Automatic three-dimensional segmentation of MR images applied to the rat uterus

Ayelet Akselrod-Ballin, Erez Eyal, Meirav Galun, Edna Furman‐Haran, John M. Gomori, Ronen Basri, Hadassa Degani, Achi BrandtProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2006 · DOI

TL;DR

This paper introduces a novel, fast, and automatic 3D segmentation algorithm for MRI scans, capable of handling anisotropic data and multiple imaging modalities simultaneously. The method successfully delineates the rat uterus with high accuracy, offering a significant advancement for non-invasive in vivo studies of this crucial organ.

What Problem Does This Paper Address?

Accurately segmenting anatomical structures in 3D medical images, particularly MRI, is a fundamental yet challenging task. Manual segmentation is time-consuming, prone to errors, and difficult to reproduce, especially for complex organs like the rat uterus, which is embedded within other abdominal tissues. Existing automatic segmentation methods often struggle with anisotropic data (where resolution differs across dimensions), require significant computational resources, or are limited in their ability to integrate information from multiple MRI sequences (modalities). This paper aims to address these limitations by developing a robust and efficient algorithm for precise 3D segmentation of the rat uterus, facilitating more reliable and reproducible research.

Methods

The core of this work is an extension of the 2D Segmentation by Weighted Aggregation (SWA) algorithm to a 3D, multi-channel context. The algorithm constructs a graph where each voxel is a node, and edges connect neighboring voxels with weights reflecting their similarity. This graph is then recursively coarsened using an algebraic multigrid (AMG) approach, creating a hierarchical representation of the image. At each level of this hierarchy, a "normalized-cut-like" cost function is minimized to identify salient segments (regions that are distinct from their surroundings).

Key innovations include:

  • 3D Multiscale Approach: The algorithm operates in 3D, creating a hierarchical decomposition of the image from fine to coarse scales. This allows for the detection of structures at various levels of detail.
  • Anisotropic Data Handling: The method is designed to work with anisotropic MRI data, which is common in biological imaging. It initially performs 2D coarsening on each slice and transitions to full 3D coarsening when inter-slice distances become comparable to in-slice distances.
  • Multi-Channel Integration: The framework can seamlessly integrate data from multiple MRI sequences (e.g., T1w, T2w, PDw). Each voxel's features are represented as a vector, allowing the algorithm to leverage complementary information from different modalities.
  • Rich Feature Set: Beyond simple intensity, the algorithm computes and utilizes a rich set of features, including intensity, shape, and location, to characterize segments and guide the segmentation process.
  • Linear Time Complexity: The algorithm achieves a linear time complexity (O(N), where N is the number of voxels), making it computationally efficient for large 3D datasets.

Datasets

The study utilizes in vivo MR images of female, ovariectomized rats. The specific datasets were acquired using two MRI protocols:

  • T2 weighted (T2w): This sequence provided good contrast for the endometrium.
  • Proton density weighted (PDw): This sequence offered better visualization of the entire uterus but had lower contrast between the uterus and surrounding organs like the intestine.

The paper does not introduce a new benchmark dataset but rather uses these acquired images for evaluating the performance of their segmentation algorithm against manual delineations.

Key Insights and Ideas

  • The Power of Multi-Channel MRI: Combining data from different MRI sequences (T2w and PDw) significantly improves the segmentation accuracy of the entire uterus compared to using a single channel. This highlights the complementary nature of different MRI contrasts for anatomical delineation.
  • 3D is Crucial for Complex Structures: For organs like the rat uterus horns, which are elongated and can be partially visible in individual slices, a true 3D segmentation approach is vital for accurate and complete detection.
  • Hierarchical Segmentation for Robustness: The multiscale, hierarchical nature of the algorithm allows it to overcome local ambiguities and artifacts by considering global saliency measures, leading to more robust segmentation.
  • Efficiency Matters: Achieving linear time complexity is a significant advantage, making this method practical for routine use in research settings where rapid analysis is often required.

Key Results

  • High Accuracy: The automatic segmentation achieved high correlation with manual segmentations.
  • Endometrial segmentation using single-channel T2w data yielded a spatial overlap of 0.91±0.04 (Dice coefficient).
  • Myometrial and endometrial segmentation using multi-channel data achieved a spatial overlap of 0.91±0.03.
  • Improved Detection with Multi-Channel: The multi-channel approach significantly improved the detection of the entire uterus, with a Dice coefficient of 0.91±0.03, compared to single-channel PDw which resulted in 0.95±0.01 but with a lower overlap score of 0.57±0.14 for the entire uterus.
  • Computational Efficiency: The algorithm runs in linear time complexity (O(N)). For a 200x200x15 voxel dataset, single-channel segmentation took approximately 40 seconds on a standard P4 3GHz PC, and multi-channel segmentation took 60 seconds.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the algorithm is robust to some MRI artifacts, further improvements could be made by incorporating more high-level anatomical knowledge.
  • The current implementation is described as "non-optimal," suggesting potential for further speedups.
  • Future Work:
    • The authors suggest incorporating prior knowledge of the uterus shape (e.g., length, width, orientation) to refine the segmentation.
  • Post-processing steps based on anatomical knowledge and the properties of different channels could further enhance precision.
  • The general framework is highlighted as being applicable to other anatomical structures and tasks, including MR angiography and virtual colonoscopy.
  • Future work will also focus on using this segmentation method to analyze time courses of functional and molecular imaging of uterine tissue.

Summary updated

Abstract

We introduce an automatic 3D multiscale automatic segmentation algorithm for delineating specific organs in Magnetic Resonance images (MRI). The algorithm can process several modalities simultaneously, and handle both isotropic and anisotropic data in only linear time complexity. It produces a hierarchical decomposition of MRI scans. During this segmentation process a rich set of features describing the segments in terms of intensity, shape and location are calculated, reflecting the formation of the hierarchical decomposition. We show that this method can delineate the entire uterus of the rat abdomen in 3D MR images utilizing a combination of scanning protocols that jointly achieve high contrast between the uterus and other abdominal organs and between inner structures of the rat uterus. Both single and multi-channel automatic segmentation demonstrate high correlation to a manual segmentation. While the focus here is on the rat uterus, the general approach can be applied to recognition in 2D, 3D and multi-channel medical images.
BibTeX
@inproceedings{akselrodballin2006automatic,
  author = {Ayelet Akselrod-Ballin and Erez Eyal and Meirav Galun and Edna Furman‐Haran and John M. Gomori and Ronen Basri and Hadassa Degani and Achi Brandt},
  title = {{Automatic three-dimensional segmentation of MR images applied to the rat uterus}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2006},
  doi = {10.1117/12.653371},
}
Figure