels of supervoxel are finally mapped back to original data to have the tissue type of voxels. The performance of the proposed method is evaluated on the commonly utilized Internet Brain Segmentation Repository 18 dataset. The experiment showed promising results with insufficient trainset.
BibTeX
@article{liu2020supervoxel,
  author = {Yongfan Liu and Sen Du and Youyong Kong},
  title = {{Supervoxel Clustering with a Novel 3D Descriptor for Brain Tissue Segmentation}},
  journal = {International Journal of Machine Learning and Computing},
  year = {2020},
  doi = {10.18178/ijmlc.2020.10.3.964},
}
Figure