A novel artificial intelligence (AI) tool can help interpret and assess treatment efficacy in patients with multiple sclerosis (MS), new research suggested.
MindGlide is a deep learning model that can extract key biomarkers, including lesion load and brain volume changes, from MRI scans acquired during the care of patients with MS.
"Using MindGlide will enable us to use existing brain images in hospital archives to better understand multiple sclerosis and how treatment affects the brain," first author Philipp Goebl, PhD student, UCL Queen Square Institute of Neurology and UCL Hawkes Institute, both in London, England, said in a press release.
The study was published online on April 7 in Nature Communications.
MS affects over 2.8 million people globally, with a notable burden on younger adults. MRI biomarkers are central to phase 2 and 3 trials, with multicontrast scans used to track new or enlarging lesions and neurodegeneration. However, these protocols are time- and resource-intensive, limiting use in routine care, the investigators noted.
They added that simplified approaches using single-contrast volumetry could help extract meaningful atrophy data from existing clinical scans, expanding research opportunities and reducing trial costs. Routine monitoring focuses on lesion activity, brain atrophy -- an important marker of worsening disability -- is often underutilized due to scan heterogeneity and lack of automation tools.
MindGlide was developed to address these limitations, enabling detection of clinically relevant biomarkers in " highly heterogeneous scans independent of contrasts, resolutions, and qualities."
The investigators trained MindGlide using an initial dataset of 4247 brain MRI scans from 2934 patients with MS across 592 scanners. During this process, the tool trained itself to identify markers of MS.
The goal of the study was to validate MindGlide using 14,952 images from 1001 patients drawn from two clinical trials in primary and secondary progressive MS, as well as a routine-care MS dataset.
When tested against expert-labeled lesion volumes, the model outperformed two state-of-the-art models -- SAMSEG, which identifies and outlines lesions and brain structures across MRI contrast, and WMH-SynthSeg, which detects small bright spots on scans -- in lesion volume detection, sensitivity, and dice score.
In the clinical trials, MindGlide detected treatment effects on T2-lesion accrual and cortical and deep gray matter volume loss. Results showed that AI tool accurately assessed effects of disease-modifying therapies in clinical trials and routine-care cohorts, particularly neurodegeneration and disease activity.
"The process took 5-10 seconds per image," investigators noted.
Using dice scores, the tool accurately quantified white matter lesions, cortical and deep gray matter volumes, with MindGlide-derived volumes showing higher correlation with Expanded Disability Status Scale than SAMSEG and WMH-SynthSeg, although "differences were not statistically significant for all comparisons."
Additionally, MindGlide was better than standard tools in quantifying the differences in lesion volume between treatment and placebo groups in the primary progressive MS trial, with 5.31% in FLAIR and 4.62% in T2, and its results were the closest to ground truth labels from expert neuroradiologists (P < .001 for all).
In contrast, longitudinal SAMSEG overestimated with 10.70% in FLAIR (P = .004) and 8.81% in T2 (P = .001), whereas WMH-SynthSeg underestimated with 2.56% in FLAIR and 2.45% in T2 (P < .001 for both).
"We hope that the tool will unlock valuable information from millions of untapped brain images that were previously difficult or impossible to understand, immediately leading to valuable insights into multiple sclerosis for researchers and, in the near future, to better understand a patient's condition through AI in the clinic. We hope this will be possible in the next 5-10 years," said Goebl.
