Single-Input Multi-Output U-Net for Automated 2D Foetal Brain Segmentation of MR Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate fetal brain segmentation is crucial for diagnosing developmental abnormalities.
• Existing segmentation methods often face challenges with complex anatomical structures and image variability.
• Deep learning models like U-Net show promise but can be computationally intensive or prone to generalization errors.

Purpose of the Study:

• To develop an efficient and accurate deep learning model for fetal brain segmentation.
• To combine the strengths of U-Net and holistically nested edge detection (HED) for improved segmentation performance.
• To reduce computational cost and prediction variance compared to traditional ensemble methods.

Main Methods:

• Development of the Single-Input Multi-Output U-Net (SIMOU-Net), a hybrid deep learning architecture.
• Integration of features from multiple output levels within a single network to mimic ensemble learning.
• Training and validation on a dataset of 200 normal fetal brains (over 11,500 images) and 54 abnormal cases (over 3500 images).

Main Results:

• Achieved high segmentation accuracy on normal fetal brains with Dice coefficients of 94.2 ± 5.9% and Jaccard coefficients of 88.7 ± 6.9%.
• Demonstrated robust performance on abnormal fetal brain cases, yielding Dice coefficients of 91.2 ± 6.8% and Jaccard coefficients of 85.7 ± 6.6%.
• The SIMOU-Net effectively reduces prediction variance and generalization errors compared to standard approaches.

Conclusions:

• The proposed SIMOU-Net offers a computationally efficient and highly accurate solution for fetal brain segmentation.
• This hybrid network architecture demonstrates significant potential for clinical applications in prenatal diagnosis.
• SIMOU-Net provides a valuable tool for improving the analysis of fetal brain development and identifying abnormalities.