Optimization-Based Artifact Correction for Electron Microscopy Image Stacks

Copyright (C) {2014} {Samaneh Azadi, Jeremy Maitin-Shepard, Pieter Abbeel}

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 3 of the License.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.

The research is to be cited in the bibliography as:

1)S. Azadi, J. Maitin-Shepard, P. Abbeel, "EMISAC software release", URL: http://rll.berkeley.edu/2014_ECCV_EMISAC/

2)S. Azadi, J. Maitin-Shepard, P. Abbeel, "Optimization-Based Artifact Correction for Electron Microscopy Image Stacks", Proceedings of the European Conference on Computer Vision, 2014.

This code is developed in python to eliminate the artifacts in the electron microscopy image stacks most notably along the section axes,and produce an image with a higher quality. These artifacts are the result of variations in section thickness, sample deformations, and variations in imaging conditions.

Another application of this code is in time-lapse photography where the light correction is needed.

The first version 1.0 (June 2014) is developed by Samaneh Azadi in a colaboration with Jeremy Maitin-Shepard in the Robotic Learning Lab, EECS department, UC Berkeley under the supervision of Prof. Pieter Abbeel.

For any question about the algorithm or the code, feel free to email sazadi156@gmail.com .

This code gets the original 3D image stack and several parameters as the input and produces the normalized image as an output which will be saved in the path specified by the user. A short description of the method is as follows:

1. The 3D input image is downsampled in x- and y- directions as a coarse estimation in the optimization algorithm.

2. The 3D image is divided into the equal sized 3D partitions, and these partitions are considered as seperate images optimized in parallel. The number of these partitions in x,y are two of the input parameters.

3. For each 3D partition, we consider non-overlapping smaller 2D blocks. For each block, a scaling factor(beta) and offset(alpha) parameters are supposed which should be multiplied by and added to each voxel respectively, through an affine transformation, (x_2=beta*x_1+alpha). The size of these blocks in the x,y directions are two other input parameters. These two parameters determine the minimum size of the local intensity discrepencies which should be eliminated.

4. Difference between the affine transformations of the aside blocks along the z-direction is optimized by the L-BFGS-B algorithm. Parameters in this optimization process are the mentioned scaling factors and offsets which should be bounded. Also, a regularization term is added to the objective function to prevent defacing the image in the x,y directions.

5. All of the steps 1-4 are repeated for the actual 3D image without any downsmapling. The initial values for the optimization parmaters in this step are the outputs of the previous steps.

6. All of the steps 1-4 are repeated for the actual image but by cinsidering smaller blocks. (new_block_size=old_block_size/2). The initial values for the optimization parameters in this step are the outputs of step 5.

7. The optimized parameters generated in step 6 are smoothed with an averaging filter to remove the blocking effect produced in the first and second steps.

#Required python modules: numpy scipy h5py skimage joblib

#Usage: ./EM_normalization.py [options] <stp_x> <stp_y> <w_x> <w_y> <par_x> <par_y> <n_cpu>

options:
--grp		STRING		The group name of your input file if it is .h5, and the group name of the output
						files.[Default: 'stack'] 
--up_fg		0/1			If you do not want to upsample your input image in the z-direction
						not to have a better resolution, you should put the 'up_fg=0'.[Default:1]
--up_z		INT			If you decide to upsample the input image, 'up_z' is the parameter
						that shows the ratio of upsampling in the z-dir.[Default: 4]			
--fact		FLOAT		Determines the accuracy of the optimization process.[Default: 1e10]	
						Typical values for fact are: 1e12 for low accuracy; 1e7 for moderate accuracy; 
						10.0 for extremely high accuracy. The lower the 'fact' value, the higher the 
						running time. The values in [1e8,1e10] produce good results.
--cns_fg	0/1			If you want to increase the contrast of the output image, set the cns_fg=1.
						[Default:0]
--cns_low	FLOAT 
--cns_high	FLOAT		If you have set cns_fg=1, you can choose which low and high quantiles be removed.
						[Default:(0.00001,0.99999)]
--lm 		FLOAT 		Determines the ratio of the regularization term to the loss function. Choosing a
						very large number may prevent your code from removing the artifacts. On the other
						hand, a very small value may result in some kinds of memory in the normalized
						image. Its value depends on the application.[Default: 100]
						Example: For electron microscpy datasets the default value is preferred. For 
						time-lapse photography, it can be in the order of 10^6.
-h						Help. Print Usage.