hot-knife

Hot knife FIB-SEM slab series stitching pipeline.

General protocol

TODO Improve and describe more details (parameters etc).

1. Convert input into N5 using the run-convert scripts. Modify the croping parameters to limit the export to the data needed for alignment (chop off excessive background). Optionally, generate a scale pyramid of the exported N5 datasets using https://github.com/saalfeldlab/n5-spark/blob/master/startup-scripts/spark-janelia/n5-downsample.py .
2. Extract top and botom faces using the run-extract-face scripts.
3. Calculate a 'global' affine alignment for all stacks with the run-align-p0 script.
4. Perform incrementally refined alignments with the run-align-p[1-x] scripts. Visualize intermediate results with the ViewAlignment class. No start script provided, use IDE or write one. Would be similar to the Spark run-scripts. Tune parameters if results are unsatisfying and repeat.
5. Manually tune the distances between slabs in the aligned series using the ViewAlignedSlabSeries. No start script provided, use IDE or write one. Would be similar to the Spark run-scripts.
6. Export the full dataset using the run-export script. Optionally, generate a scale pyramid of the exported N5 datasets using https://github.com/saalfeldlab/n5-spark/blob/master/startup-scripts/spark-janelia/n5-downsample.py .

Transformations

Transformations are stored as scaled inverse coordinate lookup tables at double (float64) precision. Memory layout is row-major. The fastest running dimension is the dimension index, i.e. a 2-dimensional lookup field for a 200px(width) * 100px(height) image is a tensor of 200 * 100 * 2 elements ([2, 100, 200] for NumPy or HDF5). TODO: Since coordinate lookups are typically done for all dimensions simultaneously, this layout is not CPU cache efficient and should be replaced by dimension index being the slowest running dimension.

Vector attributes are stored as column-major arrays (i.e. (x,y,z), not (z,y,x)) which is nice for ImgLib2 and Vigra. The transformation lookup table itself is also column-major, i.e. the x-coordinate lookup is the first slice, the y-coordinate lookup is the second slice, ...

Each transformation has the attributes:

• boundsMin : double precision vector, top left corner of bounding box in unscaled world coordinates,
• boundsMax : double precision vector, bottom right corner of bounding box in unscaled world coordinates,
• scale : double precision scalar, scale of the transformation (0.5 means the transformation field is half the size of the bounding box).

boundsMin and scale define an integer offset of the lookup field:

offset[i] = floor(boundsMin[i] * scale)

To make a world coordinate lookup, slice the n+1-dimensional lookup table into n single dimension lookup tables, border extend them and interpolate (I use n-linear interpolation, but other interpolations are legal). The lookup for the transferred coordinate y[i] is

y[i] = interpolatedLookup[i](x[n - 1] * scale - offset[n - 1], ..., x[0] * scale - offset[0]) / scale