Super-Resolution Multi-Contrast Unbiased Eye Atlases With Deep Probabilistic Refinement

Ho Hin Lee, Adam M. Saunders, Michael E. Kim, Samuel W. Remedios, Yucheng Tang, Qi Yang, Xin Yu, Shunxing Bao, Chloe Cho, Louise A. Mawn, Tonia S. Rex, Kevin L. Schey, Blake E. Dewey, Jeffrey M. Spraggins, Jerry L. Prince, Yuankai Huo, Bennett A. Landman

TL;DR

This paper introduces a novel framework for creating high-resolution, unbiased atlases of the human eye from multi-contrast MRI scans. By combining deep learning-based super-resolution with a hierarchical probabilistic registration approach, they achieve significantly improved alignment of eye organs, addressing a key challenge in medical imaging for population-wide analysis.

What Problem Does This Paper Address?

The human eye's morphology, particularly the orbit and optic nerve, exhibits considerable variation across individuals. This variability makes it challenging to create a single, standardized spatial reference (an atlas) that accurately represents the population. Existing imaging modalities often have lower resolution in one direction (through-plane) compared to another (in-plane), hindering the ability to precisely delineate anatomical structures. This lack of a robust, unbiased atlas limits the feasibility of generalizing findings from individual scans to broader population studies and can impact tasks like medical image registration and cross-sectional comparisons.

Methods

The authors propose a two-stage framework:

  1. Super-Resolution Preprocessing: They employ the SMORE (Synthetic Multi-Orientation Resolution Enhancement) algorithm, a self-supervised deep learning method. SMORE leverages the higher in-plane resolution to learn and restore details in the lower through-plane resolution, aiming to create "isotropic" (equal resolution in all directions) images. This step is crucial for enhancing the visibility and distinguishability of eye organs.
  2. Hierarchical Deep Probabilistic Registration Refinement:
    • Coarse Unbiased Template Generation: Instead of registering to a single fixed template (which can introduce bias), they use an unbiased template generation method. This involves an iterative, metric-based registration (using ANTs toolkit for rigid, affine, and deformable transformations) on a small subset of subjects to create an initial, coarse, population-averaged template.
  3. Refinement with Deep Probabilistic Networks: The coarse template is then refined using a larger population of subjects. They utilize VoxelMorph-Probabilistic, a deep learning model that predicts deformation fields as smooth and invertible (diffeomorphic) transformations. This unsupervised approach allows for more expansive deformation fields, leading to better alignment of organ boundaries. They also compare this to a non-probabilistic VoxelMorph-Original.

Datasets

The study utilized de-identified volumetric MRI scans from 1842 patients obtained from ImageVU at Vanderbilt University Medical Center. These scans were acquired in four different tissue contrasts:

  • T1-weighted pre-contrast
  • T1-weighted post-contrast
  • T2-weighted turbo-spin echo (TSE)
  • T2-weighted fluid attenuated inversion recovery (FLAIR)

For each contrast, a subset of subjects was randomly selected for template generation and evaluation. Specifically, 100 subjects were used for T2W TSE and T2W FLAIR, 100 for T1W Post-Contrast, and 44 for T1W Pre-Contrast. The ground truth labels for four organ regions (optic nerve, recti muscles, globe, and orbital fat) were provided.

Key Insights and Ideas

  • Super-resolution is key for anatomical clarity: The SMORE algorithm significantly improves the visual quality and distinctiveness of eye organs in through-plane slices, making them much easier to discern and segment.
  • Unbiased template generation is critical: Moving away from single-subject templates and towards population-averaged, unbiased templates is essential for creating representative atlases, especially given the high variability in eye morphology.
  • Deep probabilistic registration enhances alignment: The VoxelMorph-Probabilistic model, with its ability to generate smooth and invertible deformations, provides a significant improvement in aligning complex anatomical structures compared to traditional registration methods.
  • Multi-contrast atlases capture different tissue properties: Generating separate atlases for different MRI contrasts allows for capturing unique tissue characteristics and provides a more comprehensive representation of the eye.

Key Results

  • Improved Dice Scores: The proposed framework, particularly with VoxelMorph-Probabilistic, achieved statistically significant improvements in average Dice scores for label transfer across all four tissue contrasts compared to a standard ANTs registration framework. For example, for T2W FLAIR, the VoxelMorph-Probabilistic method achieved an average Dice score of 0.678 ± 0.191, compared to 0.742 ± 0.160 for ANTs alone. (Note: The paper states p < 0.001 using the Wilcoxon signed-rank test compared to ANTs alone. Bold values indicate highest mean Dice score. The table shows VoxelMorph-Probabilistic achieving the highest Dice scores in most cases).
  • Qualitative Improvements: Figure 3 visually demonstrates the enhanced anatomical context and distinctiveness of eye organs after applying SMORE. Figure 4 shows consistent registration across multiple subjects. Figure 5 highlights how SMORE refines the anatomical context and makes tissues more distinguishable. Figure 6 shows that the inverse labels registered from the atlas to subject space are qualitatively similar to the original segmentation labels.
  • Hausdorff Distance: While Dice scores showed significant improvement, the average Hausdorff distances did not show a consistently significant difference with or without super-resolution. The authors suggest this might be due to the inverse label transfer involving low-resolution subject images.

Limitations and Future Work

  • Computational Cost: Generating the coarse unbiased template using ANTs is computationally intensive, taking several days for a small sample. The authors suggest an end-to-end approach would be desirable.
  • Separate Spatial Geometries: The resulting atlases are not in registration with each other, meaning each tissue contrast has its own spatial geometry.
  • Future Work: The authors suggest that the generated atlases could be used for quantitative measurements of eye morphology across variable populations and for localizing multi-scale information within the eye for projects like HuBMAP. They also note the potential for automatic labeling of structures of interest, even in images with poor quality.

Summary updated

Abstract

Purpose: Eye morphology varies significantly across the population, especially for the orbit and optic nerve. These variations limit the feasibility and robustness of generalizing population-wise features of eye organs to an unbiased spatial reference. Approach: To tackle these limitations, we propose a process for creating high-resolution unbiased eye atlases. First, to restore spatial details from scans with a low through-plane resolution compared to a high in-plane resolution, we apply a deep learning-based super-resolution algorithm. Then, we generate an initial unbiased reference with an iterative metric-based registration using a small portion of subject scans. We register the remaining scans to this template and refine the template using an unsupervised deep probabilistic approach that generates a more expansive deformation field to enhance the organ boundary alignment. We demonstrate this framework using magnetic resonance images across four different tissue contrasts, generating four atlases in separate spatial alignments. Results: For each tissue contrast, we find a significant improvement using the Wilcoxon signed-rank test in the average Dice score across four labeled regions compared to a standard registration framework consisting of rigid, affine, and deformable transformations. These results highlight the effective alignment of eye organs and boundaries using our proposed process. Conclusions: By combining super-resolution preprocessing and deep probabilistic models, we address the challenge of generating an eye atlas to serve as a standardized reference across a largely variable population.
BibTeX
@article{lee2024super,
  author = {Ho Hin Lee and Adam M. Saunders and Michael E. Kim and Samuel W. Remedios and Yucheng Tang and Qi Yang and Xin Yu and Shunxing Bao and Chloe Cho and Louise A. Mawn and Tonia S. Rex and Kevin L. Schey and Blake E. Dewey and Jeffrey M. Spraggins and Jerry L. Prince and Yuankai Huo and Bennett A. Landman},
  title = {{Super-Resolution Multi-Contrast Unbiased Eye Atlases With Deep Probabilistic Refinement}},
  journal = {Journal of Medical Imaging},
  year = {2024},
  doi = {10.1117/1.jmi.11.6.064004},
}
Figure