2.3. Image analysis  

The image-processing workflow is schematically presented in Fig. 1 ▸(b). Images were stitched in ImageJ 1.50i (Schneider et al., 2012 ▸) using the included plugin for grid-wise or collection stitching (Preibisch et al., 2009 ▸), which uses the Fourier transform phase-correlation method (Kuglin & Hines, 1975 ▸) to find translational offsets between sets of 2D or 3D images and is included in the standard distribution of ImageJ. Alternatively, the pair-wise registration plugin MosaicJ (Thévenaz & Unser, 2007 ▸) was also used successfully but it appeared less convenient because it required a manual coarse placement of the tiles prior to stitching. The stitched images were cropped around the region of interest (LCP sample). In some images, residual stripes appeared in spite of the flat-field correction. Therefore, a Fourier transform filter was applied to suppress frequencies lower than 40 pixels, as an empirical cutoff. Although images were already interpretable by eye thanks to the phase-contrast edge-enhancement effect around the crystals, a phase-retrieval step was performed following Paganin et al. (2002 ▸) in order to ease the automation of crystal detection (see Section S2 in the Supporting information). After phase retrieval, crystals appeared as distinct solid objects of the expected shape, in moderate contrast to the background. The spherical metal beads are easily identifiable thanks to the strong absorption from the metal which gives a strong contrast. Finally, images were converted into binary images with different thresholds for the detection of crystals or beads. A particle search was performed within the region of interest (LCP sample) using the built-in functionality in ImageJ, setting appropriate ranges in particle area (in pixels) and sphericity. The particle area was set to match the expected sizes of the crystals, taking into account the effective pixel size of 0.325 µm. The sphericity was used to exclude artefacts (sphericity < 0.2) and to discriminate crystals from metal beads (sphericity > 0.9). A list of results was exported as a list of 2D coordinates in the prelocation image plane for crystals or metal beads. References were either the positions of metal beads or easily identifiable features – such as silicon nitride window corners or bigger crystals – in the absence of metal beads.