Fetal-BET: Brain Extraction Tool for Fetal MRI.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Fetal MRI is vital for prenatal neurodevelopmental studies.
• Accurate fetal brain extraction is crucial for computational analysis but challenging due to fetal position, movement, and anatomical variability.
• Existing methods lack accuracy across diverse fetal MRI sequences.

Purpose of the Study:

• To develop an accurate and generalizable method for automatic fetal brain extraction from MRI sequences.
• To overcome limitations of existing techniques in handling fetal movement, varied anatomy, and diverse imaging conditions.
• To create a robust deep learning solution for precise fetal brain delineation.

Main Methods:

• A large dataset of approximately 72,000 2D fetal brain MRI images was curated, covering T2-weighted, diffusion-weighted, and functional MRI sequences.
• Deep learning models, utilizing U-Net architectures, attention mechanisms, multi-modal feature learning, and data augmentation, were developed and validated.
• The methods were trained on both normal and pathological brain data acquired with different scanners.

Main Results:

• The developed deep learning method achieved accurate fetal brain extraction on heterogeneous test data from various scanners and centers.
• The approach demonstrated robustness across different gestational stages and on pathological fetal brains.
• The method's generalizability was confirmed on independent datasets, showing high performance in diverse conditions.

Conclusions:

• The proposed deep learning solution enables precise fetal brain delineation across various fetal MRI sequences by leveraging multi-modality data.
• The model's robustness and accuracy offer significant potential for advancing fetal brain imaging analysis.
• This work provides a valuable tool for prenatal neuroimaging research and clinical applications.