By Mary Henderson
Study results presented Thursday morning demonstrate the utility of AI in assisting neuroradiologists in longitudinal assessments of the inflammation, demyelination and axonal injury in patients with multiple sclerosis (MS).
“The radiologist’s role has become increasingly important in quantifying the progression of MS and detecting subclinical activity,” said Matheus Luis Silva, MD, a neuroradiology lead at OneLaudos in Sao Paulo. “Detecting new and growing lesions and measuring brain atrophy with MRI guides therapy for MS patients.”
According to Dr. Silva, there’s a clinical need for tools to improve consistency and support assessments. “Manually measuring brain volume and counting demyelination-related lesions is time-consuming and prone to intra-observer variability. Small differences in interpretation can alter treatment choices,” he said.
The retrospective, multi-center study included 39 patients with relapsing-remitting MS who had baseline and follow-up MRI within two years of each other. In the initial assessment, three blinded, experienced neuroradiologists independently reviewed all cases at baseline and follow-up. They estimated brain atrophy and the number and location of new and persistent white matter lesions.
Dr. Silva and his team used a commercial open source AI platform to extract brain volume and lesion metrics, producing volumetric maps and assigning a progression category. After 12 weeks, radiologists re-evaluated a random subset of cases using prior AI results to assess changes in diagnostic interpretation, reading time and diagnostic confidence.
“AI does not replace radiologists’ judgment, but rather stabilizes the ground upon which that judgment is made.”
Matheus Luis Silva, MD
The researchers compared the metrics of AI and human readers, including lesion count and brain volume. They also compared evaluators to assess consistency and reproducibility.
On volumetry, both human and AI measurements showed good intra-observer consistency. The results of the repeat evaluation were high, similar to the initial findings. According to Dr. Silva, this reflected stable individual performance. AI-derived estimates also exhibited strong internal stability.
“Both human and AI measurements showed good intra-observer consistency,” he said. “AI performance showed consistency comparable to human readings with minimal dispersion and no relevant bias.”
He said establishing standards for the evaluations helped create consistent performance. For lesion count, the comparison between clinical assessment and AI-derived metrics revealed moderate correlations, particularly involving the callosal index, volumetric parameters and lesion count.
“Correlations were not universal, underscoring that AI is not redundant but complementary to expert assessment, capturing different dimensions of MS activity,” Dr. Silva said. “AI does not replace radiologists’ judgment, but rather stabilizes the ground upon which that judgment is made.”
Dr. Silva noted that human readers who agreed with AI findings were almost twice as likely to report benefits, such as improved confidence or faster reading times. The majority of those who disagreed with AI reported no perceived benefit.
“In this case, acceptance and usefulness travel together,” he concluded. “AI tools that perform automated segmentation and volumetry offer the promise of faster, more consistent analyses.”
Access the session, “The Impact of Artificial Intelligence on Longitudinal Assessment of Multiple Sclerosis: A Comparison of Radiologist Report and Effect on Blinded Retrospective Interpretation,” (R3-SSNR13-4) on demand at RSNA.org/MeetingCentral.
© 2025 RSNA.
The RSNA 2025 Daily Bulletin is the official publication of the 110th Scientific Assembly and Annual Meeting of the Radiological Society of North America. Published online Sunday, November 30 — Thursday, December 4.
The RSNA 2025 Daily Bulletin is owned and published by the Radiological Society of North America, Inc., 820 Jorie Blvd., Suite 200, Oak Brook, IL 60523.
