Arrangement: A Spatial Relation Between Parts for Evaluating Similarity of Tomographic Section

Hemant D. Tagare, Frans Vos, Conrade C. Jaffe, James S. DuncanIEEE TPAMI 1995 · DOI

Abstract

Medical tomographic images are formed by the intersection of the image plane and an object. As the image plane changes, different parts of the object come in view or drop out of view. However, for small changes of the image plane, most parts continue to remain visible and their qualitative embedding in the image remains similar. Therefore, similarity of part embeddings can be used to infer similarity of image planes. A spatial relation called "arrangement" is proposed to describe part embeddings, formulated via Voronoi diagrams. A metric for arrangements is developed using diagonal exchanges. Experiments with a cardiac MRI database validate that similarity of image planes can be inferred from part embeddings, with performance comparable to expert radiologists.

Introduction: The Challenge of Tomographic Similarity

In medical imaging, a "section" is a 2D slice of a 3D anatomical structure. Because the human body is geometrically complex, even minor tilts or shifts in the imaging plane can significantly alter the visual appearance of organs. For clinicians or researchers searching a database for "similar" images, traditional pixel-to-pixel comparisons often fail due to these geometric variations.

Figure 1: (a) 3D imaging geometry and (b) the resulting 2D tomographic section (Section 1) showing the heart chambers.

The fundamental problem addressed in this research is how a computational system can recognize that two images represent the same anatomical "view" despite shifts in shape and size. The authors hypothesize that while exact geometry changes, the qualitative relationship—specifically, which parts surround which—remains stable across similar imaging planes.

Methodology: The "Arrangement" Framework

The authors propose a formal framework rooted in computational geometry to capture these qualitative relationships:

1. Voronoi Representation

Anatomical parts, such as the chambers of the heart (left/right ventricles and atria), are represented as regions in a Voronoi diagram. In this model, two parts are defined as "neighbors" if their respective Voronoi cells share a common edge.

2. The Arrangement String

For every anatomical part, a "local arrangement" is generated—a string listing its neighbors in counterclockwise order. The "overall arrangement" of the image is the collection of these local strings, providing a topological signature of the section.

Figure 2: A 2D section from a similar imaging plane. While the shapes of the chambers (ra, rv, la, lv) have changed, their relative arrangement remains consistent.

3. Diagonal Exchange and Distance Metric

To compare two different arrangements, the authors employ a graph operation known as a "diagonal exchange" (or the $T_0$ operator). This operation swaps the adjacency of four neighboring regions. The "distance" between two images is defined as the minimum number of diagonal exchanges required to transform one Voronoi graph into the other. This is calculated using a breadth-first search on an "exchange tree," optimized with pruning strategies to ensure computational efficiency.

Experimental Results: Cardiac MRI Retrieval

The study utilized a cardiac MRI database containing 100 images where the outlines of heart chambers (left/right ventricles and atria), ascending aorta, and pulmonary artery were manually traced. Three expert cardiac radiologists provided "gold standard" similarity ratings to evaluate the algorithm's performance.

The results demonstrated a high positive correlation (0.64–0.66) between the algorithm's similarity ratings and those of human experts. In a retrieval task, expert radiologists achieved 92% accuracy, while the automated "Arrangement" method achieved 81%. While the algorithm was more "permissive" (returning a 15% false positive rate compared to the experts' 2.6%), this behavior is often considered acceptable in database search contexts where high recall is prioritized.

Discussion and Future Directions

A key insight of this work is that the neighborhood system of organs is more critical for identifying an anatomical plane than specific boundary coordinates. This is supported by experiments showing that radiologists can identify imaging planes even when organ shapes are masked, relying primarily on their relative positions. Mathematically, the authors proved that any cubic connected Voronoi diagram can be reduced to a standard "flower" shape through a sequence of exchanges, ensuring the distance metric is robust and always computable.

However, the current "Arrangement" metric is purely topological and ignores size and shape. Future work aims to integrate relative dimensions to improve precision and automate the organ segmentation process, which currently relies on manual outlines. This topological approach also holds promise for broader computer vision applications, such as defining complex directional relations between non-point objects.

Summary updated

BibTeX
@article{tagare1995arrangement,
  author = {Hemant D. Tagare and Frans Vos and Conrade C. Jaffe and James S. Duncan},
  title = {{Arrangement: A Spatial Relation Between Parts for Evaluating Similarity of Tomographic Section}},
  journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  year = {1995},
  doi = {10.1109/34.406653},
}

Loading related content...

Figure