NOTE: The Rhoana pipeline is still under development, and should not be considered stable.

Rhoana - Dense Automated Neuron Annotation Pipeline

Prerequisites: numpy http://numpy.org scipy http://scipy.org h5py http://www.h5py.org/ mahotas http://luispedro.org/software/mahotas OpenCV http://opencv.org/ pymaxflow https://github.com/Rhoana/pymaxflow fast64counter https://github.com/Rhoana/fast64counter CPLEX http://www.ibm.com/software/integration/optimization/cplex-optimizer/

The Rhoana pipeline operates in the following stages: Classify Membranes Segmentation Block dicing Window Fusion Pairwise Matching Local and Global Remapping

A simple driver program is in Control/driver.py. It takes as input a file containing a list of images to process. These should be aligned EM sections.

ClassifyMembranes/classify_image takes three arguments:

• image file
• classifier file (an example is in ClassifyMembranes/GB_classifier.txt)
• output HDF5 The HDF5 output will contain a single dataset, "probabilities", which are the per-pixel membrane probabilities.

Segment/segment.py takes two arguments:

• probabilities HDF5
• output segmentations HDF5 Output will contain two datasets, "segmentations" and "probabilities". The first is of size IxJxN, with I,J the image dimensions and N the number of generated segmentations (at various scale and smoothness, N = 30 in the current implementation). The "probabilities" dataset is just copied from the input.

Control/dice_block.py takes a number of arguments:

• imin, jmin, imax, jmax - the IJ coordinates of the block
• output.hdf5
• K segmentation HDF5 files This will cut out a block as: np.concat([im[imin:imax, jmin:jmax, :] for im in segs[K]], 4) (and a similar block for the per-pixel probabilities) It will produce two datasets, "segmentations" and "probabilities". Segmentation

WindowFusion/window_fusion_cpx.py takes two arguments:

• input block.hdf5
• output fusedblock.hdf5 This will run window fusion to reduce the IxJxNxK block to a labeled IxJxK block. Two datasets are produced, "labels" and "probabilities".

PairwiseMatching/pairwise_match.py takes 6 arguments

• two input fused blocks
• the direction they overlap (X = 1, Y = 2, Z = 3) # this may be inaccurate, currently
• the number of pixels they overlap
• two output HDF5 fused blocks Pairwise matching produces "labels", "probabilities", and "merges" datasets. The first block should always be closer to 0,0,0. The usual method is to run it first for all X-even blocks matching to their X+1 neighbor, then all X-odd blocks matching to their X+1 neighbor, then do the same for Y, then Z. After Pairwise Matching, overlapping regions should be consistent. "merges" is Lx2, with each row indicating that two labels should be merged in the final result.

(There is a similar, program pairwise_match_region_growing.py, that uses region growing in the probability maps for overlapping regions.)

Relabelabeling/concatenate_joins.py takes multiple matches blocks and extracts their merges, and Relabelabeling/create_global_map.py processes the full list of merges to create the final remap function. Relabeling/remap_block.py takes this global remap and a single block, and produces the remapped block.

Relabeling/extract_label_plane.py takes the following arguments:

• the output hdf5 path
• its IxJ size (same as the original image)
• a Z offset for the plane within the input blocks
• a set of (ibase, jbase, input block HDF5) Extract Label Plane performs rougly the following action: for ibase, jbase, infile in args: input_data = infile['labels'][:, :, Z] output_labels[ibase:ibase+input_data.shape[0], jbase:jbase+input_data.shape[1]] = intput_data