Re-thinking Mammography Transfer Learning: The Dataset-Informed Transfer Learning (DITL) Framework for Breast Cancer Screening and Lesion Diagnosis

TL;DR - DITL is a mammography transfer-learning framework that adapts training to dataset-derived sample difficulty and neighborhood structure. It reports significant improvements across large breast-density and small lesion-classification datasets.

• Weights cross-entropy per sample using nearest-neighbor label purity in self-supervised feature space.
• Uses triplet supervision with a learnable margin to improve class separation and compactness.
• Requires no focal-loss-style hyperparameter tuning and adds negligible computational overhead.
• Achieves state-of-the-art VinDR-Mammo breast-density classification and significant gains on small ROI datasets (p < 0.0001).