Volume registration of in vivo confocal microscopy datasets of the human cornea - a comparison of two novel approaches

Introduction
In vivo corneal confocal microscopy (CCM) with a computer-controlled focus drive can be used for volume reconstructions of the corneal tissue. So far, all volumetric approaches remain limited to 3D reconstructions from single CCM focus stacks. This contribution presents two approaches for laterally expanded volume reconstruction from CCM image sequences where the focus oscillates along a triangular wave function.
Methods
Both proposed methods initially require an exhaustive, pairwise, correlation-based 2D image registration to eliminate motion artifacts in the recorded CCM images. The continuous, motion-corrected image sequence is separated into stacks, each representing a single rising or falling part of the oscillation. The first method implements a correlation-based 3D registration. The separate image stacks are transformed into (motion-corrected) volume images and then registered pairwise. The second method infers information on the relative image depth from the exhaustive 2D registration process. The underlying assumption is that overlapping image pairs from the same tissue depth yield higher correlation values than image pairs recorded at differing tissue depths.
... mehr
Results
The proposed methods were each applied to 40 oscillating focus datasets of the corneal epithelium of 11 healthy individuals, each dataset containing approx. 10 full focus oscillations. The 3D registration results of the first approach prove to be very robust, but the 3D correlation computations are extremely time-consuming (116.3±15.0 minutes per dataset; MATLAB implementation on Intel Core i7-6700 CPU). The second approach is much faster (6.1±4.7 minutes per dataset), but less reliable. A threshold is therefore required to exclude incorrect depth inference results, which necessarily excludes a substantial portion of correct results as well.
Conclusion
Laterally expanded volume reconstruction from CCM image sequences with oscillating focus is feasible but time-consuming. The 3D correlation-based approach may be used in research settings to examine the diagnostic potential of such registered volumes with cellular to sub-cellular resolution.