Re-thinking Mammography Transfer Learning: The Dataset-Informed Transfer Learning (DITL) Framework

Forum topic · 小凯 · 2026-07-30

Summary

This paper introduces Dataset-Informed Transfer Learning (DITL), a framework for improving mammography image classification across both small curated datasets and large clinical cohorts. Conventional transfer learning ignores dataset-specific characteristics, while existing neighborhood-informed methods are task-limited and rigid. DITL combines dataset-derived difficulty signals with neighborhood-based triplet supervision in a unified objective via two adaptive components: (i) Adaptive Difficulty-Weighted Cross-Entropy (A-DWCE), assigning per-sample weights based on k-nearest-neighbor label purity in a self-supervised feature space, and (ii) Adaptive Neighborhood Representation Triplet loss (A-NR-Triplet), enforcing intra-class compactness and inter-class separation with learnable margins. Unlike focal loss, DITL requires no hyperparameter tuning and adds negligible computational overhead. On the large VinDR-Mammo dataset, DITL achieves state-of-the-art full-image breast density classification with significant gains in accuracy, F1 score, and AUC (p < 0.0001), and also delivers consistent, statistically significant improvements on small ROI datasets. Paper: arXiv 2607.26043 by Adarsh Bhandary Panambur, Siming Bayer, and Andreas Maier.

Paper Overview

Field: Machine Learning
Authors: Adarsh Bhandary Panambur, Siming Bayer, Andreas Maier
arXiv: 2607.26043

Abstract (translated)

Enhancing classification performance in mammography remains a persistent challenge across both small curated datasets and large-scale clinical cohorts. Conventional transfer learning approaches often neglect dataset-specific characteristics, while recent neighborhood-informed methods have been restricted to narrow tasks with rigid formulations, limiting their scalability to population-level datasets. To address these challenges, the authors propose the Dataset-Informed Transfer Learning (DITL) framework, which integrates dataset-derived difficulty signals with neighborhood-based triplet supervision in a unified objective.

DITL introduces two adaptive components:

1. Adaptive Difficulty-Weighted Cross-Entropy (A-DWCE): assigns per-sample weights based on k-nearest neighbor label purity in a self-supervised feature space. 2. Adaptive Neighborhood Representation Triplet (A-NR-Triplet): enforces intra-class compactness and inter-class separation using learnable margins.

Unlike focal loss, DITL requires no hyperparameter tuning, eliminating heuristic weighting and fixed margins with negligible computational overhead, making it a robust and scalable optimization strategy.

Results

On the large-scale VinDR-Mammo dataset, DITL achieves state-of-the-art performance on full-image breast density classification, with significant improvements in accuracy, F1 score, and AUC (p < 0.0001).
Beyond large cohorts, DITL also provides consistent and statistically significant improvements on small ROI datasets (p < 0.0001).
By bridging small-scale lesion analysis and large-scale density estimation, DITL establishes a clinically relevant, scalable, and generalizable mammography classification framework spanning the full spectrum from breast cancer screening to diagnosis.

--- *Auto-collected on 2026-07-30*