Automated MS imaging finds microstructural complexity beyond lesion burden

Multiple sclerosis lesions can look extensive on magnetic resonance imaging while containing stretches of white matter whose diffusion properties resemble relatively spared tissue. A new study of 100 people with MS suggests that this hidden microstructural complexity is not closely tied to conventional measures of disease burden: the proportion of “reference-like” tract segments within lesions showed no statistically significant monotonic association with either lesion volume or lesion count. The finding offers a possible imaging route into the clinico-radiological paradox, in which MRI-visible lesions do not consistently predict a patient’s symptoms or disability.

Researchers built an automated pipeline that first delineates MS plaques on fluid-attenuated inversion recovery, or FLAIR, MRI and then examines diffusion tensor imaging, or DTI, along fiber pathways passing through those lesions. The DTI analysis used fractional anisotropy, which describes how directional diffusion is organized, and mean diffusivity, which reflects the overall mobility of water. Together, these measures can reveal tissue changes that are not apparent from lesion size alone, including axonal loss, myelin damage and edema, although they do not directly identify a specific biological process.

The lesion-detection component was a modified 2.5D U-Net trained and evaluated on 115 FLAIR scan series from 75 people with MS. It processed three neighboring slices together, combining local anatomical context with lower computational demand than a fully three-dimensional network. Residual blocks, attention gates and deep supervision were incorporated to improve feature propagation, emphasize subtle lesions and learn patterns at multiple scales. Patient-level splitting prevented scans from the same individual entering both training and testing sets, an important safeguard when some participants contributed several examinations.

At inference, the model generated 24 rotated and flipped predictions and combined them through a self-ensembled lesion-fusion strategy. The final configuration used a 0.5 intersection threshold, a 0.4 connected-union threshold and a minimum lesion size of 15 voxels. On the held-out test set, the network achieved a Dice similarity coefficient of 0.732, with a 95% confidence interval of 0.689 to 0.777. Its voxel-wise true-positive rate was 0.781 and lesion-wise true-positive rate was 0.831, while volume correlation reached 0.954, with a 95% confidence interval of 0.914 to 0.995.

Those detection figures were nominally higher than values reported for a contextual benchmark: a Dice coefficient of 0.682, voxel-wise true-positive rate of 0.613, lesion-wise true-positive rate of 0.542 and volume correlation of 0.864. The comparison was descriptive rather than a head-to-head trial, and the stronger sensitivity came with a trade-off. Positive predictive value was 0.710 in the proposed model, and the lesion-wise false-positive rate was 0.285. The authors therefore emphasized the high volume correlation, which indicates that the extra detections were relatively small and did not substantially distort estimated whole-brain lesion load.

For the microstructural analysis, the researchers reconstructed whole-brain fiber pathways using constrained spherical deconvolution and deterministic tractography. Streamlines were seeded in white matter with fractional anisotropy above 0.3, tracked with a 0.5-mm step size and a maximum turning angle of 25 degrees, and terminated when they entered voxels with fractional anisotropy below 0.20. Only streamlines between 30 and 250 mm were retained. The diffusion maps and reconstructed pathways were then registered into FLAIR space so that each intralesional segment could be assigned the mean fractional anisotropy and mean diffusivity of the voxels it traversed.

To define what counted as reference-like, the investigators used a subject-specific internal-capsule segment of the corticospinal tract. Atlas registration located this region, and any reconstructed pathway intersecting an MS lesion was excluded, leaving lesion-spared streamlines as the internal reference. Because the two DTI measures vary naturally between people, fractional anisotropy and mean diffusivity were normalized within each participant using the combined distribution of lesion-spared reference and intralesional segments. Euclidean distance was then used within that individualized feature space, rather than as a direct cross-patient biological scale.

Six operational approaches were evaluated: K-means, Gaussian mixture models, agglomerative clustering, spectral clustering, density-based clustering with noise, or DBSCAN, and an adaptive threshold method based on the reference distribution. After fitting each model, the cluster whose centroid was closest to the subject’s lesion-spared reference was designated the reference-like intralesional segment set. The percentage of lesion-crossing segments assigned to that set became the proposed measure. Results depended markedly on the algorithm: median reference-like proportions were 26% for K-means and 22% for the Gaussian mixture model, whereas DBSCAN and agglomerative clustering sometimes produced estimates near either 0% or 100%.

Across all six methods, no association with macroscopic burden survived statistical correction. For total lesion volume, Spearman correlation coefficients ranged from −0.04 to 0.25, but all adjusted p-values were at least 0.161. For lesion count, the coefficients ranged from −0.01 to 0.20, with all adjusted p-values at least 0.212. K-means and Gaussian mixture models produced the most stable proportions, while spectral clustering showed a bimodal-like pattern and sometimes classified every tract as reference-like. The study therefore supports a dissociation between the amount of visible lesion and the operationally measured microstructural similarity within lesions, rather than proving that either quantity predicts clinical outcomes.

Several limitations constrain how the result should be interpreted. The study was a retrospective secondary analysis of two public, de-identified imaging datasets, and the DTI cohort contained no patient-level disability scores or cognitive assessments, preventing direct testing of the proposed index against symptoms. Its sample of 100 people may also limit generalizability. In addition, comparison with the internal-capsule reference measures similarity to a highly intact, strongly anisotropic white-matter region; it is not a universal measure of healthy tract biology. Uncorrected spatial variations in diffusion encoding can bias mean diffusivity and tensor orientation, while segmentation false positives could influence downstream tract measurements. The investigators identified longitudinal studies with clinical outcomes, test–retest assessment and spatial diffusion calibration as necessary next steps before the metric can be evaluated as a predictor of disability or disease progression.

Subject of Research: Automated segmentation and diffusion-tensor analysis of microstructural heterogeneity in multiple sclerosis lesions

Article Title: Revisiting the MS paradox: automated 2.5D FLAIR segmentation and DTI microstructural profiling reveals dissociation between lesion burden and reference-like intralesional tract profiles

Article References: Toscano-Flores, E. A., Mireles-Ramírez, M. A., Guerrero-García, J. D. J., & Alvarez-Padilla, F. J. (2026). Revisiting the MS paradox: automated 2.5D FLAIR segmentation and DTI microstructural profiling reveals dissociation between lesion burden and reference-like intralesional tract profiles. Frontiers in Human Neuroscience, 20, Article 1901702. https://doi.org/10.3389/fnhum.2026.1901702

Image Credits: AI Generated

DOI: 10.3389/fnhum.2026.1901702

Keywords: multiple sclerosis, MRI lesion segmentation, diffusion tensor imaging, FLAIR, tractography, microstructure, clinico-radiological paradox, deep learning, white matter, Revisiting, paradox, automated