An automatic image-based registration method for correlative light-electron microscopy
Correlative light-electron microscopy (CLEM) enables to combine information on cell dynamics, studied with light microscopy (LM) techniques, and cell ultrastructure provided by electron microscopy systems (EM), for a better understanding of cell mechanisms.
Registration of LM and EM modalities is an open and difficult problem since LM and EM images are very different both in field-of-view, pixel size, image size, and appearance. We will present a user-friendly image-based automatic registration method to overlay LM and EM images.
It comprises three steps: 1) Laplacian of Gaussian (LoG) representation of images with an adaptive associated scale (or blurring), which provides more comparable appearance for the LM and EM images; 2) Search of the region corresponding to the LM region of interest (ROI) in the EM image (or the other way around), using a patch-based exhaustive search method. Several similarity criteria have been compared, based on the LoG-value histograms and Local Directional Pattern (LDP) features; 3) Pre-registration of LM and EM images using the shift component given by the previous step. The rotation between LM and EM images can then be computed in two different ways: if there is a visible alignment of elements in the image, an axis can be fitted through these elements in both LM and EM and the angle between the axis is computed. Otherwise, the rotation angle is estimated through an exhaustive search using mutual information. The registration is finally completed by a refinement using mutual information and an affine geometric transformation to overlay both images. This approach is able to locate the LM-ROI in the EM image, or conversely the EM patch in the LM image, as validated by experiments performed on real 2D CLEM image sets supplied by Institut Curie. The overall method requires no parameter tuning and no specific knowledge to be used. The user has just to specify the bounding box of the ROI in the LM (or EM) image as input. We will also present preliminary results on 3D CLEM images.