Introduction
In medical imaging, an atlas serves as a standardized reference for comparing anatomy across individuals. However, creating an accurate atlas for the human eye is challenging due to significant morphological variability across populations—influenced by factors such as sex, ethnicity, and conditions like myopia. Furthermore, clinical MRI scans are typically anisotropic, offering high resolution in the axial plane but suffering from significant "through-plane" blurriness in coronal and sagittal views. This anisotropy makes mapping small, critical structures like the optic nerve or recti muscles extremely difficult.
Methodology
The researchers developed a three-stage "coarse-to-fine" framework to generate high-resolution, unbiased eye atlases.
1. Super-Resolution (SMORE)
To address resolution imbalance, the pipeline utilizes the Synthetic Multi-Orientation Resolution Enhancement (SMORE) algorithm. This self-supervised deep learning model uses high-resolution in-plane slices to learn how to restore missing details in low-resolution through-plane slices, effectively "hallucinating" detail without requiring external high-resolution ground truth data.
2. Coarse Unbiased Template Generation
Using the Advanced Normalization Tools (ANTs) toolkit, the team performed iterative metric-based registration on a subset of 25 subjects. This process averages individual anatomies into a common space, creating an initial "rough draft" atlas free from single-subject bias.
3. Deep Probabilistic Refinement
To scale the atlas to a larger population (75 additional subjects) efficiently, the framework employs VoxelMorph-Probabilistic. This unsupervised model predicts deformation fields to align individual scans with the template, refining organ boundaries and ensuring the atlas represents broader population variability.
Datasets and Evaluation
The study leveraged a dataset of 1,842 de-identified volumetric scans from the ImageVU repository at Vanderbilt University Medical Center. Cohorts were curated for four MRI tissue contrasts: T1-weighted (pre- and post-contrast), T2-weighted Turbo-Spin Echo (TSE), and T2-weighted Fluid-Attenuated Inversion Recovery (FLAIR). Ground truth labels included the optic nerve, recti muscles, globe, and orbital fat.
Key Results and Insights
- Quantitative Gains: Deep probabilistic refinement significantly improved Dice scores (a measure of spatial overlap) across all contrasts compared to standard registration (p < 0.001). For T2W FLAIR, the average Dice score reached 0.681.
- Visual Clarity: Super-resolution preprocessing dramatically improved the definition of the optic nerve and recti muscles in coronal and sagittal views.
- Hybrid Efficiency: By combining traditional registration (ANTs) for initial templates with deep learning (VoxelMorph) for large-scale refinement, the authors achieved a balance of unbiased accuracy and computational speed.
- Multi-Contrast Utility: The creation of separate atlases for different contrasts is vital, as different MRI sequences highlight distinct tissues (e.g., fat vs. fluid), allowing researchers to select the most appropriate tool for specific anatomical studies.
Conclusion and Future Work
While the current framework produces high-quality atlases, a limitation remains: the four contrast-specific atlases are not yet registered to each other. Future work aims to develop end-to-end deep learning approaches to streamline the initial template generation, which currently remains a computational bottleneck. These atlases are slated for integration into the Human BioMolecular Atlas Program (HuBMAP) to support cellular-level mapping of the eye.
Summary updated