60
S. Ghosh et al.
2.1 Experimental Data Acquisition and Image Processing
Acquisition of high-fidelity micro- and nano-scale 3D data for a single crystal
Rene’88DT superalloy with γ − γ microstructure is done using an automated,
high-throughput focused ion beam (FIB) coupled with a high-resolution scanning
electron microscope (SEM) in a TriBeam system [48]. The resulting dataset used in
microstructure reconstruction is available in [46]. The FIB process serially sections
layers of the specimen in the [001] direction, ablating approximately 20 nm of
material in each pass. Individual section images are subsequently extracted using the
SEM with a backscattered electron detector, repeating the process until the desired
volume is scanned. The dataset used in the generation of a 3 × 4 × 5 μm virtual
material volume is from 182 grayscale section images, each containing 1996 × 1596
pixels. This yields a resolution of 2.5 nm between pixels and 20 nm between slices.
A pipeline of automated image processing techniques is necessary to generate 3D
virtual microstructures from this dataset. This pipeline converts the experimentally
derived image stack into a fully segmented 3D voxelized representation following
a four-step procedure that includes (i) slice registration and alignment; (ii) voxel
level cleanup; (iii) feature segmentation; and (iv) artifact removal. Once a 3D virtual
representation of the experimental data is processed and segmented, statistical
distributions of microstructural descriptors, viz., feature size, shape, orientation,
neighbor distance, etc., are generated from the dataset. The fidelity of the statistical
distributions depends on the robustness of the preliminary image processing.
The collection of image slices from serial-sectioning must be assimilated into
a 3D voxelized binary structure, representing γ matrix and γ precipitate phases.
To achieve this objective, the slices in the image stack must be aligned, followed
by preliminary image processing and thresholding for binarization. Major steps
in this process include (i) image slice alignment; (ii) background subtraction; (iii)
local smoothing; (iv) contrast enhancement through unsharp masking; (v) minimum
cross-entropy thresholding; and (vi) scanning direction normalization with image
interpolation. A representative slice, corresponding to the output of the segmentation
process, is shown in Fig. 2b. The difference in data resolution between the sectioning
direction (z) (∼20 nm) and the in-plane directions (x, y) (∼2.5 nm) necessitates
additional image slices to ensure the same distance between voxels in all directions.
Consistent resolution in the x, y, and z directions allows the usage of filters that
operate with uniform spacing, such as the watershed algorithm and microstructural
statistics extraction. The aligned and binarized images should subsequently undergo
an image interpolation method to equalize section spacing in all directions. A
consistent resolution of 2.5 nm between voxels in all directions is obtained by
inserting (
d bp
d ip
− 1) additional slices between images. Here d bp = 20 nm is the
resolution between planes, and d ip = 2.5 is the in-plane resolution. The surrounding
images of the new slices are converted into distance to boundary maps, representing
the distance to the nearest boundary of a precipitate. These are then linearly
interpolated and thresholded as detailed in [30, 47].
S. Ghosh et al.
2.1 Experimental Data Acquisition and Image Processing
Acquisition of high-fidelity micro- and nano-scale 3D data for a single crystal
Rene’88DT superalloy with γ − γ microstructure is done using an automated,
high-throughput focused ion beam (FIB) coupled with a high-resolution scanning
electron microscope (SEM) in a TriBeam system [48]. The resulting dataset used in
microstructure reconstruction is available in [46]. The FIB process serially sections
layers of the specimen in the [001] direction, ablating approximately 20 nm of
material in each pass. Individual section images are subsequently extracted using the
SEM with a backscattered electron detector, repeating the process until the desired
volume is scanned. The dataset used in the generation of a 3 × 4 × 5 μm virtual
material volume is from 182 grayscale section images, each containing 1996 × 1596
pixels. This yields a resolution of 2.5 nm between pixels and 20 nm between slices.
A pipeline of automated image processing techniques is necessary to generate 3D
virtual microstructures from this dataset. This pipeline converts the experimentally
derived image stack into a fully segmented 3D voxelized representation following
a four-step procedure that includes (i) slice registration and alignment; (ii) voxel
level cleanup; (iii) feature segmentation; and (iv) artifact removal. Once a 3D virtual
representation of the experimental data is processed and segmented, statistical
distributions of microstructural descriptors, viz., feature size, shape, orientation,
neighbor distance, etc., are generated from the dataset. The fidelity of the statistical
distributions depends on the robustness of the preliminary image processing.
The collection of image slices from serial-sectioning must be assimilated into
a 3D voxelized binary structure, representing γ matrix and γ precipitate phases.
To achieve this objective, the slices in the image stack must be aligned, followed
by preliminary image processing and thresholding for binarization. Major steps
in this process include (i) image slice alignment; (ii) background subtraction; (iii)
local smoothing; (iv) contrast enhancement through unsharp masking; (v) minimum
cross-entropy thresholding; and (vi) scanning direction normalization with image
interpolation. A representative slice, corresponding to the output of the segmentation
process, is shown in Fig. 2b. The difference in data resolution between the sectioning
direction (z) (∼20 nm) and the in-plane directions (x, y) (∼2.5 nm) necessitates
additional image slices to ensure the same distance between voxels in all directions.
Consistent resolution in the x, y, and z directions allows the usage of filters that
operate with uniform spacing, such as the watershed algorithm and microstructural
statistics extraction. The aligned and binarized images should subsequently undergo
an image interpolation method to equalize section spacing in all directions. A
consistent resolution of 2.5 nm between voxels in all directions is obtained by
inserting (
d bp
d ip
− 1) additional slices between images. Here d bp = 20 nm is the
resolution between planes, and d ip = 2.5 is the in-plane resolution. The surrounding
images of the new slices are converted into distance to boundary maps, representing
the distance to the nearest boundary of a precipitate. These are then linearly
interpolated and thresholded as detailed in [30, 47].