Unsupervised Segmentation Algorithms' Implementation in ITK for Tissue Classification via Human Head MRI Scans

Shadman Sakib, Md. Abu Bakr SiddiquearXiv 2019 · arXiv

TL;DR

This paper explores unsupervised tissue classification of human head MRI scans using ITK. The authors implement and compare Otsu thresholding, Bayesian classification, and Bayesian classification with Gaussian smoothing, finding that while all methods show promise, Bayesian classification with Gaussian smoothing offers a good balance of performance and computational cost for differentiating brain tissues like white matter, gray matter, and cerebrospinal fluid.

What Problem Does This Paper Address?

Accurate segmentation of brain tissues from MRI scans is crucial for a wide range of applications, including diagnosis, treatment planning, and monitoring neurodegenerative disorders. However, this task is challenging due to inherent difficulties in MRI data, such as intensity inhomogeneity (variations in signal intensity across the image), partial volume effects (where a single voxel contains multiple tissue types), and noise. These issues can lead to ambiguous boundaries and misclassification, making robust and automated segmentation methods highly desirable.

Methods

The authors implemented three unsupervised tissue classification algorithms within the ITK (Insight Toolkit) framework:

  • Otsu Thresholding: This is a classic, non-parametric method that automatically determines an optimal threshold by maximizing the inter-class variance between foreground and background pixels. It's based on analyzing the image's histogram.
  • Bayesian Classification: This statistical approach models the probability distribution of different tissue classes. It uses Bayes' theorem to assign each voxel to the class with the highest posterior probability, often employing a Gaussian Mixture Model (GMM) for probability density estimation.
  • Bayesian Classification with Gaussian Smoothing: This method builds upon the Bayesian classifier by applying a discrete Gaussian filter to the posterior probabilities before assigning a label. This smoothing step aims to reduce noise and improve the continuity of segmented regions.

The implementation focused on classifying four tissue types: white matter (WM), gray matter (GM), cerebrospinal fluid (CSF), and background.

Datasets

The study utilized five grayscale head MRI scans. No new datasets were introduced, and the paper does not specify if these scans are publicly available or part of a specific collection.

Key Insights and Ideas

  • The challenge of CSF: Differentiating cerebrospinal fluid (CSF) from the background was identified as a particularly difficult task due to similar intensity values, leading to potential misclassification.
  • Smoothing improves boundaries: The Bayesian classification with Gaussian smoothing demonstrated better results in producing smoother edges between segmented tissues compared to the other methods.
  • Algorithm performance trade-offs: While Otsu thresholding is computationally inexpensive and simple, Bayesian methods, especially with smoothing, offer more nuanced classification, particularly for complex tissue boundaries.
  • Number of classes matters: The authors found that initially trying to classify into three classes yielded unsatisfactory results, and increasing the number of classes improved the classification of CSF.

Key Results

The paper presents qualitative results in the form of tables showing the output of each algorithm for different MRI scans.

  • Otsu Thresholding: Showed reasonable performance, particularly in identifying CSF in the center of the brain. However, it struggled with smooth edges.
  • Bayesian Classification: Performed reasonably well but also exhibited some limitations in edge smoothness and CSF classification.
  • Bayesian Classification with Gaussian Smoothing: Demonstrated better results in identifying CSF located between gray matter and the skull, and generally produced smoother tissue boundaries.

The authors note that neither pre-processing nor post-processing was applied, which might have contributed to some misclassifications.

Limitations and Future Work

  • Misclassification of CSF: A small portion of CSF was sometimes misclassified as background due to similar intensity values.
  • Edge smoothness: The connecting edges between different tissues were not always smooth, especially with the basic Otsu and Bayesian methods.
  • Need for pre/post-processing: The authors suggest that applying pre-processing (e.g., noise reduction) or post-processing techniques could further improve the classification accuracy.
  • Number of classes: The study highlights that indicating a higher number of classes in the algorithm can lead to better results, suggesting further exploration of optimal class numbers.
  • Augmenting with other techniques: The authors believe that combining these methods with other image processing techniques could enhance their performance.

Summary updated

BibTeX
@misc{sakib2019unsupervised,
  author = {Shadman Sakib and Md. Abu Bakr Siddique},
  title = {{Unsupervised Segmentation Algorithms' Implementation in ITK for Tissue Classification via Human Head MRI Scans}},
  year = {2019},
  eprint = {1902.11131},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1902.11131},
}
Figure