Location Registration and Recognition (LRR) for Longitudinal Evaluation of Corresponding Regions in CT Volumes

Michal Sofka, Charles V. Stewart

Abstract

In longitudinal clinical studies, monitoring specific lung nodules across multiple CT scans is critical for assessing disease progression. Traditional global deformable registration methods are often computationally prohibitive and prone to errors when significant anatomical changes occur. This paper introduces the Location Registration and Recognition (LRR) algorithm, which reframes registration as a local recognition task. By utilizing 3D structural features and a robust refinement pipeline, LRR achieves sub-2.0 mm accuracy in approximately 10 seconds per location, offering a high-performance alternative for interactive clinical use.

Introduction

Longitudinal imaging involves comparing scans of the same patient taken at different time points. A primary challenge is the precise alignment of specific regions of interest (ROIs), such as potentially cancerous nodules. While global deformable registration aims to align every voxel in a 3D volume, it is often inefficient for local monitoring. Such methods are slow and can fail if the lungs exhibit significant changes due to disease progression or variations in breath-hold depth.

The LRR framework addresses these issues by treating registration as a local "recognition" problem. Given a point in an initial scan, the system identifies the corresponding neighborhood in a subsequent scan and calculates the necessary local transformation to align them.

Methodology

The LRR pipeline moves away from dense voxel-based methods toward a feature-based approach:

  1. Feature Extraction: The algorithm identifies 3D structural features, classifying them into "sheet-like" (membranes), "tube-like" (vessels), or "corner-like" (junctions) categories.
  2. 3D Shape Contexts: For corner-like keypoints, a 195-component descriptor vector is generated. This 3D adaptation of shape contexts bins the distribution of nearby features into spherical coordinates, providing a highly distinctive signature for anatomical landmarks.
  3. Indexing and Hypothesis Generation: A k-d tree is utilized to rapidly find potential matches in the target scan for keypoints near the ROI.
  4. Robust Refinement: A robust version of the Iterative Closest Point (ICP) algorithm refines the alignment. Crucially, it is designed to ignore "outliers"—features present in one scan but absent in the other due to physical tissue changes.
  5. Decision Classifier: A Support Vector Machine (SVM) performs a final sanity check, evaluating alignment quality metrics to decide if the region was correctly recognized.
Visualization of local registration results using checkerboard overlays and feature point distributions.

Experimental Setup

The method was evaluated on pairs of screening and diagnostic lung CT scans taken roughly one year apart. The dataset included volumes up to 310 x 250 x 90 voxels with slice thicknesses ranging from 0.7 mm to 5.0 mm. These cases often featured visible physical changes in lung tissue, providing a rigorous test for registration robustness.

Results

The LRR algorithm demonstrated significant improvements in both speed and reliability over traditional methods:

  • Efficiency: The implementation runs in approximately 10 seconds per location on standard hardware, making it suitable for real-time clinical interaction.
  • Accuracy: The system achieved an average RMS error of under 2.0 mm, matching or exceeding the performance of state-of-the-art global methods like "Diffeomorphic Demons."
  • Reliability: The algorithm successfully aligned 97.3% of initial locations. The SVM verification step proved highly effective, identifying failures with minimal false positives.
Quantitative evaluation of registration success and failure across multiple test pairs.

Discussion and Conclusion

The core innovation of LRR is the shift from signal processing to object recognition. By focusing on a 50mm neighborhood, the algorithm avoids the computational burden and potential distractions of distant lung deformations. While the current reliance on "corner-like" features may limit performance in low-texture regions, the efficiency and accuracy of LRR represent a significant step forward for longitudinal medical imaging. Future work may explore adaptive neighborhood sizes and the exploitation of spatial relationships between multiple nodules to further enhance performance.

Summary updated

Abstract

The algorithm described in this paper takes (a) two temporally-separated CT scans, I 1 and I 2, and (b) a series of locations in I 1, and it produces, for each location, an affine transformation mapping the locations and their immediate neighborhood from I 1 to I 2. It does this without deformable registration by using a combination of feature extraction, indexing, refinement and decision processes. Together these essentially “recognize” the neighborhoods. We show on lung CT scans that this works at near interactive speeds, and is at least as accurate as the Diffeomorphic Demons algorithm [1]. The algorithm may be used both for diagnosis and treatment monitoring.
BibTeX
@inproceedings{sofka2008location,
  author = {Michal Sofka and Charles V. Stewart},
  title = {{Location Registration and Recognition (LRR) for Longitudinal Evaluation of Corresponding Regions in CT Volumes}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2008},
  doi = {10.1007/978-3-540-85990-1\_119},
}
Figure