Construction of Abdominal Probabilistic Atlases and Their Value in Segmentation of Normal Organs in Abdominal CT Scans

Hyunjin Park, Alfred O. Hero Iii, Peyton H. Bland, Marc Kessler, Jongbum Seo, Charles R. Meyer

Introduction

Automated organ segmentation in abdominal CT scans is a critical task for radiation therapy planning and diagnostic imaging. However, the high anatomical variability of the abdomen—where organs shift and change size significantly between individuals—presents a major challenge compared to more stable regions like the brain. Probabilistic atlases, which map the likelihood of finding specific organs at given coordinates, are a common solution. Traditionally, these atlases are constructed by selecting a single "target" scan and warping others to match it. This study addresses the "target selection" problem: identifying a template scan that minimizes anatomical bias to improve segmentation accuracy.

Methodology

The researchers developed a method to identify the "least biased" target scan using Multidimensional Scaling (MDS). The process involved:

  1. MDS-Based Target Selection: Pair-wise registrations were performed between 24 abdominal CT scans. For each pair, the "bending energy"—a metric of the deformation required for alignment—was calculated. MDS was then used to map these distances into a 3D coordinate space. The scan closest to the mean geometry was designated the "best" target, while the most distant was the "worst."
  2. Optimized Registration: Geometric transforms were handled via Thin-Plate Splines (TPS) with 43 control points, using Mutual Information (MI) as the similarity measure.
  3. Segmentation Validation: Two methods were tested: simple atlas thresholding and a Maximum A Posteriori (MAP) approach utilizing Markov Random Fields (MRF) for spatial smoothing.

The dataset consisted of 24 scans with expert manual segmentations of the liver, spinal cord, and kidneys. A leave-one-out validation strategy was employed.

Results and Discussion

The choice of target scan significantly influenced atlas quality and segmentation performance. Visually, the atlas constructed from the "best" target exhibited much lower spatial variance, resulting in sharper organ boundaries compared to the "worst" target atlas.

Figure 1: Comparison of probabilistic atlases generated using the 'best' (a) and 'worst' (b) target scans. The 'best' atlas shows significantly sharper organ definitions.

Quantitatively, the "best" atlas achieved a mean Area Under the Curve (AUC) of 0.979, statistically outperforming the "worst" atlas (AUC = 0.962, p=0.0021). The MAP algorithm further improved results, achieving a false positive fraction of 0.0027 and a false negative fraction of 0.0236 when paired with the optimal atlas.

Figure 4: ROC curves demonstrating the impact of atlas quality on segmentation performance. The top curve (best atlas) shows superior true positive rates at low false positive fractions.

These findings suggest that the underlying anatomical model's quality is as vital as the complexity of the segmentation algorithm. While the MDS approach is computationally intensive for large datasets, it provides an objective, mathematical framework for template selection.

Conclusion and Future Work

This study demonstrates that MDS-based target selection effectively reduces bias in abdominal probabilistic atlases, leading to superior automated segmentation. Future work will focus on scaling this approach to larger populations and developing "target-free" construction methods to better handle topological variations and pathological changes.

Summary updated

Abstract

A good abdominal probabilistic atlas can provide important information to guide segmentation and registration applications in the abdomen. Here we bui …
BibTeX
@article{park2010construction,
  author = {Hyunjin Park and Alfred O. Hero III and Peyton H. Bland and Marc Kessler and Jongbum Seo and Charles R. Meyer},
  title = {{Construction of Abdominal Probabilistic Atlases and Their Value in Segmentation of Normal Organs in Abdominal CT Scans}},
  journal = {IEICE Trans. Information and Systems},
  year = {2010},
  doi = {10.1587/transinf.e93.d.2291},
}
Figure