60/100

Moderate evidence

Cite
More
EmailLinkedInX

Study of 361 stroke patients finds AI and visual brain-scan ratings agree in 90.3%

Study title: Regional correspondence between visual Fazekas ratings and AI-derived white matter hyperintensity burden in acute ischemic stroke: a retrospective cross-sectional study

Evidence type
Observational study
Sample
n = 361
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Regional correspondence between visual Fazekas ratings and AI-derived white matter hyperintensity burden in acute ischemic stroke: a retrospective cross-sectional study

A retrospective cross-sectional study of 361 patients with acute ischemic stroke compared visual Fazekas ratings of white matter hyperintensities on brain MRI with AI-derived volume measurements. Reader 1's total Fazekas score correlated strongly with AI-measured total burden (ρ = 0.880), and 326 patients (90.3%) were classified the same way as high or not high burden.

Read the study abstract

Abstract Background Visual Fazekas ratings assess white matter hyperintensity (WMH) severity, whereas artificial intelligence (AI)-derived segmentation provides continuous measurements. We quantified regional and total visual–AI correspondence, inter-reader reliability, and high-burden concordance in acute ischemic stroke. Methods We included 361 patients with acute ischemic stroke undergoing T2-weighted fluid-attenuated inversion recovery (T2-FLAIR) magnetic resonance imaging (MRI). Two readers independently as…

57Gain side

How the two sides scoreClose call

55Problem side

The gain

In a retrospective study of 361 acute stroke patients, AI-measured white matter hyperintensity burden matched visual Fazekas high-burden classification in 326 patients (90.3%, κ = 0.781).

The problem

In the same 361-patient stroke MRI study, AI measurements correlated only moderately with visual ratings of deep white matter disease (ρ = 0.588), and agreement on high burden varied from κ 0.462 to 0.801 across thresholds.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

Agreement was not uniform. Correlation with deep white matter disease was moderate (ρ = 0.588), and when the thresholds defining high burden were changed, κ ranged from 0.462 to 0.801. The authors conclude that visual ratings and quantitative AI measurements are complementary rather than interchangeable.

The results bear on how stroke imaging might be assessed, since a coarse visual scale and a continuous AI volume do not always label the same patient the same way. A second reader's replication gave similar results.

Main points

  1. The study included 361 patients with acute ischemic stroke who had T2-FLAIR MRI; median age was 67 and 258 (71.5%) were men.
  2. Reader 1's ratings correlated with AI periventricular WMH at ρ = 0.840 and with total WMH at ρ = 0.880.
  3. Correlation with deep WMH was weaker, at ρ = 0.588 (95% CI 0.514–0.656).
  4. Inter-reader reliability between the two readers was excellent, with quadratic weighted κ of 0.869 to 0.926.
  5. Visual high burden was a total Fazekas score of 4 or more; AI high burden was the upper third of the cohort, 1.20% of intracranial volume.
  6. Across alternative thresholds, κ ranged from 0.462 to 0.801.

The full story

Two readers independently assigned periventricular, deep and total Fazekas scores on T2-FLAIR MRI. The AI-derived WMH volumes were normalized to each patient's intracranial volume, and the team used Spearman correlations, robust linear trends, weighted κ and binary agreement statistics.

Agreement was strongest for total and periventricular disease. Under the primary thresholds, 326 of 361 patients agreed (90.3%) with κ = 0.781 (95% CI 0.708–0.847), and a second reader's replication produced similar results.

The deep white matter findings and the 2 × 3 threshold-sensitivity analysis show where the two approaches diverge. The authors conclude that visual ratings and quantitative measurements are complementary rather than interchangeable.

What this doesn’t fix

The AI high-burden cut-off was cohort-specific (the upper third, 1.20% of intracranial volume), and agreement changed when thresholds were varied, so the results depend on the thresholds chosen.

Sources

1 on file
  1. BMC Neurologydoi:10.1186/s12883-026-05594-x
    Peer-reviewed2026-11-17

Talk it over

Three questions to start on, two more if you have time
AI matched visual high-burden ratings in 326 of 361 patients, yet correlation with deep white matter disease was only ρ = 0.588. How would you weigh the strong overall agreement against that weaker result?

Start with a question

Study of 361 stroke patients finds AI and visual brain-scan ratings agree in 90.3%https://truvace.com/item/regional-correspondence-between-visual-fazekas-ratings-and-ai-derived-white-matt

  1. AI matched visual high-burden ratings in 326 of 361 patients, yet correlation with deep white matter disease was only ρ = 0.588. How would you weigh the strong overall agreement against that weaker result?

    Answer in the thread
  2. The AI high-burden cut-off was the upper third of this cohort, 1.20% of intracranial volume, and κ ranged from 0.462 to 0.801 across thresholds. What does that say about how far these results would carry to another hospital?

    Answer in the thread
  3. If you were a stroke clinician reading these MRI scans, in which situations would you rely on the Fazekas score, the AI volume or both, given the authors call them complementary?

    Answer in the thread
If you have more time
  • The study was retrospective and cross-sectional, with a median age of 67 and 258 men among 361 patients. What patients would you want included before judging how well it applies?

    Answer in the thread
  • What follow-up measure, such as patient outcomes, would show whether the visual score or the AI volume matters more for care?

    Answer in the thread

What readers say