Stevens INI researchers use generative AI to fill gaps in brain wiring maps

Researchers from the Mark and Mary Stevens Neuroimaging and Informatics Institute (Stevens INI) at the Keck School of Medicine of USC have developed an artificial intelligence (AI) model that helps fill gaps in brain wiring maps made with diffusion MRI, a type of scan that traces the brain's nerve fiber pathways. This could help researchers study how key brain connections change with aging and disease.

The model, described in a study published in Medical Image Analysis, targets the brainstem and the orbitofrontal region just above the eyes, two areas where distortions in the scanner's magnetic field can erase much of the signal. The team trained and tested the model using scans from 1,315 participants in the UK Biobank and 679 participants in the Lifespan Human Connectome Project Aging study (HCP-Aging), two large studies of mostly older adults.

“In a few parts of the brain, MRI scans can come back with pieces of information missing, a bit like a photo with a corner torn off,” said Shuo Huang, co-first author of the study and a postdoctoral fellow at the Stevens INI. “Our model looks at the intact tissue around each gap, along with what it learned from scans without those gaps, and estimates what most likely belongs there.”

Where brain scans lose the thread

Diffusion MRI tracks how water moves through brain tissue. Because water tends to move along nerve fibers, the scans reveal the paths fibers take and how brain regions connect.

Some regions are especially prone to signal loss. The problem lies with the scan, not the brain tissue, and standard correction tools cannot fully fix it. In the brainstem, which connects the brain to the spinal cord, these gaps can disrupt the trace of the corticospinal tract, a major pathway that carries motor signals. 

Borrowing a technique from AI image generators

The new model, FOD-Diffusion, builds on a type of generative AI called a diffusion model, which creates images by starting with random noise and removing it step by step. 

The data it restores, called fiber orientation distributions, consist of 45 linked 3D maps of the brain for each scan. Each map captures one part of the fiber pattern, and together they show which directions fibers run at each point. The model fills in one map at a time, drawing on all the others so the pieces fit together, and uses the well-measured tissue around each gap as a guide.

 “Most AI image tools are designed for ordinary pictures, but brain wiring data is more like a stack of dozens of connected maps that all have to agree with one another,” said Lujia Zhong, co-first author of the study and a PhD candidate at the Stevens INI. “We designed our model to fill in each map while keeping the whole stack consistent.”

Closer matches and more complete pathways 

To test accuracy, the team hid sections of the brainstem in scans with minimal signal loss and asked the model to reconstruct them. Across 43 test participants, the model's main fiber directions differed from the original data by about 1.4 degrees on average. The closest of six other AI approaches tested was off by about 4.5 degrees.

In 1,172 additional UK Biobank participants, brainstem fiber information corrected with standard tools alone was significantly weaker than in participants with minimal signal loss. After the AI model restored the data, the difference was no longer significant. Results were similar in the orbitofrontal region among 981 participants. 

In participants with severe signal loss, traces of the corticospinal tract from standard-corrected data were significantly less complete through the brainstem, while traces from AI-restored data did not differ significantly from those of participants with minimal signal loss. 

Getting more from a single scan direction

MRI scanners can collect diffusion images in two opposite directions, front to back and back to front. Combining both helps correct distortions. The HCP-Aging study collected complete scans in both directions, letting the team compare approaches.

“Scan time is limited, so many of the largest brain imaging studies collect a full set of scans from only one direction,” said Yonggang Shi, PhD, senior and corresponding author and a professor of neurology, biomedical engineering, and electrical and computer engineering. “In our tests, filling in the gaps with the AI model in those one-direction scans produced better maps of a key movement pathway than combining scans from both directions. That means researchers may be able to get more out of brain scans they have already collected.”

An estimate, not a measurement

The researchers note that restored data are informed estimates, not recovered measurements. Because no human diffusion scan is entirely distortion-free, they judged accuracy against participants with minimal signal loss rather than a perfect reference. The method still needs testing on other scanners and settings and has not been evaluated for patient care.

The team plans to apply the model to other large datasets, including the Alzheimer's Disease Neuroimaging Initiative and the Health and Aging Brain Study-Health Disparities. The model's code is publicly available on GitHub.

“Some of the brain's most important pathways run through regions that have been especially difficult to image,” said Arthur W. Toga, PhD, director of the Stevens INI. “By sharing this tool openly, the team is giving researchers a way to strengthen analyses of the large datasets the field depends on.”

About the study

 The study's authors are Shuo Huang, Lujia Zhong, and Yonggang Shi of the Stevens INI. Huang and Zhong contributed equally to the work. All three are also affiliated with the USC Viterbi School of Engineering.

This work was supported by the National Institutes of Health under grants R01EB022744, RF1AG077578, R01AG064584, U19AG078109, and P30AG066530. The research was conducted using the UK Biobank Resource under approved project 25641.