AAU-net

AAU-net segments breast lesions from ultrasound images using a hybrid adaptive attention U-Net architecture that integrates a Hybrid Adaptive Attention Module (HAAM) to improve segmentation accuracy and robustness.

Key Features:

• Hybrid Adaptive Attention Module (HAAM): Replaces traditional convolutions and combines channel self-attention and spatial self-attention to capture multi-scale features and handle irregular lesion shapes and blurred boundaries.
• Adaptive Feature Selection: Dynamically guides selection of optimal representations in both channel and spatial dimensions to manage similar intensity distributions and variable tumor morphologies in ultrasound images.
• Enhanced Generalization Performance: Demonstrates superior performance and robustness compared to several state-of-the-art deep learning segmentation methods across three public breast ultrasound datasets.
• Flexibility and Integration: HAAM is designed for integration into existing network frameworks to enhance segmentation modules without altering core architectures.

Scientific Applications:

• Breast lesion segmentation: Automatic and stable segmentation of breast lesions from ultrasound images.
• Tumor characterization support: Improves lesion delineation to aid radiologists in distinguishing benign from malignant tumors and supporting early detection and treatment planning.
• Clinical research and multi-dataset evaluation: Enables robust segmentation in studies that require generalization across variable image quality and lesion morphology.

Methodology:

AAU-net replaces conventional convolution operations with the HAAM, which integrates channel self-attention and spatial self-attention and performs adaptive feature selection in channel and spatial dimensions; performance was evaluated via experiments and robustness analysis on three public breast ultrasound datasets against state-of-the-art segmentation methods.