Super-resolution multi-contrast unbiased eye atlases with deep probabilistic refinement.

Area of Science:

• Medical imaging
• Computational anatomy
• Ophthalmology

Background:

• Significant variations in human eye morphology, particularly the orbit and optic nerve, hinder the creation of generalized spatial references.
• Existing methods struggle to accurately represent diverse population-specific eye features in a standardized atlas.

Purpose of the Study:

• To develop a robust process for generating high-resolution, unbiased eye atlases.
• To overcome limitations in generalizing population-wise eye organ features to a common spatial reference.

Main Methods:

• Applied deep learning-based super-resolution to enhance low-resolution scan details.
• Generated an initial unbiased reference using iterative metric-based registration.
• Refined the template with an unsupervised deep probabilistic approach for improved organ boundary alignment.

Main Results:

• Demonstrated the framework using magnetic resonance images across four tissue contrasts.
• Achieved significant improvement in average Dice scores for organ alignment compared to standard registration.
• Highlighted effective alignment of eye organs and boundaries through the proposed process.

Conclusions:

• Successfully addressed the challenge of creating standardized eye atlases for variable populations.
• Combined super-resolution preprocessing and deep probabilistic models for enhanced atlas generation.
• The developed framework provides a valuable tool for consistent analysis of ocular structures.