Fully convolutional network ensembles for white matter hyperintensities segmentation in MR images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• White matter hyperintensities (WMH) are common in aging brains and linked to neurological disorders.
• Accurate detection of WMH is crucial for diagnosis and monitoring.
• Current detection methods can be time-consuming and subjective.

Purpose of the Study:

• To develop and evaluate an automated deep learning system for WMH detection.
• To assess the algorithm's performance and generalizability across different scanners.
• To make the developed software and models publicly available.

Main Methods:

• Utilized deep fully convolutional networks and ensemble models.
• Employed fluid attenuation inversion recovery (FLAIR) and T1 magnetic resonance (MR) imaging modalities.
• Validated the algorithm in the WMH Segmentation Challenge at MICCAI 2017.

Main Results:

• Achieved 1st rank in the WMH Segmentation Challenge.
• Obtained an average Dice score of 80%, precision of 84%, and Hausdorff distance of 6.30 mm.
• Demonstrated state-of-the-art performance and good generalization across scanners.

Conclusions:

• The proposed deep learning method is effective and robust for automated WMH detection.
• The system shows strong potential for real-world clinical applications.
• Public availability of the software and models facilitates further research and adoption.