Wei Liu at SCI Institute of University of Utah
Wei Liu's recent picture

Wei Liu

Ph.D. Student
Scientific Computing and Imaging Institute
University of Utah

Room 2720, Warnock Engineering Building
72 S Central Campus Drive
Salt Lake City, UT 84112

Office Phone: 801-585-0611
Email: weiliu at sci dot utah dot edu

You reached here probably because of my work on functional network of human brain. Now I work at corporate research lab of ExxonMobil, New Jersey. Driven by the industrial curiosity and business needs, I am studying deep learning. I am particularly interested in segmentation/objection detection/parsing of 3D volume images with little or no training data.


November 2014: I accepted the job offer of ExxonMobil and will work at the Corporate Strategic Research lab in New Jersey soon.

June 1st: Our NeuroImage paper of hierarchical Markov random field was accepted. It's a happy coincidence that our daughter was born on that day, and I received the notification email in labor room.

February 2014: I moved to UCAIR and start to work as a postdoc.

January 2014: I will soon start my post doc work with Dr. Brian Chapman at Utah Center for Advanced Imaging Research.

Nov 25, 2013: I successfully defended my Ph.D. Here is the slides of the talk during the thesis defense.

Nov 26, 2013: Our paper "Active Learning Based Graph Cuts Segmentation For Modeling Pathological Anatomy: A Study On TBI Imaging" is accepted by ISBI 2014.

I am near the end of my PhD at the Scientific Computing and Image Institute (SCI) at the University of Utah, and am starting to look for jobs. My interests include medical imaging, machine learning, computational statistics, and general computer vision problems.

At SCI, I work with Dr. Tom Fletcher as a research assistant and study functional magnetic resonance (fMRI) imaging and brain's functional connectivity/networks. I'm particularly interested in the consistent functional network estimation from dataset of multiple subjects by using graphical model, hierarchical model and Bayesian methods. My research statement can be found here, and my thesis proposal is here.

Before joining SCI, I earned both the B.S and M.S. degrees at Jilin University, China, and then worked at Lucent Technologies, Nanjing, China (now Alcatel-Lucent after the merger) as a software engineer until June 2008. A quick summary of my Lucent work can be found here. There are also a few certificates and awards for my works and community service that can be found here.



My main research is identifying functional connectivity of human brain by using resting-state functional MRI data. The spontaneous fluctuation of the blood oxygenation level-dependent signal measured by fMRI is a rich source of information for understanding the functional organization of human brain. The abnormal functional pattern is also correlated with some neural diseases. So a good estimates of the individual and the group's functional networks also help find the cause and possible diagnosis of these disease.

The mathematical tools I used include graphical model, Bayesian method, Markov random field and hierarchical models. Graphical model is a diagrammatic representation of multivariate probabilistic distributions. With a graphical model, one can visualize the structure of a probabilistic model, improve old model and design new models. Also conditional independence can be obtained easily by just inspecting the graph. Graphical model can be used in Bayesian data analysis, where the observed data, the hidden variables and the parameters in the prior and hyper-prior distribution all integrated into a hierarchical structure. Other problems that graphical model can solve include various network problems, such as flow networks, electrical power networks, etc.

Markov random field (MRF), as a statistical model defined on undirected graph, is a principal method to model the soft constraints/dependency between variables. As a regularization approach, the MRF used in our work can also be used in more general situations where the problem can be represented by a spatiotemporal process. The hierarchical model used in the fMRI analysis can also be generalized to represent the variation and shared information within and between groups of data. This is especially useful for today's large data analysis when data comes from various sources.


Pulmonary Vessel and Airway Extraction for Phenotyping Study

At UCAIR, we start working on CT Angiography images of human's lung. We are interested in the topological properties of the airways, arteries and veins. With the imaging data from thousands of normal people and patients with pulmonary hypertension, we aim to study the statistical variations of these structures among healthy people and patients. In the future, we also aim the correlation between the anatomical structures and the patients' gene information.
high-dimensional MRF

Full brain Pairwise functional connectivity with regularization of Markov random field

We estimate the pairwise functional connectivity within the whole brain volume of a single subject. Functional MRI data, as a sequence of 3D volumes sampled on various time points, often has noise and confounding factors such as scanner noise, head motions, physiological signals like heart beating, breathing and blood vessel. We use Markov random field (MRF), a undirected graphical model to represent the spatial context information. Because of the pairwise correlations, we redefine the MRF in the 2p (p = 3) correlation space. The connection/no-connection is estimated from a two-class Gaussian mixture model, with MRF as a prior distribution of connectivity variables.

Identify Consistent, Spatially Coherent Multiple Functional Networks.

Functional network, or functional system of the human brain is the set of spatially remote regions or voxels having correlated time series signal as measured by functional MRI. In this project, we define the functional network estimation as an un-supervised image segmentation problem. We use a generative Bayesian model, where the prior distribution of the network label maps are defined as the Markov random field to reflect our knowledge on the spatial context information. The data likelihood is defined as a von Mises-Fisher (vMF) distribution. vMF is similar to an isotropic multivariate Gaussian distribution, except that the variables are in a hyper-sphere. The method is data-driven, as all the parameters are estimated in a expectation maximization framework. Compared to the previous connectivity analysis, this model can estimate more than one functional system without a priori seed region.
hierarchical mrf

Hierarchical Model For Group Study

Functional MRI dataset includes multiple subjects, and functional network are often identified from a group of subjects. The functional patterns between subjects even in the same healthy control group are different, although they show a significant amount of similarity. This hierarchical model aims to estimate both group map and individual subjects maps. In order to model the variations between subjects, we assume a group level, and the hidden subject network maps are generated from the group. The hierarchical model's inference is more difficult than single-subject analysis, and the pooled analysis (where all subject data are just pooled and a single group map is estimated).

We define a joint graph that includes the network label variables of both group and all the subjects. The within-subject links model the spatial smoothness within single subject, and the inter-level links model the similarity and variations between group and subjects, implicitly model the between-subject relationships. We use a Bayesian approach and Markov chain Monte Carlo sampling for the inference of the model.

hierarchical mrf

Allen Brain Institute Gene Study

Allen Brain Institute provide the human brain's gene data of 6 subjects. For each subject, around 900 tissue samples in the brain's cortex and subcortex area are analyzed and the micro-array gene data of around 60K probes are available. We have two goal in this fun projects: 1) To explore the gene expressions and see if the expression values are distributed according to the samples' spatial locations. 2) Since the brain's intrinsic networks are also spatially distributed, we aim to explore the relationship between the genes spatial distributions and the functional networks' spatial distribution. The second task is particularly challenging as the gene expressions and resting-state fMRI data are not from same subjects.
hierarchical mrf

TBI Image Segmentation

Traumatic brain injury images segmentation is difficult task due to the lack of the prior knowledge of the possible locations of the lesion area. Automatic approaches often result in large amount of false positive and false negative segmentations. We belive a good interactive tool is the key to solve such a problem. Instead of passive waiting for human's input, the algorithm actively ask questions that it thinks the answers will be most informative. We also represent the set of voxels by the super-voxels, or voxel patches with a multiscale framework. Furthermore, we use Markov random field for modeling the smoothness of the components within normal and lesion regions, and also the lesion regions themselves. For optimization we use a variational Bayesian method for within-class segmentation and graphcuts for between-class segmentation. See Bo Wang's project page for more information.
Autism patient brain connectivity

Autism Classification with Functional Connectivity

Autism spectrum disorder(ASDs) patients have been shown to have abnormal functional connectivities in their resting-state functional networks. The large open access data repository , Autism Brain Imaging Data Exchange (ABIDE) provides collected functional MRI datasets from individuals with ASDs and typical controls (TC). The resting-state fMRI experiments do not need to design complex tasks for probing the various characteristic of ASDs, and its test-retest reliability also shows it's a good tool for understanding the reason of and diagnosis of ASDs. We define this as a linear regression problem, with connectivity variables as independent variables, and the binary score (ASD vs. TC) as the dependent variables. Due to the huge number of independent variables, we need some regularization method such that only the most import features/connectivities are selected.

Lecture Notes

Courses and seminars I have taken:


W. Liu, S. Awate, P.T. Fletcher. “A Functional Networks Estimation Method of Resting-State fMRI Using a Hierarchical Markov Random Field,” In NeuroImage, 100, 520--534.

W. Liu, S. Awate, P.T. Fletcher. “Group Analysis of Resting-State fMRI by Hierarchical Markov Random Fields,” In Medical Image Computing and Computer-Assisted Intervention (MICCAI 2012), Lecture Notes in Computer Science (LNCS), Vol. 7512, pp. 189--196. 2012.
ISBN: 978-3-642-33453-5
DOI: 10.1007/978-3-642-33454-2_24

W. Liu, S. Awate, J. Anderson, D. Yurgelun-Todd, P.T. Fletcher. “Monte Carlo expectation maximization with hidden Markov models to detect functional networks in resting-state fMRI,” In Machine Learning in Medical Imaging, Lecture Notes in Computer Science (LNCS), Vol. 7009/2011, pp. 59--66. 2011.
DOI: 10.1007/978-3-642-24319-6_8

W. Liu, P. Zhu, J.S. Anderson, D. Yurgelun-Todd, P.T. Fletcher. “Spatial Regularization of Functional Connectivity Using High-Dimensional Markov Random Fields,” In Medical Image Computing and Computer-Assisted Intervention (MICCAI 2010), Vol. 14, pp. 363--370. 2010.
PubMed ID: 20879336

B. Wang, W. Liu, M. Prastawa, A. Irimia, P. M. Vespa, J. D. van Horn, P. T. Fletcher, G. Gerig. "Active Learning Based Graph Cuts Segmentation For Modeling Pathological Anatomy: A Study On TBI Imaging,", In IEEE International Symposium on Biomedical Imaging, 2014, In Press.

During my master degree study at Jilin University, I did some preliminary work on kernel methods of machine learning. Here are two small pieces of papers (hard to find the online information, though):

Wei Liu, Hexin Chen, Mianshu Chen. “Kernel based Optimal Iterative Discriminant Analysis” in International Conference on Computational Intelligence, Robotics and Autonomous Systems, Singapore, December, 2003. ( PDF)

CHEN Mianshu , Chen He-Xin, LIU Wei, “A New Method for Resolving the Uncorrelated Set of Discriminant Vectors,” in Chinese Journal of Computers, July 2004. (PDF)


While not staring at the fMRI data in the lab, I do some outdoor activities. I started cycling since 2011. This is a good way of exploring the Utah's natural scene and blending into local people. I also try to play badminton again that I have left off for six years. My 10-year-old racket is still in good shape. Other than that, I occasionally play tennis. One of my favorite games is soccer, though it might be too dangerous to play in Utah. I started skiing seriously since 2012, although most time I ski during the night at the Brighton ski resort -- the daytime season pass is just too expensive for poor grad students. My skills of these games are all around the beginner-intermediate level.

Update 2012: This year I started to learn basic of photography.

Update 2013: Last year's photography is not quite successful, even I bought a DSLR. However, this summer I eventually learned to swim. No panic in deep water any more. Voila!

2014: We had a baby girl on June 1st! We gave her name Sophie. Interestingly, she was born on the Duanwu festival of the Chinese lunar Calendar. Photos are coming soon.