Cerebrovascular segmentation from TOF-MRA based on multiple-U-net with focal loss function.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neuroscience

Background:

• Accurate segmentation of cerebrovascular structures is crucial for diagnosing cerebrovascular diseases.
• Manual segmentation by clinicians is complex and prone to uncertainty.
• Automated methods are needed to improve efficiency and accuracy.

Purpose of the Study:

• To develop an automated algorithm for segmenting cerebral vessels from Time-of-flight Magnetic Resonance Angiography (TOF-MRA) data.
• To enhance the accuracy and reliability of cerebrovascular segmentation compared to manual methods.
• To introduce a novel Multiple-U-net (M-U-net) approach for improved segmentation performance.

Main Methods:

• Normalization and multi-directional slicing (axial, coronal, sagittal) of TOF-MRA data.
• Training three individual U-net models with a focal loss function to address sample imbalance.
• Employing voting feature fusion and connected domain analysis for post-processing.

Main Results:

• The proposed M-U-net algorithm achieved a Dice Similarity Coefficient (DSC) of 88.60% on the verification dataset.
• The M-U-net model demonstrated 87.93% DSC on the testing dataset.
• Performance surpassed that of individual U-net models, indicating superior segmentation capabilities.

Conclusions:

• The M-U-net algorithm represents a state-of-the-art approach for TOF-MRA cerebrovascular segmentation.
• Feature fusion from multiple U-net models effectively complements segmentation results.
• This automated method offers a promising tool for the diagnosis of cerebrovascular diseases.