The CellOrganizer project provides tools for
- learning generative models of cell organization directly from images
- storing and retrieving those models
- synthesizing cell images (or other representations) from one or more models
In this section of the tutorial we are going to use CellOrganizer to
- train a framework (cell and nuclear membrane) model from simple 3D shapes
- synthesize an instance from the trained model and export as a Wavefront .obj file
First we use a CellOrganizer helper function to generate a collection of toy images. These simply geometries can then be used to train a generative model.
The first step is to seed the random number generator to obtain reproducibly-random instances.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% SEED EXAMPLE
seed = 3;
Stream.create( 'mt19937ar', 'seed', seed );
RandStream.setDefaultStream( state );
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
The helper method generate_ellipsoid returns a 3D array with centered ellipsoid.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% GENERATE SIMPLE GEOMETRIES
if ~exist( './synthetic_images' )
mkdir( './synthetic_images' );
end
for i=1:1:25
options.image_size = 256;
a = randi([1/4*options.image_size 1/2*options.image_size]);
b = randi([1/4*options.image_size 1/2*options.image_size]);
c = randi([1/4*options.image_size 1/2*options.image_size]);
filename = ['./synthetic_images/synthetic' num2str(sprintf('%04d',i)) '_0.tif'];
if ~exist( filename )
disp(['Making image ' filename]);
img = generate_ellipsoid(a,b,c,options);
img2tif( img, filename, 'lzw' );
end
a = randi([1/2*options.image_size 3/4*options.image_size]);
b = randi([1/2*options.image_size 3/4*options.image_size]);
c = randi([1/2*options.image_size 3/4*options.image_size]);
filename = ['./synthetic_images/synthetic' num2str(sprintf('%04d',i)) '_1.tif'];
if ~exist( filename )
disp(['Making image ' filename]);
img = generate_ellipsoid(a,b,c,options);
img2tif( img, filename, 'lzw' );
end
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
The resulting image collection will look something like this (this should be edited to remove some whitespace. Also, the subimages should be larger; it's hard to see the differences between each subimage.)
For convenience we will provide you with the image collection.
Now, we use img2slml, the main model training function. img2slml takes five inputs:
- a flag for the dimensionality of the data (2D/3D)
- nuclear images
- cell images
- (optional) protein images
- options used to change default model settings.
We can use the latter method to train a nuclear and cell membrane model from those geometries. The next block shows how to train the model (what is the latter method? what kind of model are we training?)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TRAIN MODEL WITH SIMPLE GEOMETRIES
clear options
% location of TIF files and resolution
dna = './synthetic_images/*_0.tif';
cell = './synthetic_images/*_1.tif';
options.model.resolution = [0.049, 0.049, 0.2000];
%training flag is set to framework because want to train a model for
%nuclear and cell shape
options.train.flag = 'framework';
%train model at full resolution
options.downsampling = [1 1 1];
%combination of supported model class and type
options.nucleus.name = 'ellipsoids';
options.nucleus.class = 'nuclear_membrane';
options.nucleus.type = 'cylindrical_surface';
%combination of supported model class and type
options.cell.name = 'ellipsoids';
options.cell.class = 'cell_membrane';
options.cell.type = 'ratio';
%output filename
options.model.filename = 'model.mat';
%helper options
options.verbose = true;
options.debug = true;
%CellOrganizer call
img2slml( '3D', dna, cell, [], options );
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
img2slml automatically saves the trained model's parameter structure in a reasonable default location. A save-path can be passed in the options struct.
>> ls model.mat
model.mat
model.mat contains nuclear and cell membrane models
>> load model
>> model.nuclearShapeModel
ans =
name: 'ellipsoids'
surface: [1x1 struct]
height: [1x1 struct]
type: 'cylindrical_surface'
resolution: [0.0490 0.0490 0.2000]
id: ''
>> model.cellShapeModel
ans =
name: 'ellipsoids'
meanShape: [1x1 struct]
modeShape: [1x1 struct]
eigens: [1x1 struct]
cellnucratio: [1x1 struct]
nucbottom: [1x1 struct]
type: 'ratio'
resolution: [0.0490 0.0490 0.2000]
id: ''
The trained model can then be used to generate synthetic instances.
slml2img is the main function for image synthesis. This function takes two inputs
- a cell array containing paths to the models from which we want to synthesize
- options used to change default synthesis settings
We can use the latter method to sample from the distributions and a synthesize nuclear and cell membrane instance. Next, we synthesize an image: (not sure what the latter method is. what are 'the distributions'?)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% SYNTHESIZE IMAGE FROM MODEL
model_file_path = './model.mat';
clear options
options.targetDirectory = pwd;
%output folder name
options.prefix = 'examples';
%number of images to synthesize
options.numberOfSynthesizedImages = 1;
%save images as TIF files
options.output.tifimages = true;
%compression for TIF output
options.compression = 'lzw';
%do not apply point-spread-function
options.microscope = 'none';
%render Gaussian objects as discs
options.sampling.method = 'disc';
%overlap frequency model and generate a single object
options.numberOfGaussianObjects = 1;
%generate framework
options.synthesis = 'framework';
%generate Wavefront obj. files
options.output.blenderfile = true;
options.output.blender.downsample = [1 1 1];
%helper options
options.verbose = true;
%CellOrganizer call
slml2img( {model_file_path}, options );
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Note the flags for generating .obj files:
%generate Wavefront obj. files
options.output.blenderfile = true;
options.output.blender.downsample = [1 1 1];
which tell CellOrganizer to save the synthetic image as indexed polygon meshes that can be imported into Blender for use with CellBlender.
The files will be saved in the default examples folder
>> ls examples
cell1
>> ls examples/cell1
cell.mtl cell.tif nucleus.obj
cell.obj nucleus.mtl nucleus.tif
The synthetic image will look like:
