The abstract is read at the publisher; the summary is JClub's.
A single nnU-Net trained on 2D and isotropic 3D FLAIR achieved mean Dice 0.73 (SD 0.10) for WMH segmentation in an internal ADNI/WWISH/UK Biobank test set.
- Why it matters: WMH are pooled across aging, CSVD, and dementia cohorts with different FLAIR protocols, yet it is unclear whether segmentation accuracy, error, and downstream clinical associations stay stable across 2D versus 3D acquisitions and algorithm choices.
- What they did: The model used FLAIR plus T1 inputs and was tested on 92 held-out scans (66 participants) and an independent ADNI3 3D set (n = 25), with voxel Dice, volume agreement, and lesion detection benchmarked against SHIVA-WMH, segcsvdWMH, and FreeSurfer WMH-SynthSeg.
- The result: Performance was similar on 2D versus 3D FLAIR (Dice 0.75 vs. 0.71), bias was smaller than comparators in WWISH paired scans (1.8 mL vs. 4.7–9.9 mL), and model-derived WMH volume predicted incident dementia (OR 1.40 per SD, 95% CI 1.11–1.77; likelihood-ratio p = 0.009).
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.