DV-DCNN: Dual-view deep convolutional neural network for matching detected masses in mammograms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Mammography uses mediolateral oblique (MLO) and craniocaudal (CC) views for breast cancer screening.
• Current deep learning models often process these views separately, limiting the analysis of mass relationships.
• Direct feature concatenation fails to capture nuanced variations in mass appearance between views.

Purpose of the Study:

• To develop a dual-view deep convolutional neural network (DV-DCNN) for matching masses across MLO and CC mammogram views.
• To improve the robustness and accuracy of breast cancer mass detection by establishing correspondence between detected masses.
• To enhance the diagnostic capabilities of radiologists by providing a more comprehensive analysis of mammographic findings.

Main Methods:

• The DV-DCNN model employs tied dense blocks for feature extraction from paired image patches.
• A neighborhood patch matching component analyzes relationships between features from different views.
• Key layers include cross-input neighborhood differences, patch summary features, and across-patch features for higher-level representation.

Main Results:

• The DV-DCNN was evaluated on the CBIS-DDSM and INbreast public datasets.
• Experimental results demonstrate superior performance compared to existing deep learning models.
• Integrating the DV-DCNN into a mass detection framework significantly improved detection outcomes.

Conclusions:

• Matching potential masses between different mammographic views enhances detection robustness.
• The proposed dual-view deep learning model effectively matches masses, increasing detection accuracy.
• This approach leads to a reduction in false positive rates for breast cancer screening.