Fast Groupwise Registration Using Multi-Level and Multi-Resolution Graph Shrinkage

Pei Dong, Xiaohuan Cao, Pew‐Thian Yap, Dinggang ShenScientific 2019 · DOI

Abstract

Groupwise registration aligns a set of images to a common space. It can however be inefficient and ineffective when dealing with datasets with significant anatomical variations. To mitigate these problems, we propose a groupwise registration framework based on hierarchical multi-level and multi-resolution shrinkage of a graph set. First, to deal with datasets with complex inhomogeneous image distributions, we divide the images hierarchically into multiple clusters. Since the images in each cluster have similar appearances, they can be registered effectively. Second, we employ a multi-resolution strategy to reduce computational cost. Experimental results on two public datasets show that our proposed method yields state-of-the-art registration accuracy with significantly reduced computational time.

Introduction

In medical imaging, groupwise registration is the essential process of aligning a large set of images—such as MRI scans from diverse patient cohorts—into a single, unbiased common space. This alignment is critical for longitudinal studies and for identifying subtle anatomical changes associated with neurodegenerative diseases like Alzheimer’s.

However, traditional registration methods face two primary hurdles. First, anatomical diversity makes it mathematically difficult to "warp" healthy brains and those with severe atrophy into the same space without significant errors. Second, the computational cost of aligning hundreds of high-resolution 3D volumes is prohibitive. Many existing methods rely on registering images to a "mean" template, which can become blurry and poorly defined, leading to degraded results.

The HMRML Framework

To address these challenges, we introduce the HMRML (Hierarchical Multi-level and Multi-resolution Graph Shrinkage) framework. This approach combines hierarchical clustering with graph-based registration to optimize both speed and accuracy.

Overview of the HMRML framework pipeline.

The framework follows a structured pipeline: input images undergo preprocessing before being organized via hierarchical image clustering. A hierarchical graph set is then constructed, followed by groupwise registration performed from the bottom of the hierarchy to the top. Finally, the system outputs the deformation fields required for alignment.

Hierarchical Graph Shrinkage Strategy

The core innovation of HMRML lies in its "divide and conquer" strategy. Instead of treating all images as a single distribution, the framework uses Affinity Propagation (AP) to group similar-looking images into clusters. It is significantly easier and more accurate to register images within a cluster of similar anatomical structures than to bridge large anatomical gaps in one step.

Hierarchical graph structure showing multi-level deformation pathways across image manifolds.

Within each cluster, images are treated as nodes in a graph. We employ "graph shrinkage" to move images along a manifold toward a group center. This occurs at multiple levels:

  1. Intra-graph registration: Aligning images within individual clusters (Level 1).
  2. Inter-graph registration: Aligning cluster "exemplars" to higher-level representatives (Level 2 and Level 3).

By using exemplar images to bridge clusters, the method avoids the "blurry template" problem common in iterative averaging methods. Furthermore, a multi-resolution strategy is employed, starting registration on down-sampled images for a quick "rough draft" before refining details at full resolution.

Experimental Results

The HMRML method was evaluated on two public neuroimaging datasets: LONI LPBA40 (40 healthy adults) and ADNI (100 subjects, including 50 with Alzheimer’s disease). The ADNI dataset represents a significant challenge due to the structural atrophy present in diseased brains.

Accuracy and Visual Quality

HMRML outperformed standard techniques like GroupMean and ABSORB, as well as the previous state-of-the-art graph method, HUGS. On the challenging ADNI dataset, HMRML achieved a Dice ratio (overlap accuracy) of 89.7%, compared to 81.8% for the conventional GroupMean approach. Visually, the resulting group mean images are sharper, preserving fine anatomical details that other methods typically blur.

Computational Efficiency

The most dramatic improvement was in processing speed. On the LONI dataset, HMRML was 15 times faster than HUGS. On the complex ADNI dataset, the speedup exceeded 25 times, reducing the estimated computation time from over 1,000 hours to just 43 hours.

Conclusion

HMRML provides a robust, high-speed framework for large-scale clinical studies. By leveraging hierarchical clustering and graph shrinkage, it makes the processing of massive brain image databases feasible with higher accuracy than previously possible. Future work may involve incorporating advanced shape features into the clustering phase or integrating deep learning-based registration agents to further enhance performance.

Summary updated

Abstract

Groupwise registration aligns a set of images to a common space. It can however be inefficient and ineffective when dealing with datasets with significant anatomical variations. To mitigate these problems, we propose a groupwise registration framework based on hierarchical multi-level and multi-resolution shrinkage of a graph set. First, to deal with datasets with complex inhomogeneous image distributions, we divide the images hierarchically into multiple clusters. Since the images in each cluster have similar appearances, they can be registered effectively. Second, we employ a multi-resolution strategy to reduce computational cost. Experimental results on two public datasets show that our proposed method yields state-of-the-art registration accuracy with significantly reduced computational time.
BibTeX
@inproceedings{dong2019fast,
  author = {Pei Dong and Xiaohuan Cao and Pew‐Thian Yap and Dinggang Shen},
  title = {{Fast Groupwise Registration Using Multi-Level and Multi-Resolution Graph Shrinkage}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-48491-9},
}
Figure