Mercurial > repos > imgteam > imagej2_skeletonize3d
view imagej2_analyze_skeleton_jython_script.py @ 0:f6df6830d5ec draft
"planemo upload for repository https://github.com/bgruening/galaxytools/tree/master/tools/image_processing/imagej2 commit b08f0e6d1546caaf627b21f8c94044285d5d5b9c-dirty"
author | imgteam |
---|---|
date | Tue, 17 Sep 2019 16:57:15 -0400 |
parents | |
children | 768825d9034a |
line wrap: on
line source
import jython_utils import math import sys from ij import IJ from sc.fiji.analyzeSkeleton import AnalyzeSkeleton_ BASIC_NAMES = [ 'Branches', 'Junctions', 'End-point Voxels', 'Junction Voxels', 'Slab Voxels', 'Average branch length', 'Triple Points', 'Quadruple Points', 'Maximum Branch Length' ] DETAIL_NAMES = [ 'Skeleton ID', 'Branch length', 'V1 x', 'V1 y', 'V1 z', 'V2 x', 'V2 y', 'V2 z', 'Euclidean distance' ] def get_euclidean_distance( vertex1, vertex2 ): x1, y1, z1 = get_points( vertex1 ) x2, y2, z2 = get_points( vertex2 ) return math.sqrt( math.pow( ( x2 - x1 ), 2 ) + math.pow( ( y2 - y1 ), 2 ) + math.pow( ( z2 - z1 ), 2 ) ) def get_graph_length( graph ): length = 0 for edge in graph.getEdges(): length = length + edge.getLength() return length def get_points( vertex ): # An array of Point, which has attributes x,y,z. point = vertex.getPoints()[ 0 ] return point.x, point.y, point.z def get_sorted_edge_lengths( graph ): # Return graph edges sorted from longest to shortest. edges = graph.getEdges() edges = sorted( edges, key=lambda edge: edge.getLength(), reverse=True ) return edges def get_sorted_graph_lengths( result ): # Get the separate graphs (skeletons). graphs = result.getGraph() # Sort graphs from longest to shortest. graphs = sorted( graphs, key=lambda g: get_graph_length( g ), reverse=True ) return graphs def save( result, output, show_detailed_info, calculate_largest_shortest_path, sep='\t' ): num_trees = int( result.getNumOfTrees() ) outf = open( output, 'wb' ) outf.write( '# %s\n' % sep.join( BASIC_NAMES ) ) for index in range( num_trees ): outf.write( '%d%s' % ( result.getBranches()[ index ], sep ) ) outf.write( '%d%s' % ( result.getJunctions()[ index ], sep ) ) outf.write( '%d%s' % ( result.getEndPoints()[ index ], sep ) ) outf.write( '%d%s' % ( result.getJunctionVoxels()[ index ], sep ) ) outf.write( '%d%s' % ( result.getSlabs()[ index ], sep ) ) outf.write( '%.3f%s' % ( result.getAverageBranchLength()[ index ], sep ) ) outf.write( '%d%s' % ( result.getTriples()[ index ], sep ) ) outf.write( '%d%s' % ( result.getQuadruples()[ index ], sep ) ) outf.write( '%.3f' % result.getMaximumBranchLength()[ index ] ) if calculate_largest_shortest_path: outf.write( '%s%.3f%s' % ( sep, result.shortestPathList.get( index ), sep ) ) outf.write( '%d%s' % ( result.spStartPosition[ index ][ 0 ], sep ) ) outf.write( '%d%s' % ( result.spStartPosition[ index ][ 1 ], sep ) ) outf.write( '%d\n' % result.spStartPosition[ index ][ 2 ] ) else: outf.write( '\n' ) if show_detailed_info: outf.write( '# %s\n' % sep.join( DETAIL_NAMES ) ) # The following index is a placeholder for the skeleton ID. # The terms "graph" and "skeleton" refer to the same thing. # Also, the SkeletonResult.java code states that the # private Graph[] graph object is an array of graphs (one # per tree). for index, graph in enumerate( get_sorted_graph_lengths( result ) ): for edge in get_sorted_edge_lengths( graph ): vertex1 = edge.getV1() x1, y1, z1 = get_points( vertex1 ) vertex2 = edge.getV2() x2, y2, z2 = get_points( vertex2 ) outf.write( '%d%s' % ( index+1, sep ) ) outf.write( '%.3f%s' % ( edge.getLength(), sep ) ) outf.write( '%d%s' % ( x1, sep ) ) outf.write( '%d%s' % ( y1, sep ) ) outf.write( '%d%s' % ( z1, sep ) ) outf.write( '%d%s' % ( x2, sep ) ) outf.write( '%d%s' % ( y2, sep ) ) outf.write( '%d%s' % ( z2, sep ) ) outf.write( '%.3f' % get_euclidean_distance( vertex1, vertex2 ) ) if calculate_largest_shortest_path: # Keep number of separated items the same for each line. outf.write( '%s %s' % ( sep, sep ) ) outf.write( ' %s' % sep ) outf.write( ' %s' % sep ) outf.write( ' \n' ) else: outf.write( '\n' ) outf.close() # Fiji Jython interpreter implements Python 2.5 which does not # provide support for argparse. error_log = sys.argv[ -8 ] input = sys.argv[ -7 ] black_background = jython_utils.asbool( sys.argv[ -6 ] ) prune_cycle_method = sys.argv[ -5 ] prune_ends = jython_utils.asbool( sys.argv[ -4 ] ) calculate_largest_shortest_path = jython_utils.asbool( sys.argv[ -3 ] ) if calculate_largest_shortest_path: BASIC_NAMES.extend( [ 'Longest Shortest Path', 'spx', 'spy', 'spz' ] ) DETAIL_NAMES.extend( [ ' ', ' ', ' ', ' ' ] ) show_detailed_info = jython_utils.asbool( sys.argv[ -2 ] ) output = sys.argv[ -1 ] # Open the input image file. input_image_plus = IJ.openImage( input ) # Create a copy of the image. input_image_plus_copy = input_image_plus.duplicate() image_processor_copy = input_image_plus_copy.getProcessor() try: # Set binary options. options = jython_utils.get_binary_options( black_background=black_background ) IJ.run( input_image_plus_copy, "Options...", options ) # Convert image to binary if necessary. if not image_processor_copy.isBinary(): IJ.run( input_image_plus_copy, "Make Binary", "" ) # Run AnalyzeSkeleton analyze_skeleton = AnalyzeSkeleton_() analyze_skeleton.setup( "", input_image_plus_copy ) if prune_cycle_method == 'none': prune_index = analyze_skeleton.NONE elif prune_cycle_method == 'shortest_branch': prune_index = analyze_skeleton.SHORTEST_BRANCH elif prune_cycle_method == 'lowest_intensity_voxel': prune_index = analyze_skeleton.LOWEST_INTENSITY_VOXEL elif prune_cycle_method == 'lowest_intensity_branch': prune_index = analyze_skeleton.LOWEST_INTENSITY_BRANCH result = analyze_skeleton.run( prune_index, prune_ends, calculate_largest_shortest_path, input_image_plus_copy, True, True ) # Save the results. save( result, output, show_detailed_info, calculate_largest_shortest_path ) except Exception, e: jython_utils.handle_error( error_log, str( e ) )