The nnU-Net based method for automatic segmenting fetal brain tissues.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Fetal brain magnetic resonance (MR) imaging enables early detection of brain pathologies.
• Accurate brain tissue segmentation is crucial for fetal brain morphology and volume analysis.

Purpose of the Study:

• To adapt and apply the nnU-Net deep learning model for precise segmentation of seven fetal brain tissue types.
• To evaluate the performance of the adapted nnU-Net on the FeTA 2021 dataset.

Main Methods:

• The nnU-Net framework was modified to handle the specific characteristics of fetal brain MR images.
• The adapted nnU-Net was trained and evaluated on the FeTA 2021 dataset for segmenting external cerebrospinal fluid, gray matter, white matter, ventricle, cerebellum, deep gray matter, and brainstem.

Main Results:

• The adapted nnU-Net outperformed other leading methods on the FeTA 2021 training data, achieving Dice, HD95, and VS scores of 0.842, 11.759, and 0.957, respectively.
• On the FeTA 2021 test data, the model achieved strong segmentation performance (Dice: 0.774, HD95: 14.699, VS: 0.875), ranking third in the challenge.
• The method demonstrated robust performance across fetal MR images of varying gestational ages.

Conclusions:

• The adapted nnU-Net provides a highly effective tool for fetal brain tissue segmentation from MR images.
• This technology can significantly assist clinicians in making accurate and timely diagnoses of fetal brain conditions.