Lawrence Livermore National Laboratory



Back to top

Back to top

Deep Learning Improves Disease Diagnosis

Livermore computer scientists and colleagues at IBM Research Almaden in San Jose, California have developed new deep-learning models to accurately diagnose diseases from x-ray images with less labeled data. The team’s paper, “Self-Training with Improved Regularization for Sample-Efficient Chest X-Ray Classification,” won the Best Paper Award for Computer-Aided Diagnosis at the 2021 SPIE Medical Imaging Conference.

The machine-learning technique, which includes novel regularization and self-training strategies, addresses some well-known challenges in the adoption of artificial intelligence for disease diagnosis. The team applied their learning approach to benchmark data sets of chest x-rays containing both labeled and unlabeled data to diagnose five different heart conditions: cardiomegaly, edema, consolidation, atelectasis, and pleural effusion. Using a framework that applies strategies including data augmentation, confidence tampering, and self-training, the researchers saw an 85 percent reduction in the amount of labeled data required to achieve the same performance as the existing state-of-the-art in neural networks trained on the entire labeled data set.

Livermore computer scientist Jay Thiagarajan explains, “We’re trying to address a fundamental problem. Data comes from different hospitals, it’s difficult to label, and experts don’t have the time to annotate it all. It’s often posed as a multi-label classification problem, where we are looking at the presence of multiple diseases in one shot. We can’t wait to have enough data for every combination of disease conditions, so we built a new technique that tries to compensate for this lack of data using regularization strategies that can make deep-learning models much more efficient, even with limited data. We’ve demonstrated that accurate models can be created with limited labeled data and perform as well or even better than neural networks trained on much larger labeled data sets.”
Contact: Jay Thiagarajan (925) 424-2255 (jayaramanthi1@llnl.gov).


Missing Physics in Explosive Hotspots Uncovered

Scientists at the Lawrence Livermore National Laboratory Energetic Materials Center and Purdue University used the Laboratory’s Quartz supercomputer to run large-scale, all-atom simulations to study the longtime relaxation properties of the kinetic and potential energy in hotspots and uncover missing physics in TATB (1,3,5-triamino-2,4,6-trinitrobenzene) hotspots and other explosives critical to managing the nation’s nuclear stockpile. Their research is featured in the March 11, 2021, issue of the Journal of Physical Chemistry Letters.

Livermore computational chemist Matthew Kroonblawd explains that “continuum-level multiphysics models used to assess safety and performance are highly empirical, making it difficult to transfer explosives models to different application conditions. The lack of transferable models is especially true for insensitive high explosives such as TATB. It is still not possible to build an explosives model from first principles, indicating that key physics aspects are missing.”

The work highlights a neglected physical aspect of the early stages of explosive hotspot formation, localized regions of elevated temperature and potential energy that accelerate chemistry. A better understanding of how hotspots form and evolve provides a route to systematically improve models used to assess stockpile performance and safety. Identifying the cause behind differences in hotspot reaction rates will inform more general explosives models and improve their predictive accuracy and transferability.
Contact: Matthew Kroonblawd (925) 422-2221 (kroonblawd1@llnl.gov).


Reduction in Marine Clouds Amplify Warming

Low-level clouds, such as the marine stratocumulus clouds responsible for gloomy San Francisco summers, are widespread over global oceans and cool the planet by shading the surface from sunlight. A new analysis of satellite cloud observations, however, indicates that global warming causes these clouds to decrease, and that, overall, their cooling effect will be modestly reduced as the concentration of carbon dioxide (CO2) in the atmosphere increases. The work, led by researchers at Lawrence Livermore National Laboratory in collaboration with colleagues from Scripps Institution of Oceanography and the NASA Langley Research Center, appears in the May 13, 2021, issue of Nature Climate Change.

The new study estimated how marine low clouds respond to natural variations in large-scale meteorological conditions, used global climate model simulations to determine how these meteorological conditions will change as atmospheric CO2 increases, and then calculated how the clouds will respond to this modified meteorological environment.

To test their method, the researchers turned to a highly unusual and extreme sea-surface warming event, or “marine heat wave,” observed in the northeast Pacific Ocean in 2015. Former Lawrence Livermore climate scientist and lead author of the study, Tim Myers says, “We showed that we could accurately predict the cloud changes detected by satellites during the marine heat wave, so we are confident we can predict how the clouds will respond to global warming.”
Contact: Mark Zelinka (925) 423-5146 (zelinka1@llnl.gov).