Artificial intelligence (AI) and multicontrast MRI postprocessing may improve detection of cortical lesions in people with multiple sclerosis (MS), according to findings published in Communications Medicine.

In the study, investigators conducted a retrospective analysis of MRI data from the phase 3 ORATORIO trial, which included 732 participants with primary progressive MS. An 80-participant subset was used to develop and evaluate a framework combining several cortical lesion–sensitive image contrasts, including fluid-attenuated inversion recovery squared (FLAIR²), T1/T2 ratio imaging, AI-generated double inversion recovery, and a newly developed multimodal cortical lesion–enhanced (MMCLE) contrast.

The researchers also trained transformer-based deep learning models to identify and segment cortical lesions across individual and combined image contrasts.

Key findings included the following:

  • In the 80-participant development cohort, investigators identified 1182 cortical lesions, corresponding to a mean of 14.8 lesions per participant.
  • Across the full 732-participant dataset, 10,366 cortical lesions were identified, with a mean of 14.4 lesions per participant.
  • MMCLE produced the highest lesion contrast-to-noise ratio and enabled blinded raters to correctly identify 86.0% of cortical lesions, with an 8.4% false-positive rate.
  • Conventional FLAIR detected 15.2% of lesions, compared with 86.0% using MMCLE.
  • AI models performed best when multiple lesion-sensitive contrasts were analyzed together rather than individually.
  • Lesion count and volume measurements showed good reproducibility across 1.5-T and 3-T MRI scanners and between 2-dimensional and 3-dimensional FLAIR acquisitions.

The findings suggest that existing clinical trial MRI datasets may contain cortical lesion information that can be recovered through postprocessing and AI-assisted analysis. The authors noted that the analysis lacked a histopathologic reference standard and included only participants with primary progressive MS. Detection of new or enlarging cortical lesions was also less accurate than cross-sectional lesion identification, supporting the need for further validation in independent and broader MS populations.

Source: Dwyer MG, Bergsland N, Bartnik A, et al. Quantifying cortical lesions in multiple sclerosis MRI datasets using multi-contrast post-processing and deep learning. Commun Med. 2026;6:365. doi:10.1038/s43856-026-01683-7