Dilated Saliency U-Net for White Matter Hyperintensities Segmentation Using Irregularity Age Map.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• White matter hyperintensities (WMH) are key biomarkers for brain aging and dementia.
• Automatic WMH segmentation is difficult due to variations in intensity, size, and shape.
• Existing U-Net models for WMH segmentation are complex.

Purpose of the Study:

• To develop a less complex U-Net architecture for WMH segmentation.
• To improve the accuracy and efficiency of WMH detection in MRI scans.
• To reduce computational resources required for WMH segmentation.

Main Methods:

• Proposed Saliency U-Net incorporating an Irregularity map (IAM) with T2-FLAIR MRI sequences.
• Applied dilated convolution to enhance Saliency U-Net for multi-context learning.
• Trained and evaluated models based on Dice coefficient and sensitivity metrics.

Main Results:

• Saliency U-Net with IAM outperformed standard U-Net.
• Dilated Saliency U-Net achieved the best Dice score (0.5588) and good sensitivity (0.4747).
• The proposed method reduced model complexity, training time, and parameter count.

Conclusions:

• Incorporating IAM with Dilated Saliency U-Net is an effective approach for WMH segmentation.
• The developed model offers improved accuracy with greater efficiency.
• This method shows promise for advancing dementia and brain aging research.