An Automated Deep Learning Model for the Cerebellum Segmentation from Fetal Brain Images.

Area of Science:

• Medical imaging
• Artificial intelligence
• Neuroscience

Background:

• Cerebellum measurements from ultrasound (US) images are crucial for assessing gestational age and identifying central nervous system abnormalities.
• Standardized, large-scale cerebellar assessments are needed to correlate fetal development with postnatal neurodevelopmental outcomes.
• Current methods for cerebellar segmentation are time-consuming and lack precision, necessitating advanced automated solutions.

Purpose of the Study:

• To introduce an innovative deep learning (DL) technique for automatic fetal cerebellum segmentation from 2D US brain images.
• To develop a semantic segmentation network, ReU-Net, specifically designed for fetal cerebellum anatomy.
• To enhance segmentation accuracy by addressing noise in US data.

Main Methods:

• Developed ReU-Net, a U-Net based semantic segmentation network incorporating residual blocks and a Wiener filter.
• Trained the model on 590 fetal US images and tested on 150 images, employing a 5-fold cross-validation strategy.
• Evaluated performance using Dice Score Coefficient (DSC), F1-score, Hausdorff Distance (HD), accuracy, recall, and precision.

Main Results:

• ReU-Net achieved high performance metrics: 91% DSC, 92% F1-score, 98% accuracy, 92% recall, and 94% precision.
• The method demonstrated a Hausdorff Distance of 25.42.
• ReU-Net significantly outperformed other U-Net based methods (p < 0.001).

Conclusions:

• The proposed ReU-Net offers an accurate and efficient automated solution for fetal cerebellum segmentation from 2D US images.
• This approach can facilitate high-throughput medical studies and broader biometric evaluations using fetal US data.
• The method holds potential for improved prediction of neurodevelopmental and growth consequences through early identification of structural abnormalities.