A Generalized Learning Based Framework for Fast Brain Image Registration

Minjeong Kim, Guorong Wu, Pew-Thian Yap, Dinggang Shen

Introduction

Deformable image registration is a fundamental task in neuroimaging, enabling the comparison of diverse brain structures within a standardized coordinate space. Despite its importance, traditional registration methods are often hampered by high computational costs and a susceptibility to local minima. These issues arise because most algorithms start from a generic initialization, requiring extensive iterative optimization to align a subject's unique anatomy to a template. This paper proposes a generalized framework that leverages machine learning to provide a subject-specific "warm-start," significantly accelerating the registration process without sacrificing accuracy.

Methods

The proposed framework functions as a wrapper around existing registration algorithms, such as HAMMER or Diffeomorphic Demons. It operates in three distinct stages:

  1. Statistical Modeling and Simulation: To build a robust predictive model, the authors first register a small set of real brain images to a template. Principal Component Analysis (PCA) is applied to the resulting deformation fields to identify primary modes of anatomical variation. To address the challenge of limited training data, the framework perturbs these PCA coefficients to generate a large synthetic dataset (12,500 samples).
  2. Learning the Mapping: The framework utilizes Support Vector Regression (SVR) to learn the non-linear relationship between image features—represented as "signature maps" based on tissue boundaries—and the corresponding PCA deformation coefficients. This allows the model to predict the necessary warp for a given brain structure.
  3. Prediction and Refinement: During the testing phase, the SVR instantly predicts an initial deformation field for a new subject. This predicted warp is applied to the template to create an Intermediate Template (IT). A standard registration algorithm is then used to resolve only the small residual differences between the IT and the subject.
Overview of the proposed registration framework, illustrating the training and testing phases.

Datasets and Evaluation

The framework was trained and validated using several neuroimaging resources:

  • Internal MR Database: 50 images for initial PCA modeling.
  • NIREP Database: 16 images with manual segmentations of 32 Regions of Interest (ROIs) for accuracy assessment.
  • OASIS Database: 50 images for supplementary testing.
  • Simulated Data: 12,500 PCA-derived samples to ensure model generalization.

Key Results

The framework was evaluated using two popular registration algorithms, demonstrating substantial improvements in efficiency:

  • Computational Speedup: Registration using the HAMMER algorithm was reduced from 90 minutes to 16 minutes (a 6X speedup). For Diffeomorphic Demons, the time dropped from 219 seconds to 71 seconds (a 3X speedup).
  • Registration Accuracy: Accuracy remained stable or improved slightly. For instance, the average ROI overlap ratio for HAMMER increased from 66% to 67%.
  • Visual Quality: The resulting average brain images were notably sharper, particularly in complex cortical regions, suggesting more precise alignment across the study cohort compared to traditional methods.

Discussion and Conclusion

The primary contribution of this work is the shift from optimizing algorithms to optimizing the starting point of the registration process. By using a fast regression model to move the "starting line" closer to the final alignment, the framework reduces the burden on iterative optimizers. Furthermore, the use of PCA-based data augmentation demonstrates that effective machine learning models can be trained even with modest initial datasets. While the current implementation relies on manual feature engineering (signature maps), the framework's generalized nature allows for future integration with modern deep learning techniques and a wider variety of registration algorithms.

Summary updated

Abstract

This paper presents a generalized learning based framework for improving both speed and accuracy of the existing deformable registration method. The key of our framework involves the utilization of a support vector regression (SVR) to learn the correlation between brain image appearances and their corresponding shape deformations to a template, for helping significantly cut down the computation cost and improve the robustness to local minima by using the learned correlation to instantly predict a good subject-specific deformation initialization for any given subject under registration. Our framework consists of three major parts: 1) training of SVR models based on the statistics of image samples and their shape deformations to capture intrinsic image-deformation correlations, 2) deformation prediction for a new subject with the trained SVR models to generate a subject-resemblance intermediate template by warping the original template with the predicted deformations, and 3) estimating of the residual deformation from the intermediate template to the subject for refined registration. Any existing deformable registration methods can be easily employed for training the SVR models and estimating the residual deformation. We have tested in this paper the two widely used deformable registration algorithms, i.e., HAMMER [1] and diffeomorphic demons [2], for demonstration of our proposed frameowrk. Experimental results show that, compared to the registration using the original methods (with no deformation prediction), our framework achieves a significant speedup (6X faster than HAMMER, and 3X faster than diffeomorphic demons), while maintaining comparable (or even slighly better) registration accuracy.
BibTeX
@inproceedings{kim2010generalized,
  author = {Minjeong Kim and Guorong Wu and Pew-Thian Yap and Dinggang Shen},
  title = {{A Generalized Learning Based Framework for Fast Brain Image Registration}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2010},
  doi = {10.1007/978-3-642-15745-5\_38},
}
Figure