Registration of Longitudinal Image Sequences with Implicit Template and Spatial-Temporal Heuristics

Guorong Wu, Qian Wang, Hongjun Jia, Dinggang ShenMICCAI 2010 · OpenReview

Abstract

Accurate measurement of longitudinal changes of anatomical structure is important and challenging in many clinical studies. Also, for identification of disease-affected regions due to the brain disease, it is extremely necessary to register a population data to the common space simultaneously. In this paper, we propose a new method for simultaneous longitudinal and groupwise registration of a set of longitudinal data acquired from multiple subjects. Our goal is to 1) consistently measure the longitudinal changes from a sequence of longitudinal data acquired from the same subject; and 2) jointly align all image data (acquired from all time points of all subjects) to a hidden common space. To achieve these two goals, we first introduce a set of temporal fiber bundles to explore the spatial-temporal behavior of anatomical changes in each longitudinal data of the same subject. Then, a probabilistic model is built upon the hidden state of spatial smoothness and temporal continuity on the fibers. Finally, the transformation fields that connect each time-point image of each subject to the common space are simultaneously estimated by the expectation maximization (EM) approach, via the maximum a posterior (MAP) estimation of probabilistic models. Promising results are obtained to quantitatively measure the longitudinal changes of hippocampus volume, indicating better performance of our method than the conventional pairwise methods.

Introduction

In longitudinal brain studies, researchers track patients over years to monitor the progression of neurodegenerative diseases like Alzheimer’s. Measuring the subtle annual changes in anatomical structures is notoriously difficult because standard image registration—the process of overlaying images for comparison—often treats each time point independently. This independence can introduce "noise," where a structure might appear to grow one year and shrink the next due to registration errors rather than biological reality.

Furthermore, most registration methods rely on a specific "template" image as a reference, which introduces bias. This paper proposes a novel framework to simultaneously align brain images from multiple subjects across multiple time points into a "hidden" common space, ensuring consistency and unbiased alignment.

Methodology: Temporal Fiber Bundles and Implicit Templates

The proposed method integrates groupwise registration (aligning many subjects) and longitudinal registration (aligning time sequences) into a unified probabilistic model.

Conceptual framework for simultaneous groupwise and longitudinal registration, showing the mapping from the mean shape Z to individual subjects and the temporal fiber bundles.

Core Components:

  • Implicit Template: Instead of selecting a single subject as a reference, the model defines a "hidden" common space. All images are warped toward this mathematical mean shape simultaneously, removing template bias.
  • Temporal Fiber Bundles: This core innovation tracks specific anatomical points across all time points for every subject. These trajectories are treated as "fibers." The model applies a smoothing constraint along these fibers, ensuring that brain deformation over time follows a continuous, biologically plausible path.
  • EM Framework: The problem is solved using an Expectation-Maximization (EM) approach. It iteratively estimates the "true" locations of brain structures in the common space and updates transformation fields to minimize "free energy," an optimization concept borrowed from physics.

Experimental Results

The researchers evaluated the method using both simulated atrophy data (5% annual hippocampal shrinkage) and real-world MRI scans from nine elderly subjects over five years.

Averaged shrinkage ratio of hippocampus volume for simulated data, comparing the proposed method against standard 3D registration.

Key Findings:

  • Superior Accuracy: In simulations with 15.86% actual atrophy, the proposed method estimated a 15.73% loss, significantly closer to the ground truth than the 12.8% detected by standard 3D registration.
  • Clinical Correlation: In the elderly brain study, the method detected 5.38% hippocampal shrinkage, aligning closely with the 5.65% measured by expert human raters. The standard method only captured 3.84%.
  • Improved Alignment: The method achieved higher overlap ratios for white matter, gray matter, and ventricles, indicating a tighter and more accurate alignment across the population.

Conclusion and Future Work

By solving for group alignment and temporal smoothness simultaneously, the model leverages information from the entire group to improve the registration of individual, noisy images. The "fiber" heuristic effectively filters out the noise inherent in independent registrations. Future work may involve applying more sophisticated kernel-based regression to model complex disease progressions and testing the framework's ability to predict the transition from Mild Cognitive Impairment (MCI) to Alzheimer's disease.

Summary updated

Abstract

Accurate measurement of longitudinal changes of anatomical structure is important and challenging in many clinical studies. Also, for identification of disease-affected regions due to the brain disease, it is extremely necessary to register a population data to the common space simultaneously. In this paper, we propose a new method for simultaneous longitudinal and groupwise registration of a set of longitudinal data acquired from multiple subjects. Our goal is to 1) consistently measure the longitudinal changes from a sequence of longitudinal data acquired from the same subject; and 2) jointly align all image data (acquired from all time points of all subjects) to a hidden common space. To achieve these two goals, we first introduce a set of temporal fiber bundles to explore the spatial-temporal behavior of anatomical changes in each longitudinal data of the same subject. Then, a probabilistic model is built upon the hidden state of spatial smoothness and temporal continuity on the fibers. Finally, the transformation fields that connect each time-point image of each subject to the common space are simultaneously estimated by the expectation maximization (EM) approach, via the maximum a posterior (MAP) estimation of probabilistic models. Promising results are obtained to quantitatively measure the longitudinal changes of hippocampus volume, indicating better performance of our method than the conventional pairwise methods.
BibTeX
@inproceedings{wu2010registration,
  author = {Guorong Wu and Qian Wang and Hongjun Jia and Dinggang Shen},
  title = {{Registration of Longitudinal Image Sequences with Implicit Template and Spatial-Temporal Heuristics}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2010},
  url = {https://openreview.net/forum?id=bXTtOZByDD},
}
Figure