Convolutional Neural Network for Automated FLAIR Lesion Segmentation on Clinical Brain MR Imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neurology

Background:

• Brain lesions on FLAIR MRI are common.
• Accurate segmentation and volumetric quantification are crucial for disease assessment.
• Current automated methods have limitations.

Purpose of the Study:

• Develop an automated deep learning method for FLAIR lesion segmentation.
• Quantify lesion volumes accurately on clinical brain MRIs.
• Evaluate performance against manual segmentation and existing automated techniques.

Main Methods:

• Adapted a U-Net convolutional neural network architecture for 3D brain MRIs.
• Trained the network on 295 brain MRIs for automated FLAIR lesion segmentation.
• Validated performance using Dice scores, sensitivity, specificity, and volumetric measurements on 92 cases.

Main Results:

• Achieved high FLAIR lesion segmentation performance (median Dice score, 0.79) across diverse lesion types.
• Outperformed existing automated methods (Dice, 0.56 and 0.41) and neared human performance (Dice, 0.81).
• Demonstrated strong correlation in lesion volume prediction (ρ = 0.99) and accuracy across various scanners and acquisition parameters.

Conclusions:

• A 3D U-Net-based convolutional neural network enables high-performance automated FLAIR segmentation in brain MR imaging.
• The method is robust across various pathologies and imaging parameters.
• Provides accurate volumetric data for disease burden assessment and radiologic reporting.