Artificial intelligence-based segmentation of perisinusoidal tissue along the superior sagittal sinus in human brain magnetic resonance imaging.

PURPOSE: Meningeal lymphatic vessels (MLVs) contribute to transporting interstitial fluid and macromolecules accruing in the brain to deep cervical lymph nodes. Dysfunction of MLVs has been associated with neurodegenerative disorders. A dense network of MLVs is embedded in the tissue immediately adjacent to the superior sagittal sinus (SSS), i.e., the perisinusoidal tissue (PT). The PT can be visualized on non-contrast-enhanced T2-FLAIR MRI. However, volumetric analysis of the PT has so far been limited to manual segmentation and was thus not feasible in larger cohorts. Therefore, we trained a deep neural network for automated segmentation of the PT along the SSS.

METHODS

We established a detailed manual segmentation protocol representing the reference standard in the evaluation. Four different expert raters performed manual segmentation of perisinusoidal hyperintensities in 35 individuals (training cohort 27, test cohort 8) based on 3D T2-FLAIR MRI. To enable automated segmentation, we trained a 3D fully convolutional neural network.

RESULTS

When comparing different human raters’ segmentations, the mean Dice-score was 0.755 (SD = 0.050), reflecting the interrater reliability. Comparison of manual segmentations and algorithm predictions yielded a mean Dice-score of 0.756 (SD = 0.047). Volumetric measures from rater and algorithm segmentations revealed a Pearson correlation coefficient of 0.927 (95% CI = 0.642–0.987).

CONCLUSION

Our findings demonstrate that volumetric analysis of the perisinusoidal FLAIR-hyperintensities containing MLVs using deep learning-based segmentation is technically feasible and achieves good accuracy, comparable to human performance. This approach aims to enable time efficient volumetric analysis of dural lymphatic structures in large-scale prospective population and interventional studies.

SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s00234-026-03912-1.