Room: ePoster Forums
Purpose: Intensity-based deformable image registration (DIR) requires minimizing an image dissimilarity metric that results in a nonlinear, nonconvex, and possibly discontinuous numerical optimization formulation, which can be problematic for gradient-based algorithms. A recent gradient-free method based on simple block-matching and the quadratic penalty function optimization algorithm has demonstrated a high spatial accuracy on inhale/exhale 4D computed tomography phases, which depict maximum voxel displacements less than 20. However, for extreme deformations, block match reference templates become inaccurate representations of estimated target image data. Thus, a data-driven block match template is proposed.
Methods: K-nearest neighbors, using normalized correlation distance, was applied to the 21x21 voxel patches centered on 2400 expert-determined landmarks placed on 8 extreme-deformation (COPDgene) breathhold CT pairs from www.dir-lab.com. The K=25 generated centroids were taken as representative block match templates. An image dissimilarity metric was defined as the mean absolute difference between the reference block correlations with the K inhale-templates and the target block correlations with the K exhale-template correlations. The structure of the data-defined dissimilarity metric is mathematically equivalent to standard block matching and is therefore suitable for the previous quadratic penalty method algorithm.
Results: Two additional COPDgene cases were registered with the quadratic penalty method and data-driven dissimilarity metric. Each case had 300 expert-determined landmarks. The mean landmark errors were 0.90 and 1.53. This represents an improvement over previous block match methods, which generated mean errors of 0.93 and 1.77 for the same cases.
Conclusion: : Results indicate that data-driven image dissimilarity metrics are feasible and capable of outperforming current methods. Moreover, the metric lends itself to already existing gradient-free algorithms. Future work includes gathering more landmark data and employing more sophisticated learning methods for template generation.