Saliency map-guided hierarchical dense feature aggregation framework for breast lesion classification using...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning, particularly convolutional neural networks, has improved breast lesion classification in breast ultrasound (BUS) images.
• Challenges remain due to complex patterns, low contrast, and fuzzy boundaries in BUS images, hindering accurate classification.
• Few methods separate foreground (lesion) and background for domain-specific feature learning and fusion.

Purpose of the Study:

• To propose a saliency map-guided hierarchical dense feature aggregation framework for enhanced breast lesion classification using BUS images.
• To improve the accuracy and robustness of automated breast lesion diagnosis.

Main Methods:

• Generated saliency maps for foreground and background using super-pixel clustering and multi-scale region grouping.
• Developed a triple-branch network with two feature extraction branches (inputting original image and saliency map) and one feature aggregation branch.
• Learned and fused foreground- and background-specific representations hierarchically under saliency map guidance.

Main Results:

• The proposed framework demonstrated superior performance compared to several state-of-the-art deep learning methods.
• Evaluated on three BUS datasets using 5-fold cross-validation, confirming its effectiveness.
• The saliency map guidance effectively improved the extraction and fusion of discriminative features.

Conclusions:

• The saliency map-guided hierarchical dense feature aggregation framework offers a promising approach for accurate breast lesion classification in BUS images.
• This method effectively addresses the challenges posed by complex patterns and low contrast in BUS imaging.
• The findings suggest a significant advancement in automated breast lesion diagnosis using deep learning.