Innovative Multi-View Strategies for AI-Assisted Breast Cancer Detection in Mammography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Mammography is crucial for early breast cancer detection but faces challenges with inter-reader variability and subtle feature interpretation.
• Deep convolutional neural networks (CNNs) offer potential for automated mammogram classification, aiming to improve diagnostic accuracy.

Purpose of the Study:

• To evaluate deep convolutional neural networks (CNNs) for automated mammogram classification.
• To introduce and assess two novel multi-view integration techniques: Dual-Branch Ensemble (DBE) and Merged Dual-View (MDV).
• To test model generalizability across diverse mammography datasets and imaging systems.

Main Methods:

• Comparative analysis of state-of-the-art CNN architectures (ResNet, DenseNet, EfficientNet, MobileNet, Vision Transformers, VGG19) on six mammography datasets.
• Implementation and evaluation of Dual-Branch Ensemble (DBE) and Merged Dual-View (MDV) multi-view integration strategies.
• Out-of-sample testing using dedicated datasets to assess model generalizability and robustness against domain shift.

Main Results:

• Both MDV and DBE strategies significantly improved classification performance.
• Under the MDV approach, VGG19 and DenseNet achieved high ROC AUC scores of 0.9051 and 0.7960, respectively.
• In the DBE setting, DenseNet and ResNet50 achieved ROC AUC scores of 0.8033 and 0.8042, demonstrating the benefits of multi-view fusion.

Conclusions:

• Multi-view fusion techniques (MDV and DBE) enhance the robustness and accuracy of CNN-based mammogram classification.
• Generalization tests highlight the critical need for diverse training datasets to mitigate the impact of domain shift.
• The findings provide practical guidance for developing reliable and widely applicable AI-assisted breast cancer screening tools.