NCSI: Real-Time Adaptive Control of Cell Behavior in Artificial 3D Cultures using Machine Learning

Status	Completed
Mentor Name	Roman Voronov
Mentor's XSEDE Affiliation	Campus Champion
Mentor Has Been in XSEDE Community	4-5 years
Project Title	Real-Time Adaptive Control of Cell Behavior in Artificial 3D Cultures using Machine Learning
Summary	Current approaches to tissue engineering fail because there is no central nervous system to orchestrate cell behavior in artificial cultures. During the Fall 2019 phase of the XSEDE EMPOWER project we submitted for a publication a novel automated microfluidics technology that integrates tissue engineering scaffolds with vascular-like channels. These channels are then used for analyzing and modulating cell behavior using a computer, which effectively acts like a brain for the artificial tissue. The analysis of the cell behavior occurs in real time and at different locations in the culture, which in turn requires the use of distributed computing in order to handle the work load. However, several problems still remain unsolved: 1) The calculation flow process requires continuous 3D imaging at a very high resolution. But when the individually-acquired tiles are stitched into a panorama, the data size becomes too large to fit into RAM (which slows down performance). Therefore, we have also been working on parallelized Matlab algorithm that would perform the image stabilization algorithm by analyzing the individual tiles (before the stitching). Currently, we have finished this code, as well as a control "naive" implementation of the stabilization for comparison. We have also generated 20 scenarios of different types of disturbances and drifts that the microscopy acquisition can experience, in order to make sure that our algorithm outperforms the "naive" implementation in call cases. With the remainder of the Fall 2019 we intend to complete the testing, and begin drafting up the manuscript that will share the results of our work with the rest of the community. After this bottleneck is resolved, we will be able to analyze the incoming stream of images in real-time, by sending them to the supercomputer. This will allow our device to respond to the imaged cell behavior, with the ultimate goal of controlling it. To that end, the Spring 2020 phase of the project will focus on developing a Model Predictive Controller (MPC) that will use a Recurrent Neural Network (RNN) in order to learn from empirical data and modulate the cell action based on real-time non-destructive feedback from 3D Lattice Light Sheet Microscopy (a Nobel Laureate technology available in the PI's lab) and chemical assaying. There will be many MPCs working collaboratoratively, with each one being responsible for a different location in the culture. Therefore, their performance will also be parallelized. The ultimate goal is to achieve the ability to culture reproducible tissue patterns, specified by the user prior to the at the beginning of the experiment. This technology will benefit patients in need of life-saving organ and tissue transplants.
Job Description	-The student will work 10 hours per week during the course of the project in Spring 2020. -The student will read the existing literature provided by the PI. -The student will develop custom Matlab codes that implement HDF5 in order to parallelize 3D rendering of microscopy images by integrating the code's output with Visit and Paraview packages using the XDMF standard. -The student will train a Recurrent Neural Network (RNN) algorithm on the microscopy images of cell behavior and on the chemical assaying data, for implementation in the Model Predictive Controller (MPC) -The student will explore XSEDE resources and learn about them in order to be able to use in this project. -The student will work as XSEDE student campus champion and promote the use of supercomputing on the NJIT campus.
Computational Resources	The project will use Texas Advanced Computing Center resources via a New Jersey Institute of Technology Campus Champion allocation as follows: -Long term tape storage for large data from Lattice Light Sheet Microscopy -Image processing of the 3D Lattice Light Sheet Microscopy data, and Visualization using Visit and Paraview packages -Training of the Recurrent Neural Network (RNN) on experimental microscopy and chemical assaying data for implementation in the Model Predictive Controller (MPC). Additionally, NJIT's local Kong cluster will be used for preliminary code development and testing
Contribution to Community
Position Type	Apprentice
Training Plan	1) The student will continue to receive training in HPC from the PI and the NJIT supercomputing staff 2) The student will continue to receive personal training from the PI in the areas of Image Processing, Machine Learning and Controls Theory 3) The student will further improve by interacting with an active research group of 2 PhDs and 3 MS students, all of whom are active in image processing and HPC. 4) The student will help the PI to develop computational educational materials as a part of the Campus Champion and Computer Aids for Chemical Engineering Education (CACHE) initiatives (the PI is an active member of the CACHE community, see http://cache50th.org/?page=youngfaculty )
Student Prerequisites/Conditions/Qualifications	Must be an undergraduate student at NJIT with experience in Matlab and Lattice Light Sheet Microscopy. Additionally, interest in the following is highly desired: - Image Processing - Machine learning - Controls Theory
Duration
Start Date	01/15/2020
End Date	04/15/2020