The Atlas of MS estimates 2.9 million people live with multiple sclerosis (MS) worldwide. For more than a century, scientists have known that MS damages more than the brain’s white matter. It also scars the brain’s outer layer, the cortex, made of gray matter. These scars, called cortical lesions, are small patches where the immune system attacks gray matter tissue. Cortical damage is closely tied to disability, thinking problems, and how fast MS progresses.
Yet routine MRI scans have rarely shown this damage clearly. Because doctors couldn’t reliably see cortical lesions, researchers couldn’t fully study what role this damage plays in disease progression. They also had no good way to tell whether MS treatments were actually slowing it.
A new study offers a way to see much more of this damage, using image processing and artificial intelligence (AI). That could finally let researchers study cortical lesions properly and check whether treatments protect against them.
Reanalyzing old scans with new tools
The study, led by the University at Buffalo, appears in Communications Medicine, a peer-reviewed journal. The researchers wanted to know whether combining several MRI-processing techniques with AI could reliably detect cortical lesions. They tested this on scans clinics already collect, without needing specialized equipment or new trials.
The researchers reanalyzed brain images collected during ORATORIO, a phase 3 trial of the MS drug ocrelizumab. That trial enrolled 732 people with primary progressive MS (PPMS), which affects roughly 10% to 15% of people with MS. It steadily worsens, without the flare-ups and remissions seen in other MS types.
The researchers wanted to know whether combining several MRI-processing techniques with AI could reliably detect cortical lesions.
Large trials like this already have years of stored scans. Pulling new information from them could answer key questions without repeating costly trials.
To reveal cortical lesions on these scans, the team combined several MRI-derived images and used AI to sharpen them. One step generated an AI image resembling a specialized scan type that shows cortical lesions well. That scan type isn’t part of routine imaging. Combining that image with other enhanced scans created a composite called MMCLE — the clearest view of the cortex yet. The team then trained separate AI software to flag lesions automatically.
More lesions than conventional scans usually catch
On a closely studied set of 80 patients, the team found nearly 15 cortical lesions per person on average. Across the full 732-person dataset, they detected more than 10,000 cortical lesions in total.
To test whether the enhanced images helped, independent reviewers compared images side by side, blinded to which type they viewed. To judge whether a spot was a real lesion, a panel of experts examined all image types together and reached agreement. That expert consensus became the standard for scoring each scan type — not an examination of brain tissue after death.
Reviewers correctly identified 86% of confirmed lesions on the new combined images. That’s far ahead of two conventional scan types doctors already use. One caught only about 38% of the same lesions, and another identified just 15%.
They also rarely mistook healthy tissue for damage. Fewer than 1 in 10 spots flagged as a lesion on the combined images turned out to be a false alarm. The real lesions also stood out far more clearly. Contrast between damaged and healthy tissue was more than twice as strong on the combined images as on standard scans. That makes lesions easier to spot, whether by a human reviewer or by the AI software itself.
The researchers also tested the approach in various ways: they rescanned some patients using different scanners, magnet strengths, and scan protocols. The method counted a similar number and size of lesions each time. Agreement among the results was in the high 80s to low 90s percent — evidence that they don’t depend on one specific machine.
Not every lesion was equally easy to find. Most lesions detected, 82%, sat at the border between gray and white matter. Far fewer, just 5.5%, sat entirely beneath the cortex’s surface — a type called subpial lesions. Earlier tissue studies link subpial lesions especially strongly to disability, and they remained the hardest type to detect.
Why this matters for future MS trials
Doctors have recognized cortical lesions since the late 1800s, but MRI couldn’t reliably see them. They were left out of official MS diagnostic criteria until this century. Many MS drugs developed over the past 20 years mainly target relapses and white matter. That’s largely because researchers had no way to measure effects on gray matter.
That’s the biggest opportunity this work opens up. If researchers can reliably measure cortical lesions on ordinary scans, future trials could test something new. They could check whether treatments protect gray matter, not just white matter. Researchers have largely been unable to evaluate that before.
Robert Zivadinov, the study’s senior author, said this in a University at Buffalo press release: “Detecting previously invisible cortical lesions on conventional legacy MRI scans has major implications for MS research and clinical care.” Co-author Michael Dwyer called it “a real success story for applying AI in the medical arena.”
“Detecting previously invisible cortical lesions on conventional legacy MRI scans has major implications for MS research and clinical care.”
— Robert Zivadinov, MD, PhD
Professor of Neurology, University at Buffalo
Because the method works on stored scans, researchers could revisit old trial data now. They could ask whether past treatments already affected cortical lesions, without new studies or costly scanners.
That said, this study is about methodology, not treatment effects. The researchers didn’t test whether ocrelizumab changed cortical lesions in the ORATORIO trial. That analysis is still to come.
What this study can’t tell us yet
The study has real limits. It’s a retrospective, exploratory analysis. As noted above, the researchers had no pathology-based reference standard to confirm every lesion. They did check their results against a small number of scans from “7 Tesla” MRI scanners. These are far more powerful, research-only machines that can show cortical lesions more clearly than standard hospital scanners. But only a handful of patients were scanned this way, so it’s a limited comparison.
Even so, the method still detects only a fraction of all cortical damage — missing most of the hardest-to-see subpial lesions.
Every patient studied had PPMS. It’s unclear whether the approach works as well in other MS forms, like relapsing-remitting or secondary progressive MS. Those forms involve different lesion patterns.
The method performed worse at tracking new or enlarging lesions over time than at spotting existing ones. That gap matters, since tracking change is exactly what’s needed to measure whether a treatment works.
Funding is worth noting: Genentech, which makes ocrelizumab and sponsored the original ORATORIO trial, also partly funded this study. Two co-authors are Genentech employees, and the senior author reports financial ties to several drug companies, including Genentech. That doesn’t undermine the peer-reviewed findings, but it’s relevant context for a study built on a manufacturer’s own trial data.
The bottom line
There’s still no cure for MS, though early treatment can slow the disease and ease symptoms. This combined imaging and AI approach won’t change anyone’s treatment today. It’s a research tool, not a diagnostic one.
Future studies still need to confirm it works well across MS types and can reliably track lesions over time. If it does, researchers could finally see whether treatments protect gray matter, not just white matter. That would fill a real gap in what we know about MS drugs already in use.
See our other news articles about artificial intelligence in medical research and care.
