Split-Attention U-Net: A Fully Convolutional Network for Robust Multi-Label Segmentation from Brain MRI.

Area of Science:

• Neuroimaging and Computational Neuroscience
• Artificial Intelligence in Medical Imaging

Background:

• Accurate multi-label brain segmentation from MRI is crucial for neurological analysis.
• Existing segmentation algorithms can be complex, leading to delays in neuroimaging findings.

Purpose of the Study:

• To introduce Split-Attention U-Net (SAU-Net), a novel deep learning architecture for efficient brain MRI segmentation.
• To improve the speed and accuracy of neuroimaging biomarker extraction.

Main Methods:

• Developed SAU-Net, a convolutional neural network incorporating split-attention blocks, pyramid-level skip pathways, and evolving normalization layers.
• Employed a pre-training and fine-tuning strategy using original and modified FreeSurfer labels for efficient training on heterogeneous data.
• Evaluated SAU-Net on nine diverse neuroimaging datasets.

Main Results:

• SAU-Net demonstrated superior segmentation accuracy and reliability compared to state-of-the-art methods across nine evaluation datasets.
• The proposed learning strategy effectively utilized heterogeneous neuroimaging data with minimal manual annotations.
• SAU-Net exhibited a significantly faster processing runtime.

Conclusions:

• SAU-Net offers a robust and efficient solution for multi-label brain MRI segmentation.
• Its robustness to neuroanatomical variability and swift processing enable rapid access to accurate neuroimaging biomarkers.
• SAU-Net holds significant potential for clinical applications requiring timely neuroimaging analysis.