Spatial Correspondence between Prone and Supine CT Colonography Images: Creating a Reference Standard

Thomas Hampshire, Emma Helbren, Andrew Plumb, Holger Roth, Steve Halligan, David J. HawkesAbdominal Imaging 2013 · DOI

TL;DR

This paper tackles the challenge of aligning CT colonography (CTC) images taken with patients in different positions (prone and supine), which is crucial for developing accurate automated registration algorithms. They introduce a novel method where expert readers collaboratively create a "ground truth" reference standard by manually matching anatomical landmarks (haustral folds) across these image pairs.

What Problem Does This Paper Address?

Accurately aligning medical images from different acquisition states is a fundamental problem in medical image analysis. In the context of CT colonography (CTC), patients are often scanned in both a prone (face down) and a supine (face up) position. This is done to help distinguish between actual polyps and colonic contents or folds, as a true polyp will appear in the same location in both scans, while a fold or content might shift. However, the colon is a deformable organ, and its shape changes significantly between these two positions. This deformation makes it very difficult to automatically register (align) the prone and supine images. To evaluate and improve automated registration algorithms, researchers need a reliable "ground truth" or reference standard that accurately maps corresponding points between the prone and supine scans. Existing methods for creating such standards are limited, often relying on the locations of polyps, which are not always present or sufficient in number.

Methods

The core of this paper is the development of a method to create a high-quality reference standard for prone-supine CTC image registration. The process involves two main stages:

  1. Haustral Fold Segmentation: They first automatically identify and segment haustral folds, which are the characteristic ridges in the colon wall. This is achieved using a curvature-based metric applied to a surface mesh of the colon's interior. A graph cut segmentation algorithm is employed to distinguish between fold and non-fold regions based on curvature measurements.
  2. Manual Haustral Fold Matching (Reference Standard Creation): This is the novel contribution. They developed a custom graphical user interface (GUI) designed to aid expert readers in matching corresponding haustral folds between the prone and supine images. The GUI provides multiple synchronized views:
  3. External View: Shows the colon's outer surface.
  4. Internal View: Simulates a "virtual colonoscopy" from within the colon lumen.
  5. Unfolded View: Maps the entire colon surface onto a 2D plane, useful for seeing the global structure. Three readers (one computer scientist and two radiologists) independently used this GUI to match haustral fold centers. To ensure accuracy and build consensus, they used a dual-monitor setup for simultaneous visualization and employed navigation controls for detailed inspection. Discrepancies were resolved through consensus meetings.

Datasets

The study utilized 17 anonymized CT colonography (CTC) patient cases. These datasets were previously used in other publications by the authors. For 12 of these cases, the colon was well-distended in both prone and supine views. The remaining 5 cases exhibited local colonic collapse in one or both views, which is noted as a potential challenge. The authors have made this reference standard, along with the original CTC data, publicly available for the research community.

Key Insights and Ideas

  • Haustral folds as reliable landmarks: The paper highlights that haustral folds, despite being deformable, can serve as robust anatomical landmarks for establishing spatial correspondence when carefully matched.
  • Consensus-driven reference standard: The method emphasizes the power of expert consensus in creating a high-fidelity reference standard, especially for complex and subjective tasks like matching deformable structures.
  • Importance of multi-view visualization: The custom GUI's multi-view approach (external, internal, unfolded) is crucial for providing readers with sufficient context and confidence to make accurate correspondences.
  • Publicly available benchmark: The creation and release of a substantial, manually curated reference standard for prone-supine CTC registration is a significant contribution to the field, enabling better algorithm development and validation.

Key Results

  • A reference standard consisting of 1743 corresponding haustral fold pairs was established across the 17 CTC datasets.
  • On average, 102.5 haustral fold matches per case were achieved, providing a much denser set of correspondence points than typically available from polyp locations.
  • The comparison of individual reader labels against the final consensus standard showed varying levels of agreement, with "strong agreement" in 936 instances and "weak agreement" in 569 instances.
  • The final consensus standard excluded unresolved matches, resulting in a resolved rate of 58.7% of all initial potential matches across the three readers.

Limitations and Future Work

  • Unresolved matches: A significant number of haustral folds (1226 out of 2969 initial potential matches) remained unresolved, indicating the inherent difficulty in matching these deformable structures, especially in cases with colonic collapse.
  • Reader variability: While consensus was reached, there were discrepancies between individual readers, highlighting the subjective nature of the task and the need for clear guidelines and training.
  • Future Work: The authors suggest that this reference standard can be used to validate registration algorithms. Further work could focus on improving the automated segmentation of haustral folds and exploring more sophisticated methods for handling colonic collapse to increase the number of matched points. The availability of the dataset is a direct invitation for others to build upon this work.

Summary updated

BibTeX
@inproceedings{hampshire2013spatial,
  author = {Thomas Hampshire and Emma Helbren and Andrew Plumb and Holger Roth and Steve Halligan and David J. Hawkes},
  title = {{Spatial Correspondence between Prone and Supine CT Colonography Images: Creating a Reference Standard}},
  booktitle = {Abdominal Imaging},
  year = {2013},
  doi = {10.1007/978-3-642-41083-3\_5},
}
Figure