Visual Fazekas Ratings Closely Mirror AI White Matter Hyperintensity Burden
Visual Fazekas scores closely matched the amount of white matter damage measured by artificial intelligence (AI) in patients with acute ischemic stroke. However, differences between brain regions suggest that these two methods provide useful but different information and should not be treated as interchangeable.
The visual Fazekas scale measures how severe white matter changes appear on brain scans. AI segmentation measures their volume as a continuous number. Researchers wanted to find out how closely the two methods matched across different brain regions and overall. They also studied how consistently different readers scored the scans and how often both methods identified patients with a high amount of white matter damage.
The study looked back at medical records and included 361 patients with acute ischemic stroke who had undergone T2-weighted fluid-attenuated inversion recovery (FLAIR) magnetic resonance imaging. The median age was 67, and 258 patients, or 71.5 percent, were men.
Two human readers independently gave each scan separate Fazekas scores for areas around the brain's ventricles, deeper areas of white matter, and the brain overall. At the same time, an AI system measured the volume of white matter hyperintensities and adjusted these measurements for the size of each person's skull cavity. The researchers used several statistical methods to measure how closely the results matched, including Spearman correlations, HC3-robust linear trends, quadratic weighted kappa statistics, agreement between two categories, and a matrix testing two-by-three different cutoff combinations.
The researchers used different rules to define a high amount of white matter damage for each method. For visual scoring, a total Fazekas score of 4 or higher counted as high burden. For AI measurements, high burden meant having a total white matter hyperintensity volume, adjusted for intracranial volume, within the highest third of the study group. This cutoff was 1.20 percent.
The first reader's visual scores closely matched the AI measurements for white matter hyperintensities around the brain's ventricles. The Spearman correlation was 0.840, with a 95 percent confidence interval of 0.807 to 0.866. The match was even stronger for the overall amount of white matter damage, with a correlation of 0.880 and a 95 percent confidence interval of 0.851 to 0.902. However, the match was only moderate for damage in deeper areas of white matter. The correlation was 0.588, with a 95 percent confidence interval of 0.514 to 0.656. All these relationships were statistically significant, with P values below 0.001.
The two human readers also showed excellent agreement with each other. Their quadratic weighted kappa values ranged from 0.869 to 0.926. Using the main cutoff rules, the visual scores and AI measurements agreed for 326 of the 361 patients. This was an agreement rate of 90.3 percent, with a kappa value of 0.781 and a 95 percent confidence interval of 0.708 to 0.847. Using the second reader's scores produced similar results. However, testing different cutoff values gave kappa values ranging from 0.462 to 0.801.
Overall, this study found that visual Fazekas scores and AI measurements matched closely when assessing total white matter damage and changes around the brain's ventricles. They matched less closely when assessing deeper brain tissue. The level of agreement also changed depending on the cutoff values used. The researchers therefore concluded that visual Fazekas scores and AI-based measurements provide complementary information and should not be treated as interchangeable clinical tools.
Subject of Research: How visual Fazekas scores compare with AI-based measurements of white matter hyperintensities in patients with acute ischemic stroke
Article Title: Regional correspondence between visual Fazekas ratings and AI-derived white matter hyperintensity burden in acute ischemic stroke: a retrospective cross-sectional study
Article References: Liao, Q., Tang, Y., Lian, H., Wang, L., & Liu, J. (2026). Regional correspondence between visual Fazekas ratings and AI-derived white matter hyperintensity burden in acute ischemic stroke: a retrospective cross-sectional study. BMC Neurology. https://doi.org/10.1186/s12883-026-05594-x
Image Credits: AI Generated
DOI: 10.1186/s12883-026-05594-x
Keywords: acute ischemic stroke, Fazekas scale, white matter hyperintensities, artificial intelligence, T2-FLAIR MRI, periventricular hyperintensities, deep white matter, intracranial volume, inter-reader reliability, quadratic weighted kappa, Spearman correlation, lesion segmentation, neuroimaging, cross-sectional study, threshold sensitivity, magnetic resonance imaging
