Benign Share Benefit to Malignant: Balanced Mixing on Feature Space for Imbalanced Breast Cancer Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Imbalanced mammography datasets can bias deep learning models, leading to under-recognition of less frequent malignant or high-risk categories.
• Accurate classification of Breast Imaging Reporting and Data System (BI-RADS) categories is crucial for effective breast cancer diagnosis.

Purpose of the Study:

• To propose B2M (Benign Share Benefit to Malignant), a model-agnostic framework to address imbalance-aware multi-class BI-RADS classification in C-View mammography.
• To improve the recognition of less frequent malignant and high-risk BI-RADS categories within imbalanced datasets.

Main Methods:

• B2M employs a two-phase training strategy combining dual sampling (instance-based and class-balanced) with feature-space mixing.
• Phase I trains the model with dual sampling to balance class representation while preserving diversity.
• Phase II fine-tunes the model using feature-space mixing and a soft-target regularization objective for smoother decision boundaries.

Main Results:

• B2M was evaluated on an imbalanced C-View mammography cohort (EMBED dataset) using stratified 5-fold cross-validation.
• The ResNeXt-50 model integrated with B2M achieved superior balanced accuracy and Macro-F1 scores compared to other oversampling and mixing methods.
• While B2M introduces an approximate 2.81× offline training overhead, it does not increase inference costs.

Conclusions:

• B2M demonstrates potential as an effective framework for imbalanced multi-class BI-RADS classification in C-View mammography.
• Further external and prospective validation is necessary before widespread clinical adoption.