AI finds hidden MS brain lesions on standard scans

Clinical takeaway: This method runs on standard MRI sequences already collected in routine MS imaging at 1.5T and 3T, so it could eventually widen access to cortical-lesion assessment that normally requires 7T or specialized scans. It is currently validated for research use, rather than clinical diagnosis.

Cortical lesions in the brain track closely with disability and cognitive decline in multiple sclerosis. Yet clinicians have rarely been able to detect them, since conventional MRI captures white-matter damage far better than gray. This blind spot matters, because most disease-modifying drugs act on the white-matter lesions that clinicians can routinely see. But these treatments have done far less for disability. Researchers have now shown that image post-processing and AI can pull cortical lesions out of standard MRI scans.

Cortical lesions show up in up to 96% of people with MS on 7T imaging, and they have been part of the diagnostic criteria since 2017. Their use in practice has stalled anyway, because conventional MRI cannot resolve them and the alternatives are impractical. 7T scanners are rare, and specialized sequences add scan time with no standardization across machines. Against that barrier, the researchers asked whether combining post-processing methods with AI could surface the lesions on scans already in hand.

Applied to scans already collected for a past MS drug trial, the pipeline detected 10,366 cortical lesions, roughly 14 per patient, on images never meant to show them. The method works by generating several enhanced views from the standard scans, each processed to make cortical lesions stand out.

The team tested detection two ways. In a human blinded review, raters using the best-performing view correctly flagged 86% of known lesions while falsely identifying 8.4%, and on that view lesions stood out more than twice as sharply as on conventional scans. Separately, an automated AI detector showed why combining views matters: reading all the enhanced views together, it caught 84.6% of lesions, well above the 68.9% it managed on the single best view alone.

The source scans came from ORATORIO, a phase 3 ocrelizumab trial in primary progressive MS. Researchers ran detailed lesion analysis on 80 patients and trained the deployable model on all 732 patients, using images from 21 scanner models at 1.5T and 3T. Detection stayed stable across scanners. Because no histopathology or 7T comparison was available, how many true lesions the method catches or misses could not be confirmed.

Before the method reaches the clinic, it needs validation against a definitive reference. Histopathology or 7T imaging could serve, but this study had neither. That step would establish its true accuracy. It also needs testing in relapsing-remitting MS patients. The cohort here was entirely primary progressive. More immediately, the approach could be useful to reanalyze scans from existing trials to see how current drugs act on the cortical lesions tied to disability.

"We have all been very frustrated, knowing that these cortical lesions were there but not being able to see them," said Michael Dwyer, PhD, associate professor of neurology and biomedical informatics at the University at Buffalo and first author of the study.

He added, "Generative AI is very powerful because it can look between the scans and detect tiny differences between them. Because it sees those minor discrepancies, AI can reveal that there's something going wrong there, that the tissue is not behaving like healthy tissue."

Source: Dwyer MG, et al. (2026 Jul 7) Commun Med. Quantifying cortical lesions in multiple sclerosis MRI datasets using multi-contrast post-processing and deep learning