Learning global dependencies based on hierarchical full connection for brain tumor segmentation.

Area of Science:

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

Background:

• Brain tumor segmentation (BTS) is challenging due to high variability in tumor appearance, shape, and location.
• Existing attention mechanisms, like spatial attention and self-attention, have limitations in capturing global dependencies effectively for BTS.
• Convolution-based spatial attention struggles with global dependencies, while self-attention is memory-intensive for high-resolution feature maps.

Purpose of the Study:

• To introduce a novel Hierarchical Fully Connected (H-FC) module for learning global dependencies in BTS.
• To address the limitations of existing attention mechanisms in handling global dependencies and GPU memory constraints.
• To improve the performance of attention mechanisms in brain tumor segmentation by incorporating global dependency learning.

Main Methods:

• Proposed a Hierarchical Fully Connected (H-FC) module designed to learn global dependencies.
• H-FC hierarchically learns local dependencies at various feature map scales using fully connected layers.
• The module approximates global dependencies by combining these hierarchical local dependencies, requiring minimal GPU memory.

Main Results:

• The H-FC module demonstrated superior performance compared to Attention Gate and SAM (in CBAM) in BTS tasks.
• Significant improvements were observed across most performance metrics, particularly in Hausdorff Distance.
• Experiments confirmed H-FC is lightweight, with low computational and parameter overhead.

Conclusions:

• The novel H-FC module effectively learns global dependencies for brain tumor segmentation.
• Experiments on the BraTS2020 dataset validated the effectiveness and lightweight nature of H-FC.
• The study confirmed the importance of global dependencies in low-level feature maps for enhancing BTS performance.