Unsupervised Registration Refinement for Generating Unbiased Eye Atlas.

Area of Science:

• Medical Imaging
• Anatomy
• Computer Vision

Background:

• Demographic variations complicate eye organ localization in large-scale imaging surveys.
• Accurate anatomical feature identification is crucial for population-based eye studies.

Purpose of the Study:

• To develop a robust method for stable registration and transfer of eye organ data across diverse populations.
• To create an unbiased eye atlas template for generalized eye organ context.

Main Methods:

• Generation of an unbiased eye atlas template using an iterative approach on 20 subject scans.
• Hierarchical coarse-to-fine registration pipeline utilizing metric-based registration and a deep probabilistic network for unsupervised refinement.
• Validation using computed tomography (CT) scans from 100 de-identified subjects.

Main Results:

• The developed pipeline achieved stable transfer of eye organs, well-localized in a high-resolution atlas space.
• Demonstrated a significant 2.37% improvement in Dice score for inverse label transfer performance.
• Qualitative representations confirmed accurate transfer of organ context and generalization of morphological variations.

Conclusions:

• The unbiased eye atlas and hierarchical registration method effectively address anatomical variability in eye imaging.
• This approach enhances the localization of eye organs for population analysis and morphological studies.
• The method shows applicability in generalizing morphological variations across patient cohorts.