Summary: An international team of scientists and clinicians has created a generative artificial intelligence framework that reveals cortical lesions previously hidden in routine MRI scans. By synthesizing subtle, sub-visual differences across multiple MRI contrasts, the AI acts like a computational lens that extracts diagnostic information from standard images and exposes an otherwise invisible layer of multiple sclerosis (MS) pathology.
Key Facts
- The clinical blind spot: Although modern MS therapies have substantially slowed disability progression over the past decade, their development and monitoring have emphasized reductions in white matter lesions because cortical gray matter lesions remained largely undetectable on standard clinical MRI.
- The spatial loophole: Individual MRI slices of the cortex can appear normal, but generative AI models analyze relationships across multiple image contrasts. By detecting tiny, sub-visual discrepancies in tissue behavior, the AI synthesizes a missing pathological picture of the cortex.
- MMCLE breakthrough: The team integrated several advanced image-processing approaches into a new protocol called MMCLE (Multimodal Cortical Lesion Enhancement), designed to boost cortical lesion visibility on conventional scans.
- More than 11,000 lesions uncovered: Applied to the ORATORIO trial dataset, MMCLE identified roughly 15–20 previously invisible cortical lesions per patient, totaling over 11,000 lesions across the cohort—far beyond the white matter markers visible on original scans.
- Works with legacy MRI data: Because MMCLE operates on standard clinical MRI contrasts, hospitals and clinics do not need to buy costly new imaging hardware. Existing scans, both historic and current, can be reprocessed to reveal a patient’s fuller disease burden.
- Implications for trials and drug development: Senior author Robert Zivadinov emphasizes that revealing cortical pathology in legacy trial data could reshape how pharmaceutical companies evaluate past results and design future therapies that specifically protect gray matter and address cognitive decline.
Source: University at Buffalo
One of the unsettling realities about multiple sclerosis is that the portion of the brain most informative about disease severity and future impact—the cortex—has often been invisible to clinicians using routine MRI.
Gray matter involvement has long been recognized as a key driver of MS progression and cognitive impairment. Yet standard clinical MRI protocols have traditionally been optimized to detect bright, conspicuous lesions in white matter, leaving cortical lesions largely hidden. As a result, many newer drugs have been evaluated and monitored primarily on their effect on white matter damage.
In a new paper published in Communications Medicine, researchers led by the University at Buffalo describe how they used generative AI and multimodal post-processing to expose these hidden cortical lesions by reanalyzing existing MRI data.
The ability to visualize cortical lesions on routine scans, the team says, is a major advance because cortical damage is closely linked to cognitive decline and long-term disability in MS.
“Detecting previously invisible cortical lesions on conventional legacy MRI scans has major implications for both MS research and clinical care,” says Robert Zivadinov, MD, PhD, senior author and director of the Buffalo Neuroimaging Analysis Center. “Being able to see these hidden indicators of disease progression and cognitive impairment is an important step forward.”
Although cortical lesions were recognized by pathologists and later incorporated into diagnostic criteria, clinical MRI historically lacked the contrast and resolution needed to display them reliably. Histopathology of postmortem tissue long showed cortical damage, but there was no reliable way to visualize that damage in living patients with standard scans.
Seeing ongoing damage that was previously hidden
“We’ve all been frustrated knowing these cortical lesions exist but not being able to see them,” says Michael G. Dwyer, PhD, first and corresponding author and associate professor of neurology and biomedical informatics. “Pathology told us the cortex was affected, but conventional MRI couldn’t show it. Our AI-based approach pulls those signals out of routine images so clinicians and researchers can finally access these data.”
The methods used build on prior international work and focus on extrapolating information from how multiple contrasts of the same brain relate to one another—information that is not visible on any single image. The team combined existing post-processing contrasts such as FLAIR squared and T1/T2 ratio with AI-derived synthetic contrasts and added their new multimodal cortical lesion enhancement (MMCLE) contrast. They then applied transformer-based semantic segmentation to automate lesion detection and delineation.
These techniques were tested on MRI scans from the large phase III ORATORIO clinical trial (which evaluated the drug ocrelizumab) that included more than 700 participants. While the original scans primarily showed white matter lesions, the MMCLE workflow consistently revealed roughly 15–20 cortical lesions per participant, yielding more than 11,000 cortical lesions across the dataset.
“Generative AI is powerful because it evaluates the subtle differences between multiple images,” says Dwyer. “It identifies tiny discrepancies that indicate tissue is not behaving like healthy cortex and synthesizes what had been missing on conventional images.”
The research team included clinicians and scientists from academia and industry, including contributors from Genentech. Zivadinov emphasizes that the breadth of expertise in the collaboration helped bring the work to fruition.
“Revealing such extensive invisible pathology will influence how we review past trial data and how future studies are designed,” he says. “This opens the door to therapies targeted at protecting gray matter and preserving cognition.”
Co-authors from the University at Buffalo include Niels P. Bergsland, PhD; Alexander Bartnik, PhD; and Dejan Jakimovski, MD, PhD. Additional contributors include Samantha Noteboom, Menno M. Schoonheim, Martijn D. Steenwijk (MS Center Amsterdam), and Jinglan Pei and David Clayton (Genentech).
Funding: The research was supported in part by Genentech.
Key questions answered
Q: Why have cortical (gray matter) lesions remained invisible on routine hospital MRI?
A: Standard clinical MRI provides strong contrast for conspicuous white matter lesions but lacks the contrast and resolution to separate subtle cortical changes from surrounding healthy tissue on a single image. Histopathology has long shown cortical damage, but until now clinicians lacked a practical imaging method to visualize it in living patients.
Q: How can generative AI “see” lesions not visible on the original scans?
A: The AI analyzes multiple MRI contrasts together and evaluates the mathematical relationships and tiny differences between them. By detecting consistent sub-visual discrepancies in how cortex tissue behaves across contrasts, the AI synthesizes a representation of cortical lesions that are not discernible on any single original image.
Q: What does this mean for someone living with MS?
A: This advancement can immediately improve clinical assessment because it works on legacy MRI scans already present in hospital archives. Clinicians can reprocess existing scans to gain a more complete picture of disease progression and cognitive risk. For research and drug development, the ability to quantify cortical damage in historic trial datasets may help explain treatment effects and guide development of therapies that protect gray matter.
Editorial notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full by the editorial team.
- Additional explanatory context was added by staff for clarity.
About this AI and multiple sclerosis research news
Author: Ellen Goldbaum
Source: University at Buffalo
Contact: Ellen Goldbaum – University at Buffalo
Image: Image credited to Neuroscience News
Original research: Open access. Title: “Quantifying cortical lesions in multiple sclerosis MRI datasets using multi-contrast post-processing and deep learning” by Michael G. Dwyer et al., Communications Medicine. DOI: 10.1038/s43856-026-01683-7
Abstract
Quantifying cortical lesions in multiple sclerosis MRI datasets using multi-contrast post-processing and deep learning
Background
Multiple sclerosis affects both white and gray matter of the central nervous system. Despite strong evidence that cortical lesions contribute to progression and cognitive decline, most clinical trials have been unable to evaluate cortical lesions because standard MRI protocols do not visualize them reliably. Recent post-processing approaches—synthetic contrasts and AI-based enhancements—have shown promise in improving cortical lesion detection on conventional MRI and enable reanalysis of existing trial data to address mechanistic and treatment-related questions.
Methods
The authors assessed the feasibility of combining and extending several promising post-processing methods in a unified framework using data from the multicenter phase 3 ORATORIO trial (n=732; development subset n=80). They evaluated FLAIR squared, T1/T2 ratio, an AI-derived double inversion recovery (AI-DIR), and introduced the combined MMCLE contrast. Transformer-based semantic segmentation was applied to automate detection and delineation of cortical lesions.
Results
At baseline, the unified approach detected an average of 14.8 ± 20.72 cortical lesions per participant, achieving an 86.0% true positive rate and an 8.4% false positive rate for blinded MMCLE review, using simultaneous interpretation of all contrasts as the reference. The methods demonstrated high reproducibility across field strengths and acquisition types (ICC 88.8–92.5%).
Conclusions
The study confirms that cortical lesions can be visualized and quantified on conventional MRI using multi-contrast post-processing and deep learning. Simultaneous use of multiple contrasts improves quantification and enables automated lesion detection, opening opportunities for revisiting legacy datasets and improving clinical assessment and trial design.