Medical Imaging Data Compression Using Adaptive Space Filling Curves

Summary

Magnetic Resonance Imaging (MRI) is one of the most widely used medical imaging data/modalities in neuroimaging. MRI data are stored as 3 dimensional (3D) tomographical data. Functional magnetic resonance imaging (fMRI) data, which has 3D plus a 1D time dimension, usually result in computation of 3D brain spatial activation maps. These activation maps can identify different brain networks involved in processing a task, stimulus, etc. For fMRI analyses, 3D spatial data of the brain are usually mapped to 1D for further analysis, and sometimes smoothened and/or compressed using 3D kernels. These operations have been done traditionally by using pre-defined linear ordering schemes or kernels. Linear ordering is far from optimal in the sense that it does not retain the structure of the 3D brain. Finding an optimal mapping which uses an optimal space-filling curve could retain the structure of the brain much better [1]. In this project, we propose finding a data-adaptive space filling curve, heuristics of which were defined in [1], and apply it to compression of fMRI brain activation maps from a schizophrenia study, and evaluate its compression performance. Due to the large size of 3D neuroimaging data, finding an optimal space-filling curve and applying it to compression requires computational power. Success of this project's outcome would improve any subsequent analyses that neuroimaging researchers perform, which includes brain's network functional connectivity analysis, brain disease/disorder classification, and pseudo computed tomography image generation from MRI data [2].
References:
[1] Sakoglu et al. "In Search of Optimal Space-Filling Curves for 3-D to 1-D Mapping: Application to 3-D Brain MRI Data," Proceedings of ISCA 6th International Conference on Bioinformatics and Computational Biology (BICOB), pp. 61-66, March 2014, Las Vegas, NV.
[2] Kuljus et al. "Comparison of hidden Markov chain models and hidden Markov random field models in estimation of computed tomography images," Communications in Statistics: Case Studies, Data Analysis and Applications 4, no. 1 (2018): 46-55.

Job Description

The student will -read the necessary literature provided by the PI, -meet with the PI regularly, i.e. at least weekly, -assist in development of the NUC algorithm -implement the algorithm in MATLAB scientific programming language -run the programs on an EXSEDE supercomputing resources -write monthly progress reports -help PI publish any results (this may go beyond the dates of the project) in conference or journals

Computational Resources

XSEDE supercomputing resources will be utilized to run the implementation of the developed algorithms which will be implemented in MATLAB. The PI has already access to TACC via EXSEDE until 02/2021, therefore existing access will be utilized; however, if necessary, new EXSEDE resources may be requested.

Contribution to Community

Position Type

Apprentice

Training Plan

The PI has been training a volunteer undergraduate student at UHCL on this research topic since January 2020. Upon successful approval of the proposed project by XSEDE EMPOWER, the student will take existing available XSEDE trainings, including XSEDE HPC workshop (UHCL is also a location), focusing on big data. The PI will also train the student on MRI datasets and intermediate/advanced MATLAB programming. Basic skills of MATLAB programming, as well as fundamental knowledge of linear algebra and signals and systems is expected.

Student Prerequisites/Conditions/Qualifications

MATLAB programming skills.
Positive attitude, proactivity, willingness to learn, and good math and programming skills, good communication skills.