Robust registration under large image misalignment using an iterative step-aware transformer with application to...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Corneal confocal microscopy (CCM) provides high-resolution images of the corneal subbasal nerve plexus (SNP).
• Manual raster scanning for CCM mosaicking often results in misalignment and limited visualization of larger anatomical landmarks like the corneal whorl.
• Existing deep learning methods struggle with the large misalignments common in CCM imaging.

Purpose of the Study:

• To develop a robust image registration method for CCM to enable high-quality mosaic construction.
• To improve the visualization of corneal structures for enhanced disease differentiation.

Main Methods:

• Proposed a hybrid cross- and self-attention transformer-based registration network to model long-range spatial correspondences.
• Introduced the Iterative Step-Aware Transformer Registration Network (ISATR-Net) for progressive misalignment reduction.
• Curated a new CCM dataset with 1,349 image pairs from 91 patients across various corneal conditions.

Main Results:

• ISATR-Net demonstrated robust affine alignment even under challenging conditions with large misalignments.
• The proposed method outperformed state-of-the-art affine registration techniques on the curated CCM dataset.
• ISATR-Net effectively enhances mosaic construction for improved anatomical landmark visualization.

Conclusions:

• ISATR-Net offers a significant advancement in automated image registration for CCM.
• The developed method addresses limitations of existing deep learning approaches for CCM mosaicking.
• This work facilitates more comprehensive analysis of corneal pathologies through improved imaging.