Microstructure of a freeze-dried aqueous solution of maltodextrin DE 12 (15 % w/w)

X-ray microtomography (µ-CT) data of the dried microstructure of an aqueous maltodextrin DE 12 solution with 15 % (w/w) solid content. The sample was freeze-dried in-situ inside the tomograph and scanned at high resolution after the completion of primary drying. The dataset contains a cuboid section of the reconstructed tomogram, both as reconstructed grayscale data and as a binarized (solid/void) image that resolves the pore space of the dried cake.

The measurement is part of the experimental campaign published in Gruber et al. (2024), see Related work.

Data files

The dataset includes 2 files, which are as follows:

• img@raw_res@790x790x650_vxl@3.87592x3.87592x3.87592_uint16.raw.zip
  • Contents: Reconstructed grayscale image data of the microstructure of a freeze-dried aqueous maltodextrin DE 12 solution of 15 % (w/w) solid content
  • Interpretation: voxel value = reconstructed X-ray attenuation (higher value = higher attenuation, i.e. solid)
  • Resolution: 790 × 790 × 650 voxels
  • Voxel size: 3.87592 µm (in each dimension)
  • Encoding: uint16
  • Archive contents: single headerless binary file img@raw_res@790x790x650_vxl@3.87592x3.87592x3.87592_uint16.raw (811 330 000 bytes uncompressed)
• img@binarized_res@790x790x650_vxl@3.87592x3.87592x3.87592_uint8.raw.zip
  • Contents: Binarized image data of the same section
  • Interpretation: solid = 1, voidage = 0
  • Resolution: 790 × 790 × 650 voxels
  • Voxel size: 3.87592 µm (in each dimension)
  • Encoding: uint8
  • Archive contents: single headerless binary file img@binarized_res@790x790x650_vxl@3.87592x3.87592x3.87592_uint8.raw (405 665 000 bytes uncompressed)

Both files describe the identical section and share the same voxel grid, so the binarized image can be used directly as a mask for the grayscale image. The section covers a physical volume of approximately 3.06 × 3.06 × 2.52 mm³.

Data format

The .raw files are plain binary dumps without any header, padding or metadata. All information needed to read them is encoded in the file name:

img@<type>_res@<Nx>x<Ny>x<Nz>_vxl@<Lx>x<Ly>x<Lz>_<class>.raw

with the voxel edge lengths <Lx>, <Ly>, <Lz> given in µm and <class> naming the integer type of a single voxel.

• Byte order: little-endian
• Voxel order: column-major (Fortran order), i.e. the first index varies fastest, followed by the second and then the third

Reading the data in MATLAB

res   = [790, 790, 650];
fname = 'img@binarized_res@790x790x650_vxl@3.87592x3.87592x3.87592_uint8.raw';

fid = fopen(fname, 'r');
img = fread(fid, prod(res), '*uint8');
fclose(fid);

img = reshape(img, res);

For the grayscale file, use '*uint16' instead of '*uint8'.

Reading the data in Python

import numpy as np

res = (790, 790, 650)
fname = "img@binarized_res@790x790x650_vxl@3.87592x3.87592x3.87592_uint8.raw"

img = np.fromfile(fname, dtype=np.uint8).reshape(res, order="F")

For the grayscale file, use dtype=np.uint16.

Sample preparation and freeze-drying

Maltodextrin (Glucidex 12D, Roquette, France) was dissolved in demineralized water to a solid content of 15 % (w/w) and stirred for 1 h at 300 rpm at ambient temperature.

Freezing and drying were carried out in-situ, inside the tomograph, using the freeze-drying stage described by Hilmer et al. (2020), which holds a sample of 7 mm diameter and 5 mm height. The sample was frozen directly in the stage at a cooling rate of 6 K/min, held at 10 °C for 10 min to equalize the initial conditions, then brought to the target shelf temperature of −11 °C and held for 30 min before drying. Primary drying was performed at a chamber pressure of 10 Pa and continued until all ice had sublimed. No annealing step was applied.

The drying process itself was followed by time-resolved scans at a voxel size of 15.5 µm for about 2 h (not part of this dataset). The high-resolution scan published here was acquired directly afterwards, when all ice had sublimed.

Image acquisition

The measurement was carried out on a TESCAN DynaTOM system (TESCAN XRE, Ghent, Belgium), in which a horizontally rotating gantry carries source and detector so that the in-situ stage remains static during acquisition.

Reconstruction was performed with the Panthera software (TESCAN XRE, Ghent, Belgium) using an FDK filtered back-projection algorithm. The settings recorded for this particular reconstruction are a Shepp-Logan filter with a filter attenuation of 0.3, a zero-padding factor of 3, bilinear interpolation, a scatter correction of 0.01 and no beam-hardening correction.

The reconstructed tomogram has a resolution of 1935 × 1935 × 741 voxels. The published section is a cuboid crop of that tomogram of 790 × 790 × 650 voxels. Its first voxel is voxel (567, 569, 1) of the tomogram (1-based indices, i.e. a zero-based offset of (566, 568, 0)), so the section spans the tomogram voxels 567–1356 along the first axis, 569–1358 along the second axis and 1–650 along the third axis. It was placed inside the dried cake so that it is free of sample boundaries. The rotation axis runs along the third axis through voxel (402, 400) of the section (1-based), i.e. close to its centre. The grayscale data contain weak ring artifacts around this axis, which were removed during binarization (see Image processing).

Image processing

The lamellar pore walls are only a few voxels thick, and the image contrast varies with height. The binarized image was therefore derived from the grayscale image by comparing every voxel with its local background, scaled by the local contrast, so that a single threshold is valid for the whole section (MATLAB Image Processing Toolbox):

1. Gaussian denoising with a standard deviation of 1 voxel (imgaussfilt3), which yields the image G
2. Subtraction of the local background, D = G − G_bg, where G_bg is G smoothed with a Gaussian of 6 voxels standard deviation (imgaussfilt3)
3. Ring artifact removal: for each slab of 10 slices along the third axis, the median of D over all voxels at the same distance from the rotation axis (radial bins of 0.5 voxels) gives a radial profile. The narrow part of this profile (profile minus its moving median over 10 voxels) is subtracted from D.
4. Normalization by the local contrast, N = D / D_rms, where D_rms is the square root of D² smoothed with a Gaussian of 18 voxels standard deviation (imgaussfilt3)
5. Global thresholding: voxels with N > 0.6 are solid
6. Removal of solid objects smaller than 50 voxels (26-connectivity, bwareaopen) and filling of void objects smaller than 50 voxels (6-connectivity, bwareaopen)

The resulting binarized image has a solid volume fraction of 27.8 %. The largest connected solid object (26-connectivity) contains 99.8 % of all solid voxels, i.e. the solid forms one connected network apart from a few small isolated objects.

No downsampling was applied at any point, so the binarized image is on the native voxel grid of the reconstruction.

Note that this binarization differs from the Python-based workflow used for the morphological analysis in Gruber et al. (2024), which applied Otsu thresholding followed by a morphological opening. The binarized image published here is therefore not identical to the one underlying the pore statistics reported in that article.

Related work

The freeze-drying experiments and the tomographic measurement campaign this sample belongs to are described in:

Gruber, S., Greiner, J., Eppink, A., Thomik, M., Coppens, F., Vorhauer-Huget, N., Tsotsas, E., & Foerst, P. (2024). Pore shape matters – In-situ investigation of freeze-drying kinetics by 4D XCT methods. Food Research International, 193, 114837. https://doi.org/10.1016/j.foodres.2024.114837