Hybrid Atlas Building with Deep Registration Priors

Nian Wu, Jian Wang, Miaomiao Zhang, Guixu Zhang, Yaxin Peng, Chaomin Shen

Abstract

Registration-based atlas building often poses computational challenges in high-dimensional image spaces. In this paper, we introduce a novel hybrid atlas building algorithm that fast estimates atlas from large-scale image datasets with much reduced computational cost. In contrast to previous approaches that iteratively perform registration tasks between an estimated atlas and individual images, we propose to use learned priors of registration from pre-trained neural networks. This newly developed hybrid framework features several advantages of (i) providing an efficient way of atlas building without losing the quality of results, and (ii) offering flexibility in utilizing a wide variety of deep learning based registration methods. We demonstrate the effectiveness of this proposed model on 3D brain magnetic resonance imaging (MRI) scans.

Introduction

In medical imaging, an atlas serves as a representative average template of a population's anatomy. It is a fundamental tool for structural segmentation and longitudinal studies of disease progression. However, constructing an unbiased atlas is traditionally a computationally intensive task. It requires an iterative process of registering every individual image in a dataset to a dynamic average. Given the high dimensionality of medical images, this gradient-based optimization can take hours or days for large cohorts. While deep learning (DL) offers potential speedups, many existing DL models lack the flexibility to adapt to new datasets without extensive retraining.

Methods: The HAB Framework

The proposed Hybrid Atlas Building (HAB) framework bridges the gap between classical optimization and deep learning. The core innovation is the use of pre-trained registration networks as "priors" within a joint objective function.

The HAB framework iteratively updates the atlas using a pre-learned registration prior and a closed-form update equation.

The algorithm operates through an alternating minimization strategy:

  1. Registration Step: A pre-trained DL model (e.g., VoxelMorph, Quicksilver) predicts the transformation fields between the current atlas and the subject images.
  2. Atlas Update Step: A closed-form mathematical solution updates the atlas image based on the predicted transformations.

This modular approach allows the framework to be "agnostic" to the specific registration architecture used, enabling researchers to "plug and play" the latest state-of-the-art registration models without modifying the underlying mathematical framework.

Experimental Evaluation

The researchers evaluated the HAB model using 230 T1-weighted 3D brain MRI scans from the OASIS dataset. Images were preprocessed to $128^3$ resolution. The dataset was split into 130 volumes for training the registration priors and 100 volumes for testing the atlas-building process.

Results and Discussion

The primary advantage of the HAB framework is its computational efficiency. By replacing the iterative gradient-based registration search with a single-pass DL inference, the primary bottleneck of atlas construction is eliminated.

Comparison of GPU computation time across different atlas building methods, demonstrating the significant speedup achieved by HAB variants.

As shown in the performance comparison, HAB variants achieve orders-of-magnitude speedups compared to traditional methods like FLASH or iterative DL methods like LagoMorph. Despite this speed, anatomical accuracy remains high. Quantitative metrics, such as Normalized Cross-Correlation (NCC), remained in the 0.90 to 0.92 range, comparable to slower state-of-the-art methods. Qualitatively, the resulting atlases are sharp and preserve intricate anatomical details, avoiding the blurriness common in other fast approximation techniques.

Conclusion

The HAB framework provides a scalable and flexible solution for rapid atlas generation. Future work will focus on generalizing these priors to other anatomical regions beyond the brain and scaling the framework to handle full-resolution clinical scans. Additionally, the framework's modularity suggests potential applications in multi-atlas building and longitudinal anatomical modeling.

Summary updated

Abstract

Registration-based atlas building often poses computational challenges in high-dimensional image spaces. In this paper, we introduce a novel hybrid atlas building algorithm that fast estimates atlas from large-scale image datasets with much reduced computational cost. In contrast to previous approaches that iteratively perform registration tasks between an estimated atlas and individual images, we propose to use learned priors of registration from pre-trained neural networks. This newly developed hybrid framework features several advantages of (i) providing an efficient way of atlas building without losing the quality of results, and (ii) offering flexibility in utilizing a wide variety of deep learning based registration methods. We demonstrate the effectiveness of this proposed model on 3D brain magnetic resonance imaging (MRI) scans.
BibTeX
@inproceedings{wu2022hybrid,
  author = {Nian Wu and Jian Wang and Miaomiao Zhang and Guixu Zhang and Yaxin Peng and Chaomin Shen},
  title = {{Hybrid Atlas Building with Deep Registration Priors}},
  booktitle = {ISBI},
  year = {2022},
  doi = {10.1109/isbi52829.2022.9761670},
}
Figure