LRSCnet: Local Reference Semantic Code learning for breast tumor classification in ultrasound images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Accurate breast cancer detection is crucial for early diagnosis and treatment.
• Breast ultrasound imaging is a widely used modality, but classification can be challenging, especially with limited labeled data.

Purpose of the Study:

• To propose a novel Local Reference Semantic Code (LRSC) network for automatic breast ultrasound image classification.
• To address the challenge of classifying breast ultrasound images with few labeled data.

Main Methods:

• Development of a local structure extractor to identify common tumor characteristics.
• Implementation of a two-stage hierarchical encoder for generating high-level semantic codes from lesion structures.
• Utilization of a self-matching layer for final classification based on the learned semantic code.

Main Results:

• The LRSC network achieved superior performance compared to traditional methods.
• Key performance metrics included AUC (0.9540), Accuracy (0.9776), Sensitivity (0.9629), Specificity (0.93), PPV (0.9774), and NPV (0.9090).
• The proposed method also demonstrated an improvement in matching speed.

Conclusions:

• The LRSC network is effective for breast ultrasound image classification with limited labeled data.
• The two-stage hierarchical encoder learns high-level semantic codes that improve classification accuracy and generalization.
• The LRSC network offers a simpler and more effective approach for breast ultrasound analysis.