Examples: Image-Based Experimental Analysis 

Analysis from Synchrotron Data
Designing, Building and Optimizing Image Analysis of Diverse Materials

D. Ushizima, H. Krishnan, D. Parkinson

Our goal is to design and develop science-driven algorithms and software exploiting advances in applied mathematics, computer vision, machine learning, visualization, and statistics for quantitative analysis from 2D, 3D and 3D+time digital images. High-resolution images have turned into a common input to data analysis in several scientific domains. This process involves quantification over large data files, containing hundreds of slices of 3D objects, e.g. to measure the porosity of these volumes. Our goal is to collaborate with facilities that design nondestructive techniques to image materials and build software that can quantify properties in the interior of solid objects, including information on their 3D geometries. This quantification will support modeling of the fluid dynamics into the pore space of the host object.


Example: Analysis of Porous Material Analysis for Carbon Sequestration:

Facility: Beamline 8.3.2 of Advanced Light Source. 


Beamline 8.3.2 is for high resolution synchrotron-based X-ray micro-tomography. This beamline generated data at  ~100 GB/week in 2010, increasing to 2-5 TB/week in 2012. Upcoming sensors can acquire images at frame rates almost 1,000 times faster.

Our task here is to recover dense material from reconstructed micro-CT data. The figures in this page illustrates some of several experiments we analyzed. This one has investigate use of calcite precipitates to alter flow and soil strengthening. Important in CO2 sequestration and storage, subsurface remediation other geotechnical engineering processes.


We have analyzed the output of reconstructed parallel beam projected data, as illustrated in the figures in this page, using 3 main steps. These include a new automated parameter tuning method for noise reduction using 3D bilateral filters; a new method to minimize over-segmentation in 3D statistical region merging [1,2] and permeability calculation and comparison using maximum flow curves [1]. 


We have also tested our algorithms using simulated data. The reason is that  microCT images of porous materials are seldom composed by homogeneous particles, and they also present artifacts that can interfere in the calculation of permeability. In order to isolate permeability models from issues associated to CT imaging and granularity, we generate synthetic datasets of randomly distributed sphere packing using numerical simulation of Stokes flow. The simulation algorithm creates numerical constructions of jammed packed bead beds, with identical and nonoverlapping spheres, and following algorithms proposed by Salvatore Torquato, later modified by Todd Weisgraber (LLNL). Link for further information and C++ code used in our experiments. The use of synthetic beads is only the first step toward describing porous media, using a new descriptor based on topological analysis, that we named max-flow curves [1].


To provide these tools to other users, we have developed and introduced the software package Quant-CT [1,2,3,4] for automated filter parameter tuning and automated biphasic material segmentation. This has been made available to all NERSC users as a plugin to ImageJ/Fiji (software disclosure filed  Apr. 2013.) The software was tested in several different datasets (largest stack size ~ 10GB), with varying granularities.



Our focus now is to develop algorithms that identifies the components of multi-phase materials. Several of the filters and pre-processing tools can be accelerated immensely by using new paradigms that explore computer architectures such as those within Open-CL. These are some of the new challenges we have ahead.

Simulation and experimental data:   (left) simulation of glass bead packing used for algorithmic verification, (middle) glass bead-packed column in biogenic mixture under induced calcite precipitation, using microbe S. pasteurii , and (right) crack detection from ceramics.

Additional Videos:

Iron-coated sand stone (collab. Peter Nico, ESD)

Biogenic Mixture in glass bead-package (collab. Jonathan Ajo-Franklin, ESD)

Crack detection from ceramic under load at temperatures above 1,600oC (collab. Hrishikesh Bale, Dep. Materials Science and Eng., UCB)

[1] Ushizima, D.M., Morozov, D, Weber, G.H., Bianchi, A.G.C. and Bethel E.W., "Augmented topological descriptors of pore network", in: IEEE Trans. Comp Graph. and Vis. USA. [pdf]

[2] Ushizima, D.M., Bianchi, A.G.C., deBianchi, C. and Bethel E.W., "Material science image analysis using Quant-CT in ImageJ", in: ImageJ User and Developer Conference 2012, Oct, Luxembourg, LX. 

[3] Ushizima, D.M., Ajo-Franklin, J., Macdowell, A., Morozov, D., Nico, P., Parkinson, Bethel E.W, Sethian J.A., "Statistical segmentation and porosity quantification of 3D x-ray microtomography", in SPIE Optics and Photonics: XXXIV Applications of Digital Image Processing, Vol.8135-1, pp.1-14, Aug 2011, San Diego, CA, USA. [pdf]

[4] Ushizima, D. M., "Analysis and Visualization of High-throughput Experiments", in: Workshop on Driving Discovery: Visualization, Data Management, and Workflow Techniques - Advanced Photon Source, Argonne National Laboratory.

Example: Particle Image Analysis:
Particle Toxicity - An Answer that Relies on Multiple Scales

K. Odziomek, M. Haranczyk, D. Ushizima

Material informatics focuses on problems at the interface between material design and computation chemistry in order to discover and improve the models, design and/or utility of chemical compounds. An imaging modality for information-driven chemistry and material science is scanning electron microscopy (SEM), which can access information about the microarchitecture of compounds as well as the morphology of nanoparticles. The information from SEM images allows quantification of structural/chemical properties, relevant on predictive analytics of nanoparticles toxicity.


Our goal is to quantify the risks posed by engineered nanoparticles, particularly hydroxyapatites. This project investigates computational tools based on computer vision algorithms, applied to SEM images to predict toxicity of nanoparticles based on their morphology among other structural/physical properties. We have analyzed SEM images to obtain particle information (particle shape/size/surface/etc), and we are building a statistical model that will correlate both nano and microscale parameters with a biological response to those particles (e.g. toxicity). 



This model will be used to predict toxicity for new particles during their engineering process. Our goal is a workflow environment to analyze images, extract statistical parameters and classify samples. Current images come from a cooperation with the European Union, which involves different laboratories; this necessarily means that  our samples are subjected to different amount of artifacts and sample preparation issues.


In parallel with this on-going project, we are building other collaborations to test the approach on consistent series of images (SEM,TEM etc) presenting particles and their corresponding (measured) physical property  (e.g. solubility, absorption, etc). 

Example: Simulation of Filamentous Networks
Applying Machine Learning for Statistical Image Analysis

D. Ushizima and T. Perciano

Original synthetic image

Synthetic image with noise

Reconstructed image

Filaments are fundamental structures permeating materials. They define a channel network that influences material porosity and crack development. When detected from images, the filaments enable essential quantification and analysis of  processes related to the material. The challenge is the extraction of this type of structure from microscopic images due to filament complex spectral and spatial characteristics. Our project aims to

develop and apply machine learning techniques to tackle filament detection, measurement, and classification of filament networks.


Markov random fields theory provides a convenient and consistent way to model context dependent entities as image pixels/voxels or primitives. This is possible through the characterization of mutual influences between these entities. Our previous work [1, 2] uses this idea to detect thin structures from 2D images. The same algorithm can be used to extract blood vessels from 3D synthetic images [3], such that the algorithm is applied to 2D slices of the 3D images. Lacking complete 3D contextual information, which is essential for the MRF approach, the connectivity information leads to incomplete reconstructions (as observed in Fig.1-3). The goal of this project is to invent a new approach to deal with 3D structures, and apply to several microscopic images, including simulated data.





[1] T. Perciano, F. Tupin, R. Hirata Jr., R. M. Cesar, A hierarquical Markov random field for road network extraction and its application with optical and SAR data. International Geoscience and Remote Sensing Symposium (IGARSS), 2011, 1159-1162.

[2] H. Sportouche, F. Tupin, J.-M. Nicoloas, C.-A. Deledalle, How to combine TerraSAR-X and Cosmo-SkyMed high resolution images for a better scene understanding?. International Geoscience and Remote Sensing Symposium (IGARSS), 2012, 178-181.

[3] M. A. Galarreta-Valverde, M. Macedo, C. Mekkaoui, M. P. Jackowski, Three-dimensional synthetic blood vessel generation using stochastic L-systems, SPIE Medical Imaging, 2013.