MR-CT multi-atlas registration guided by fully automated brain structure segmentation with CNNs.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Radiology

Background:

• Computed tomography (CT) is crucial for identifying brain anomalies, but its soft tissue contrast limits detailed anatomical segmentation.
• Magnetic resonance imaging (MRI) is typically used for precise brain region division, posing challenges for CT-based diagnostics.

Purpose of the Study:

• To develop a fully automated method for registering MRI brain atlas data to CT scans.
• To enable detailed, automated anatomical reporting of brain CT examinations.

Main Methods:

• A multi-atlas registration approach was employed, propagating anatomical information from an MRI atlas to CT scans.
• Lateral ventricles and brain volume masks from CT images were used to guide the registration.
• Convolutional neural networks (CNNs) were validated for automated segmentation to enhance registration accuracy.

Main Results:

• The registration method achieved high accuracy, with mean Dice similarity coefficients of 0.92 (ventricles) and 0.99 (parenchyma) using manual segmentation guidance.
• Using CNN-based automated segmentation for guidance resulted in mean Dice values of 0.87 (ventricles) and 0.98 (parenchyma).

Conclusions:

• The proposed approach offers a fully automated solution for registering MRI atlases to CT scans, providing detailed anatomical insights.
• CNN-based segmentation effectively generates masks for brain ventricles and volume, significantly aiding the registration process.