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

This page gives a high level overview of our research on Location Registration and Recognition (LRR). For more details, please refer to our article published in MICCAI 2008 proceedings.

Contents

• Overview
• Motivation and Intuition
• Results
• Summary
• Bibliography
• Download Source Code

Overview

The algorithm described in this paper takes (a) two temporally separated CT scans, and , and (b) a series of locations in , and it produces, for each location, an affine transformation mapping the locations and their immediate neighborhood from to . 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.

Motivation and Intuition

• Given:	image volumes and
  	set of locations from .

  
• Goal: find, for each , the affine transformation , which best aligns neighborhood with a region of .

Algorithm Outline

Figure 1: Diagram of the Location Registration and Recognition system. The initial transform maps the region surrounding the location from image into image . The transform is refined into accurate alignment in the estimation stage using correspondences between image features. If the verification step decides that the alignment is correct, the algorithm finishes. Otherwise, new initialization is generated.

Results

Figure 2: Examples of LRR (1st and 3rd column) vs. deformable registration (2nd and 4th column). Agreement of both results (a), and examples where LRR alignment is better (b). Features detected in fixed (blue) and moving (red) images drive the registration and robust estimation handles outliers (e.g. the case with changes on the bottom right which is partially aligned in the lower half).

(a)	(b)

Figure 3: Examples of nodule alignments shown as a checkerboard image alternating fixed and mapped moving axial slices. Images in the upper row have superimposed fixed features (blue) and mapped moving features (red). LRR correctly aligns nodules of various shapes and sizes in neighborhoods with different structural complexity.

Summary

• Algorithm for Location Registration and Recognition (LRR) without solving deformable registration first or simultaneously
• Technique to obtain initial transform using shape contexts
• Novel verification algorithm
• Handle changes within the local regions
• At least as accurate as the deformable registration
• Fast algorithm runs at near interactive speeds

     Future work:

• Combining results from multiple locations, exploring other applications

Publications and Further Reading

Bibliography

[1] Vercauteren, T., Pennec, X., Perchant, A., Ayache, N.: Non-parametric diffeomorphic image registration with the demons algorithm. In: Proceedings of the 10th International Conference of Medical Image Computing and Computer-Assisted Intervention (MICCAI 2007), Brisbane, Australia (2007) 319-326.