TW-Net: Transformer Weighted Network for Neonatal Brain MRI Segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Neonatal brain MRI segmentation is crucial for understanding brain development and neurodevelopmental disorders.
• Intensity-based segmentation methods struggle with subtle contrast differences in neonatal brains due to myelination.
• Existing convolutional neural networks lack essential in-plane long-range dependency for accurate segmentation.

Purpose of the Study:

• To develop a novel deep learning network, TW-Net, for accurate neonatal brain MRI segmentation.
• To address the limitation of lacking in-plane long-range dependency in current segmentation models.
• To improve the analysis of infant brain growth and neurodevelopmental conditions.

Main Methods:

• Proposed TW-Net, a Transformer-Weighted network with an encoder-decoder architecture.
• Integrated a Transformer module with a rotate-and-flip layer to capture in-plane long-range dependencies.
• Incorporated deep supervision and squeeze-and-excitation blocks to enhance boundary information.

Main Results:

• TW-Net demonstrated superior performance compared to state-of-the-art deep learning algorithms.
• Achieved high accuracy in multiple-label segmentation tasks.
• Validated performance on two independent public datasets in 2D and 2.5D configurations.

Conclusions:

• TW-Net is a promising method for accurate neonatal brain MRI segmentation.
• The incorporation of in-plane long-range dependency significantly improves segmentation accuracy.
• TW-Net facilitates better investigation of infant brain development and disorders.