A deep learning framework for accurate mammographic mass classification using local context attention module

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Dense breast tissue is a significant risk factor for breast cancer (BC).
• Current mammographic classification of BC is often subjective and unreliable, hindering accurate evaluation.
• Improving BC classification accuracy is critical for patient outcomes.

Purpose of the Study:

• To develop a deep learning method with a local context attention module (LCAM) for enhanced BC classification.
• To improve grading consistency and accuracy using dual mammogram views aligned with BI-RADS categories.
• To leverage local context around breast masses for more precise BC evaluation.

Main Methods:

• Identified regions of interest (ROIs) with dense tissue around breast masses from dual mammogram views.
• Utilized a convolutional neural network (CNN)-based model incorporating LCAM for feature selection and differentiation.
• LCAM inferred attention maps along channel and spatial dimensions for adaptive feature refinement.

Main Results:

• The framework was evaluated on 3020 patients across four BI-RADS categories.
• Achieved a sensitivity of 82.46% and a specificity of 91.42% in identifying BI-RADS grading for breast masses.
• Demonstrated robust performance in classifying BC using dual mammogram views.

Conclusions:

• Introduced a novel CNN-based framework utilizing dual mammogram views for BC classification.
• The LCAM effectively captures local characteristics surrounding breast masses, enhancing classification accuracy.
• The proposed method aims to improve the consistency and reliability of BC classification outcomes.