U-net combined with CRF and anatomical based spatial features to segment white matter hyperintensities

Area of Science:

• Neuroimaging
• Medical image analysis
• Computational neuroscience

Background:

• White matter hyperintensities (WMH) are crucial biomarkers for cerebral small vessel disease.
• WMH are associated with various neurodegenerative processes.
• Accurate segmentation of WMH is essential for clinical research and diagnosis.

Purpose of the Study:

• To propose a fully automatic and accurate method for segmenting white matter hyperintensities (WMH).
• To evaluate the effectiveness of a U-net architecture combined with Conditional Random Fields (CRF) for WMH segmentation.
• To investigate the impact of anatomical spatial features on segmentation performance.

Main Methods:

• A U-net deep learning architecture was employed for automated WMH segmentation.
• Conditional Random Fields (CRF) were integrated to refine segmentation outputs.
• The neural network was trained using T1 and T2-FLAIR image intensities, incorporating novel anatomical spatial features derived from T1-based brain tissue segmentation.

Main Results:

• The proposed method demonstrated superior performance compared to 8 other automated segmentation techniques, including traditional and deep learning approaches.
• The integration of anatomical spatial features significantly enhanced segmentation accuracy and overall performance.
• The U-net with CRF and anatomical features achieved the best results across most evaluation metrics.

Conclusions:

• The developed fully automatic WMH segmentation method shows high efficacy and accuracy.
• Anatomical spatial features are vital for improving the performance of deep learning-based WMH segmentation.
• This approach offers a promising tool for the assessment of cerebral small vessel disease and related neurodegeneration.