AI & TBI: Bringing artificial intelligence to the forefront of traumatic brain injury research

A child’s sledding accident. A bad football tackle. A wartime explosion near a service member. These are just a few of the ways traumatic brain injuries can occur, upending a person’s life in an instant. 

Every year, nearly 2.8 million people in the U.S. suffer a traumatic brain injury (TBI), a condition that can have life altering consequences. From emergency department visits to long-term rehabilitation, the challange  of diagnosing and treating TBI remains complex and, at times, imprecise. The reason being, our current understanding, ability to accurately diagnose, and treatment of this condition remains, largely, understudied.

With support from the UC Noyce Initiative, a multidisciplinary team of researchers from three University of California campuses—Berkeley, Davis, and San Francisco— are developing cutting-edge, machine learning that is paving the way for a new era in TBI diagnosis and prognosis. 

Together, the team of computational sciences and physician scientists are working to tackle a crucial problem: how to harness vast amounts of medical data, particularly CT scans, to improve patient outcomes.

From Data Overload to Actionable

Insights TBI diagnosis often begins with a head CT scan of the brain, yet these scans are typically classified in a binary fashion: positive (indicating hemorrhage) or negative (showing no visible trauma). However, these images contain a wealth of underutilized information that could significantly refine diagnoses and predict patient trajectories.

In short, the problem isn’t a lack of data—it’s making sense of it all.

That’s where machine learning enters the picture. The UC research team developed advanced AI-driven image processing pipelines that extract crucial features from CT scans, offering insights that go beyond what the human eye can detect. Using high-performance computing infrastructure at UC Berkeley and the National Energy Research Supercomputing Center, these machine learning models are designed to provide clinicians with precise, quantitative assessments of TBI severity, injury patterns and likely outcomes.

Collaboration at the Cutting Edge 

Led by Geoffrey Manley, M.D., Ph.D., chief of neurosurgery at Zuckerberg San Francisco General Hospital and professor at UCSF, and co-principal investigator Adam Ferguson, M.S., Ph.D., the project has leveraged a wealth of resources, including UCSF's Transforming Research and Clinical Knowledge in Traumatic Brain Injury (TRACK-TBI) dataset. This TRACK-TBI study provided a comprehensive database of clinical information, neuroimaging, blood biomarkers and patient outcomes. By integrating this dataset with a clinical TBI registry at UC Davis, the team was able to cross-validate findings and refine their machine learning models. 

“The most exciting accomplishment has been successfully bringing together physicians, biomedical scientists, and data scientists with different expertise and perspectives to work on a common goal of accelerating TBI precision medicine,” said Manley. 

He continued, “We gathered massive amounts of clinical imaging data, curated them at UCSF, and prototyped a machine learning pipeline, which is now being deployed at scale in cloud-based supercomputing infrastructure at UC Berkeley.”

Generative AI and Precision Medicine

A major breakthrough of this project has been the development of AutoCT, a generative AI-based tool that automates CT registration, segmentation and quantification. This technology not only improves the accuracy of TBI diagnosis but also reduces the workload on neuroradiologists, thereby helping prevent burnout and allowing clinicians to focus on patient care. 

Kristofer Bouchard, Ph.D., neuroscientist at Lawrence Berkeley National Laboratory and UC Berkeley, has brought his expertise in generative-AI and health to the project. 

Bouchard has played a key role in refining these AI algorithms, ensuring they not only enhance diagnostic precision but also provide interpretable results that clinicians can trust. 

“A critical first step towards personalized TBI treatments is increasing the precision with which we describe brain injury features and then predicting outcomes based on multimodal data, including imaging, blood biomarkers and medical history," he said.

This technology not only improves the accuracy of TBI diagnosis, but also reduces the workload on neurologists, thereby helping prevent burnout and allowing the clinicians to focus on patient care. 

TBI Research Lays Groundwork for Other Diseases 

Beyond the immediate benefits for TBI care, this project has laid the groundwork for future generative-AI applications in medicine. The tools developed—particularly those for integrating medical imaging with clinical data—could be extended to other complex diseases that rely on combined decision support from imaging and biomarkers. The UC Noyce Initiative’s funding has not only accelerated this research but has also fostered new collaborations and created a lasting platform for ongoing innovation, according to the team. 

The team has already begun sharing their findings with the broader scientific community. Meanwhile, the project has also nurtured new talent. Research assistants and postdoctoral scholars at UC Berkeley and UCSF have contributed to this initiative, with one research assistant now pursuing graduate studies.

"This collaboration has been a massive effort, from coordinating data curation to deploying large-scale AI models," said Manley. "With the foundation now in place, we are well-positioned to continue streamlining the workflow from raw data to actionable clinical results—an effort that could transform precision medicine for TBI and beyond."