Att-BrainNet: Attention-based BrainNet for lung cancer segmentation network.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computational neuroscience

Background:

• Current medical image segmentation models often use a single feature strategy, failing to capture lesion heterogeneity.
• This limitation reduces accuracy and robustness when segmenting diverse lesions in complex medical images.

Purpose of the Study:

• To propose a novel brain-inspired segmentation framework, BrainNet, addressing limitations of current methods.
• To enhance lesion segmentation accuracy and generalization through differentiated feature modeling.

Main Methods:

• Developed a tri-level backbone encoder-Brain Network-decoder architecture (BrainNet).
• Instantiated BrainNet with an attention-enhanced model (Att-BrainNet) featuring a Thalamus Gating Module (TGM) and Encephalic Region Networks (ERNs).
• Incorporated an S-F image enhancement module and multi-head self-attention for improved feature extraction and global modeling.

Main Results:

• Att-BrainNet demonstrated superior accuracy and generalization compared to mainstream models on lung cancer CT and multi-organ datasets.
• Ablation studies and visualizations confirmed the effectiveness of the BrainNet architecture and its dynamic scheduling strategy.
• The model successfully handles morphologically diverse lesions in complex imaging contexts.

Conclusions:

• BrainNet offers a novel structural paradigm for medical image segmentation, inspired by brain function.
• Att-BrainNet significantly improves segmentation performance by dynamically modeling lesion heterogeneity.
• The framework shows potential for broader applications in complex medical image segmentation tasks.