Addressing Clinical Ambiguity in Breast Density Assessment: A Hybrid Multi-View Deep Learning Framework for BI-RADS B...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Mammographic breast density assessment is vital for breast cancer detection and risk stratification.
• High inter-observer variability in classifying intermediate densities (BI-RADS B vs. C) necessitates automated solutions.
• Developing robust and interpretable automated techniques is crucial for improving diagnostic consistency.

Purpose of the Study:

• To present a hybrid multi-view deep learning framework for classifying BI-RADS B versus C mammographic breast density.
• To enhance the accuracy and reduce variability in breast density assessment.
• To provide an interpretable AI model for clinical decision support.

Main Methods:

• A hybrid multi-view deep learning framework using EfficientNet-B4 and U-Net was developed.
• Feature-level fusion of craniocaudal (CC) and mediolateral oblique (MLO) mammographic views was employed.
• Interpretability was achieved using Grad-CAM, and evaluation was performed on the RSNA dataset with a patient-wise split.

Main Results:

• The model achieved an accuracy of 87% and an Area Under the Curve (AUC) of 94.40%.
• Performance was statistically significant compared to reference values, validated by McNemar's and DeLong's tests.
• Qualitative evaluation by radiologists confirmed the clinical relevance of Grad-CAM highlighted regions.

Conclusions:

• Combining multi-view deep learning with explainable AI improves breast density assessment consistency.
• The proposed framework shows potential to support clinical decision-making in mammography.
• Further prospective multicenter validation is required before clinical implementation.