A University at Buffalo deep learning pipeline reads cortical lesions (damage in the brain's gray matter outer layer) on conventional MRI, exposing the gray matter damage every approved MS drug's white matter endpoint (the standard MRI benchmark MS
A patient with stable multiple sclerosis walks into a clinic, gets a routine brain MRI, and hears what they have heard for years: "Looks fine." A deep-learning pipeline from the University at Buffalo then analyzes the same images and surfaces dozens of cortical lesions a radiologist's eye never registered. The gap between the clinical verdict and the algorithm's count is the new dividing line in MS care.
A peer-reviewed paper in Nature Communications Medicine from the Buffalo Neuroimaging Analysis Center reports the method: a multi-contrast post-processing step combined with a deep-learning model that surfaces cortical, or gray-matter, lesions on conventional MRI. The same group had posted the approach as a medRxiv preprint in April 2025; the Nature paper is the peer-reviewed version. The lead investigator is Robert Zivadinov, who directs the Buffalo Neuroimaging Analysis Center, and the institutional record is on the BNAC publications page.
Standard clinical MRI cannot see cortical lesions. The T2-weighted and FLAIR sequences used to count MS lesions were designed around white-matter disease, and gray-matter lesions in the cortex produce far less signal contrast. The BNAC pipeline applies a multi-contrast post-processing step and a deep-learning model trained to identify lesion signatures in that gray-matter tissue, and the team reports the ability to quantify cortical lesions on the same legacy scans a clinic already has on file. GEN News covered the result as the first peer-reviewed demonstration that the gap is closable on existing scans.
MS therapies were designed, tested, and approved using white-matter lesion counts as a primary measure. A trial that cut white-matter lesions by 90% and left the patient progressively disabled was scored as a success in lesion terms. The Nature paper does not claim the drugs are worthless. It establishes that the imaging endpoint never measured the half of the disease driving the disability most patients report: cognitive decline, fatigue, and progressive loss of function the white-matter count failed to capture. The independent context for the algorithmic problem is the broader effort to benchmark cortical lesion segmentation, surveyed in a separate arXiv preprint on benchmarking deep-learning cortical lesion MRI segmentation in MS. That work treats cortical lesion segmentation as an open, contested benchmark, not a solved problem.
The Nature paper is a research demonstration on curated datasets, not a regulatory clearance. A clinical-assay version would need independent replication, prospective trials on real-world scanner fleets, and a defined path into radiology workflow. The model reads lesions; it does not yet prescribe treatment. The link from a countable lesion to a treatable one remains a separate, open question.
The paper turns an open question into a measurable one. The MS field has spent three decades approving drugs against a white-matter endpoint. Cortical lesion load can now be quantified on the same scans that produced those endpoints, which makes it possible to ask, retrospectively on legacy data and prospectively on new trials, whether any approved therapy touched the gray-matter compartment at all. Independent replication and adoption as a secondary endpoint in a prospective MS drug trial are the next filters. The second half of the disease was always inside the scans the standard workflow could not resolve.