Self-attention U-Net (SAU-Net): An attention-driven U-Net framework for precise brain tumor segmentation using...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Computational Neuroscience

Background:

• Brain tumor segmentation (BraTS) is crucial for diagnosis and treatment planning.
• Existing methods face challenges in accuracy and computational cost.
• Multimodal MRI data is essential for comprehensive tumor assessment.

Purpose of the Study:

• To develop a precise and efficient brain tumor segmentation technique.
• To address limitations in accuracy and computational expenses of current BraTS methods.
• To improve clinical diagnosis and treatment planning through enhanced tumor localization.

Main Methods:

• Proposed a novel Self-Attention U-Net (SAU-Net) model integrating self-attention with U-Net architecture.
• Enabled selective feature concentration and preserved spatial context for enhanced segmentation.
• Validated performance on BraTS 2018 and BraTS 2020 datasets using rigorous cross-validation.

Main Results:

• Achieved high Dice scores: 98.23% (BraTS 2018 avg.) and 98.62% (BraTS 2020 avg.).
• Demonstrated superior segmentation accuracy for whole tumor (WT), tumor core (TC), and enhancing tumor (ET).
• Reported reduced computational complexity, training/prediction times, and optimized memory usage.

Conclusions:

• SAU-Net offers a highly effective and computationally efficient solution for BraTS.
• Superior performance on benchmark datasets indicates significant potential for clinical applications.
• The model's efficiency and accuracy support practical integration into diagnostic workflows.