Multi-task Learning for Neonatal Brain Segmentation Using 3D Dense-Unet with Dense Attention Guided by Geodesic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Deep convolutional neural networks (CNNs) excel at neonatal brain MRI segmentation.
• CNNs often struggle with generalization to new datasets due to differing imaging characteristics.
• Poor generalization stems from models overfitting training data rather than learning robust features.

Purpose of the Study:

• To develop a deep learning method that improves generalization for neonatal brain MRI segmentation.
• To address the challenge of domain shift in medical imaging datasets.
• To enhance the reliability of automated segmentation across diverse neonatal brain MRI data.

Main Methods:

• A multi-task learning framework was proposed, integrating tissue segmentation and geodesic distance regression.
• A shared encoder network was regularized by simultaneously learning both tasks.
• A dense attention gate mechanism was incorporated to capture richer contextual information.

Main Results:

• The proposed method demonstrated superior performance on unseen neonatal brain MRI datasets.
• Experimental results validated the effectiveness across three datasets with varying imaging protocols and scanners.
• The multi-task approach significantly improved generalization compared to existing deep learning methods.

Conclusions:

• The developed multi-task learning method enhances the generalization capability of deep learning models for neonatal brain MRI segmentation.
• Integrating geodesic distance regression and attention mechanisms provides a robust solution for diverse imaging conditions.
• This approach offers a more reliable tool for analyzing neonatal brain development and pathology.