Space-Frequency Detail-Preserving Construction of Neonatal Brain Atlases

Yuyao Zhang, Feng Shi, Pew-Thian Yap, Dinggang Shen

Abstract

Standard brain atlases often suffer from "blurriness" because the process of averaging multiple images acts as a low-pass filter, wiping out fine anatomical details. This paper introduces a framework that reconstructs neonatal brain atlases in the space-frequency domain using wavelet transforms and sparse patch-based representation, resulting in sharper, more detailed templates that significantly improve image normalization.

Introduction

In neuroimaging, a "brain atlas" is a standard template used to compare different subjects. Creating one usually involves registering many individual brain scans to a common space and averaging them. However, because no two brains are identical and registration is never perfect, this averaging creates a "fuzzy" result, especially in the complex folds of the neonatal cortex.

Neonatal brains are particularly difficult to map because they have low spatial resolution and low tissue contrast compared to adult brains. Existing methods mostly try to fix this by improving the registration (how images are aligned), but they still use simple weighted averaging for the final fusion (how images are combined). This paper argues that the fusion step itself is where the detail is lost and proposes a way to "reconstruct" those lost high-frequency details.

Methods

The proposed framework moves away from simple averaging and instead treats atlas construction as a signal reconstruction problem across two domains:

  1. Frequency Domain (Wavelet Transform): A 3D wavelet transform decomposes images into different scales and orientations. Fine anatomical structures, such as cortical folds, become much more apparent in specific high-frequency subbands. By processing these subbands individually, the method specifically targets and preserves fine details.
  2. Space Domain (Sparse Patch Reconstruction): The algorithm operates on small 3D patches ($6 \times 6 \times 6$ voxels). For every location in the atlas, a "dictionary" of corresponding patches from all 73 subjects is constructed. Sparse Representation is then used to find the best combination of these patches to represent the underlying anatomy without the blur.
  3. Spatial Consistency: To ensure that adjacent patches are not disjointed, multi-task LASSO is applied. This mathematical constraint forces neighboring patches to have similar sparse coefficients, ensuring the final brain image is smooth and continuous.
Overview of the proposed space-frequency detail-preserving atlas construction framework.

Experimental Setup

The construction dataset consisted of 73 healthy neonatal subjects (42 males / 31 females), scanned at approximately 24 days old using T2-weighted MRI. For evaluation, a separate group of 20 healthy neonatal subjects (10 males / 10 females) scanned at 37–41 gestational weeks was used to test the atlas's performance in image normalization. The results were compared against four state-of-the-art neonatal atlases (Atlas A-D) and a baseline "simple averaging" atlas.

Results and Discussion

Qualitative inspection reveals that the proposed atlas has much clearer cortical regions and sharper boundaries between gray and white matter compared to existing templates. Quantitatively, the proposed method achieved significantly higher Dice Ratios for Gray Matter (GM), White Matter (WM), and Cerebrospinal Fluid (CSF) when used to align the test subjects. Statistical significance was confirmed via two-sample t-tests (p < 0.001).

Visual comparison of the proposed atlas against existing neonatal atlases (Atlas A-D) and simple averaging.

The core insight is that standard atlas construction acts as a low-pass filter. By operating in the wavelet domain, high-frequency "detail" is treated as a signal to be preserved rather than noise to be smoothed. By building a dictionary from multiple subjects for every voxel location, the model can "pick and choose" the most representative structural features, effectively "de-blurring" registration errors.

Conclusion

This framework demonstrates that preserving high-frequency details in the space-frequency domain leads to superior neonatal brain atlases. While computationally more expensive than simple averaging, the gains in visual clarity and normalization accuracy are substantial. Future work may extend this approach to spatio-temporal atlases to track early brain development.

Summary updated

Abstract

Brain atlases are an integral component of neuroimaging studies. However, most brain atlases are fuzzy and lack structural details, especially in the cortical regions. In particular, neonatal brain atlases are especially challenging to construct due to the low spatial resolution and low tissue contrast. This is mainly caused by the image averaging process involved in atlas construction, often smoothing out high-frequency contents that indicate fine anatomical details. In this paper, we propose a novel framework for detail-preserving construction of atlases. Our approach combines space and frequency information to better preserve image details. This is achieved by performing reconstruction in the space-frequency domain given by wavelet transform. Sparse patch-based atlas reconstruction is performed in each frequency subband. Combining the results for all these subbands will then result in a refined atlas. Compared with existing atlases, experimental results indicate that our approach has the ability to build an atlas with more structural details, thus leading to better performance when used to normalize a group of testing neonatal images.
BibTeX
@inproceedings{zhang2015space,
  author = {Yuyao Zhang and Feng Shi and Pew-Thian Yap and Dinggang Shen},
  title = {{Space-Frequency Detail-Preserving Construction of Neonatal Brain Atlases}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2015},
  doi = {10.1007/978-3-319-24571-3\_31},
}
Figure