Automatic segmentation of white matter hyperintensities: validation and comparison with state-of-the-art methods on

Philippe Tran1, Urielle Thoprakarn2, Emmanuelle Gourieux3

  • 1Qynapse, Paris, France; Equipe-projet ARAMIS, ICM, CNRS UMR 7225, Inserm U1117, Sorbonne Université UMR_S 1127, Centre Inria de Paris, Groupe Hospitalier Pitié-Salpêtrière Charles Foix, Faculté de Médecine Sorbonne Université, Paris, France.

Neuroimage. Clinical
|January 20, 2022
PubMed

Insights

A new algorithm, WHASA-3D, accurately segments white matter hyperintensities (WMH) in brain MRIs for both Multiple Sclerosis (MS) and age-related conditions. This tool offers improved performance over existing methods, aiding clinical diagnosis and monitoring.

Area of Science:

  • Medical Imaging
  • Neuroscience
  • Computer Vision

Background:

  • White matter hyperintensities (WMH) are visible on MRI, associated with Multiple Sclerosis (MS) and age-related cognitive decline.
  • Accurate WMH quantification is crucial for disease monitoring and clinical trials.
  • Manual segmentation is time-consuming and prone to variability, necessitating automated solutions.

Purpose of the Study:

  • To introduce WHASA-3D, a novel, fast, and robust automatic segmentation tool for WMH detection.
  • To validate WHASA-3D's performance on both MS lesions and age-related WMH using 3D T1-weighted and T2-FLAIR images.
  • To compare WHASA-3D against existing state-of-the-art WMH segmentation methods.

Main Methods:

  • WHASA-3D builds upon the WHASA method, incorporating non-linear diffusion and watershed parcellation, refined with geodesic dilation.
  • The algorithm was evaluated on a heterogeneous dataset of 60 subjects (dementia, MS, elderly) using volume and spatial agreement metrics.
  • Direct comparison with six supervised and unsupervised WMH segmentation algorithms was performed on the MS patient cohort.

Main Results:

  • WHASA-3D demonstrated improved performance over WHASA, with higher Dice overlap (0.67 vs 0.63) and better volume agreement (ICC 0.96 vs 0.78).
  • Compared to other methods on the MS database, WHASA-3D achieved the highest volume agreement (ICC=0.95) and average Dice (0.58) at default settings.
  • Optimized state-of-the-art methods showed improved performance, but WHASA-3D remained competitive, especially at default settings.

Conclusions:

  • WHASA-3D is a reliable and user-friendly tool for automated segmentation of both MS lesions and age-related WMH.
  • The algorithm shows significant potential for clinical implementation in routine practice and trials.
  • Further validation on larger datasets is recommended to confirm findings.

Related Concept Videos