In the quest to better understand and cure childhood diseases, scientists at St. Jude Children's Research Hospital accumulate enormous amounts of data from powerful video microscopes. To help St. Jude scientists mine that trove of data, researchers at Oak Ridge National Laboratory have created custom algorithms that can provide a deeper understanding of the images and quicken the pace of research.
The work resides in St. Jude's Department of Developmental Neurobiology in Memphis, Tennessee, where scientists use advanced microscopy to capture the details of phenomena such as nerve cell growth and migration in the brains of mice. ORNL researchers take those videos and leverage their expertise in image processing, computational science, and machine learning to analyze the footage and create statistics.
"This happens in fields where the use of advanced microscopy is common. In the past, you might have had medical researchers manually circling the cells in those images and making measurements. But we can automate that process and make it happen faster and with better detail," says Philip Bingham, leader of the Imaging, Signals, and Machine Learning (ISML) Group within the Electrical and Electronics Systems Research Division at ORNL.
The collaboration was jump-started in 2009 through the lab's directed research and development program when St. Jude recognized that it had a wealth of microscopy data but needed to extract more information and turned to the lab for solutions, says Shaun Gleason, the initial principal investigator for the collaboration and now director of ORNL's Computational Sciences and Engineering Division.
One of the biggest challenges initially was the "language barrier" between the engineers and computer scientists at ORNL with deep expertise in optics and image analytics, and the researchers at St. Jude who understood what was going on biologically at the cellular and subcellular level. "We didn't understand their terms and they didn't understand ours," Gleason says, adding that after eight years of work on various projects, "we have a strong relationship where we can easily communicate and make great progress."
ORNL's Ryan Kerekes led the work over a span of several years. He notes that although there are plenty of off-the-shelf software packages to analyze cell imagery, ORNL has created custom algorithms to do specialized measurements sought by the St. Jude faculty. "Most of our work for St. Jude is to write these algorithms for a specific task on a specific biological assay. We apply image processing methods and machine learning to develop software that's just not available anywhere else," says Kerekes, now leader of the RF, Communications, and Cyber-Physical Security Group at the lab.
Adds Gleason: "There's so much information that St. Jude has yet to quantify in their imaging datasets — tracking proteins, organelles, populations of cells — the list is seemingly endless. St. Jude's ability to image at high resolution and to tag cellular structures of interest with different fluorescent proteins is constantly improving, and they are generating more data with interesting content. That means their need for automated analysis and high-performance computing is also growing. We are helping them do what they do better and faster because of computational tools that accelerate the discovery process. And that means cures can come faster."
"Restoring Auditory Cortex Plasticity in Adult Mice by Restricting Thalamic Adenosine Signaling," published in the journal Science, details St. Jude research on brain plasticity, or the ability of the brain to change and form new connections between neurons. In this work, ORNL helped track mice brain cell electrical activity in the auditory cortex when the animals were exposed to certain tones.
ORNL researchers created an algorithm to measure electrical activations, or signals, across groups of neurons, collecting statistics and making correlations between cell activity in the auditory cortex and tones heard by the mice. The team first had to stabilize the video because it was taken while the mice were awake and moving to ensure a proper analysis was being conducted, says Derek Rose, who now leads the work at ORNL.
Machine learning is deployed in these projects when software is written to automatically recognize what neurons look like and successfully locates them, says Rose, who is assisted by postdoctoral researcher Matthew Eicholtz. Such techniques could be leveraged further as datasets grow, with images covering a wider field of view with greater cell populations, for instance.
The St. Jude scientists found some interesting results in the research: by regulating a single brain chemical, adenosine, plasticity potentially can be restored in adults to levels similar to those found in juveniles. The work holds promise for the treatment of certain neurological conditions as well as for enhancing adults' abilities to learn new languages or acquire musical skills.
Through the years, what ORNL has brought to the table is a deep understanding of imaging modalities, Bingham says. "We're dealing with a 4-D dataset and that is not a common thing. It's not economically viable for private industry to do these specialized tasks that perhaps only one or two doctors in the world may be interested in. It makes sense to come to a national lab where we offer cross-disciplinary, scientifically tailored expertise and a cost-effective solution," he says.
David Solecki, associate faculty in Developmental Neurobiology at St. Jude, says, "Our long-term collaboration with ORNL has been indispensable to developing a number of image analysis tools for neuroscience research. We highly value the collaboration."
The algorithms and methods developed at ORNL "save time and free up biologists to do what they do best in the lab, allowing them to devote their expert knowledge to analysis tasks," Solecki says.
The two Tennessee institutions plan to continue their mutually beneficial collaboration.
"When we tackle science problems like this that may typically be outside our domain, we build capabilities in terms of data architecture, advanced analytics leveraging high-performance computing, and new machine and deep-learning methods," Gleason says. "Many of the applications developed for St. Jude can be directly adopted or adjusted for other uses."
Rose says the ORNL research for St. Jude will increasingly employ deep learning techniques that can perform automatic detection of mitochondria and other structures without having to hand-select descriptive features beforehand. The result will be even more rigorous analysis and statistics from larger datasets, he says.
Gleason adds, "Looking to the future, we would like to continue doing what we're doing with St. Jude and even expand our relationship to help them solve some of their biggest challenges, while also assisting them as they build computational capabilities in-house. ORNL will still be here for the really hard problems and the bleeding-edge technology."
The most important aspect of the work is the human one, the researchers say.
"The projects are interesting, and we enjoy solving the puzzles of how to capture the best data from these images," Kerekes says. "But when we go down to St. Jude and visit with our collaborators, we eat in the same cafeteria as the patients and their families, and the goal of our research is very clear: we are trying to develop knowledge that will lead to treatments to save the lives of children. Being there really makes that hit home."
The research has been supported by ORNL's laboratory directed research and development program and St. Jude's Children's Research Hospital.
No entries found