Magnetic resonance imaging (MRI) plays a vital role in diagnosing and monitoring multiple sclerosis (MS), offering key indicators of disease progression. However, vast collections of single-contrast MRI scans housed in hospitals have largely remained underutilised for research, as extracting clinically relevant biomarkers from them has proven challenging. This is particularly significant in MS, where measuring lesion load and brain volume changes is essential for evaluating disease activity and treatment outcomes. A new deep learning model, MindGlide, addresses this issue by enabling quantitative analysis of routine MRI scans, regardless of their contrast type or quality, unlocking the potential of these data for both clinical research and care.
Expanding Research Access Through Deep Learning
MindGlide was created to overcome the technical constraints associated with MRI scans of varying contrasts and resolutions. The model was trained using a large and diverse dataset comprising 4,247 real MS brain scans and 4,303 synthetic images. These included T1-weighted and FLAIR MRI contrasts, gathered from 592 MRI scanners. The training population represented multiple MS subtypes, with patients spanning relapsing-remitting, secondary progressive and primary progressive forms. MindGlide uses a three-dimensional convolutional neural network architecture and can segment brain structures and white matter lesions from any single MRI contrast, including T2-weighted and proton density (PD) scans, even if these were not used during training. It produces segmentation results in less than a minute and requires no preprocessing by the user.
Must Read: The Role of AI in Multiple Sclerosis MRI Assessment
The model was externally validated using 14,952 scans from two progressive MS clinical trials and a routine-care paediatric MS cohort, encompassing 1,001 patients from 186 scanners. These included various MRI contrasts and slice thicknesses, highlighting MindGlide's ability to generalise across diverse clinical scenarios. This broad applicability allows for the repurposing of archived routine scans, supporting research and potentially lowering the cost of trials by reducing the need for multi-contrast acquisitions.
Clinical Relevance and Sensitivity to Treatment Effects
MindGlide was effective in detecting clinically meaningful treatment effects in both progressive and paediatric MS populations. In the secondary progressive MS (SPMS) and primary progressive MS (PPMS) trials, the model identified slower lesion volume growth and reduced grey matter loss in patients receiving treatment compared to those on placebo. These findings were consistent across several MRI contrasts, including T2-weighted, FLAIR, PD and T1-weighted scans, demonstrating the model’s ability to extract meaningful data even from contrasts not traditionally used to assess such changes.
In the routine-care paediatric cohort, MindGlide revealed that patients receiving high-efficacy treatment experienced stable lesion volumes over time, while those on moderate-efficacy treatment showed increases. It also detected cortical grey matter loss across all treatment groups, with varying rates depending on the MRI contrast and treatment type. Similarly, it was found that deep grey matter volume remained more stable in patients on high-efficacy therapy, while it declined in those on moderate treatments. These results underline MindGlide’s utility in identifying subtle changes in brain tissue over time, even from routine scans with inconsistent resolution.
Comparative Performance and Broader Applications
When compared to existing segmentation tools such as SAMSEG and WMH-Synthseg, MindGlide consistently achieved stronger performance. It yielded higher Dice scores and demonstrated superior accuracy in segmenting lesion volumes, particularly when benchmarked against expert-labelled ground truth data. MindGlide-derived volumes also showed stronger correlations with clinical disability scores, as measured by the Expanded Disability Status Scale (EDSS), across multiple contrasts and brain regions. This included both lesion loads and grey matter volumes, with significant correlations seen in cortical and deep grey matter.
In addition to its accuracy, MindGlide proved to be more robust in handling challenging clinical data. In a visual assessment of 433 routine-care MRI scans, WMH-Synthseg failed to process 15% of the scans, particularly those with thick slices, whereas MindGlide failed on just 1%. This difference highlights MindGlide’s reliability in real-world settings, where image quality and scan protocols often vary widely.
The model also demonstrated consistent results across different contrasts when measuring percentage brain volume change over time, with high intraclass correlation coefficients. It performed similarly on both two-dimensional and three-dimensional scans, though reduced sensitivity was noted in detecting brain atrophy on lower-resolution 2D images. Despite this limitation, the model’s ability to work effectively with fewer contrasts and lower-resolution data offers substantial benefits for clinical environments where high-resolution imaging may not be available.
MindGlide significantly advances the use of routine MRI scans in multiple sclerosis research and care. It enables accurate, contrast-agnostic segmentation of brain structures and lesions, allowing the repurposing of previously unusable MRI archives for quantitative analysis. This capability facilitates large-scale studies, enhances clinical trial efficiency and improves biomarker monitoring across various healthcare settings. Publicly available, MindGlide ensures widespread accessibility in academic and clinical environments and could eventually provide broader insights into neurodegenerative conditions by reusing existing imaging data.
Source: nature communications
Image Credit: Freepik
