Learning to segment fetal brain tissue from noisy annotations.

Area of Science:

• Medical imaging
• Neuroscience
• Artificial intelligence

Background:

• Accurate fetal brain segmentation is crucial for assessing neurodevelopment.
• Deep learning excels at medical image segmentation but requires extensive training data.
• Manual segmentation of numerous 3D fetal brain MRI scans is impractical.

Purpose of the Study:

• To develop an automated method for segmenting fetal brain tissues in MRI.
• To train a deep learning model using a large dataset of noisy segmentations.
• To improve the accuracy and reproducibility of fetal brain analysis.

Main Methods:

• Generated a large training dataset using automatic multi-atlas segmentation with manual error correction.
• Developed a novel label smoothing technique and loss function for training with noisy segmentations.
• Evaluated the deep learning model on 23 manually-segmented test images.

Main Results:

• Achieved average Dice similarity coefficients of 0.893 (younger fetuses) and 0.916 (older fetuses).
• Outperformed several state-of-the-art segmentation methods, including nnU-Net.
• Demonstrated accurate segmentation of transient fetal brain structures across gestational weeks.

Conclusions:

• The proposed deep learning method effectively segments fetal brain MRI, accounting for segmentation uncertainty.
• This approach provides a valuable tool for enhancing the accuracy and reproducibility of fetal brain development analysis.
• The method addresses the challenge of limited manually segmented data for deep learning model training.