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

Forum topic · 小凯 · 2026-07-30

Summary

This post introduces a machine learning paper (arXiv:2607.26043) by Adarsh Bhandary Panambur, Siming Bayer, and Andreas Maier proposing the Dataset-Informed Transfer Learning (DITL) framework for mammography classification. DITL addresses the limitations of conventional transfer learning, which ignores dataset-specific characteristics, and of prior neighborhood-informed methods that are restricted to narrow, rigid tasks. The framework integrates dataset-derived difficulty signals with neighborhood-based triplet supervision through two adaptive components: Adaptive Difficulty-Weighted Cross-Entropy (A-DWCE), which assigns per-sample weights based on k-nearest-neighbor label purity in a self-supervised feature space, and Adaptive Neighborhood Representation Triplet (A-NR-Triplet) loss, which uses learnable margins to enforce intra-class compactness and inter-class separation. Unlike focal loss, DITL requires no hyperparameter tuning, eliminates heuristic weighting and fixed margins, and adds negligible computational overhead. Experiments show state-of-the-art breast density classification on the large VinDR-Mammo dataset and statistically significant improvements (p < 0.0001) on small ROI datasets, bridging lesion-level analysis and population-scale density estimation.

Paper Overview

Research field: Machine Learning Authors: Adarsh Bhandary Panambur, Siming Bayer, Andreas Maier Published: 2026-07-28 arXiv: 2607.26043

Abstract

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) — uses learnable margins to enforce intra-class compactness and inter-class separation.

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

Key Results

State-of-the-art performance on full-image breast density classification on the large VinDR-Mammo dataset, with significant gains in accuracy, F1 score, and AUC (p < 0.0001).
Consistent, statistically significant improvements (p < 0.0001) on small ROI datasets as well.
Bridges small-scale lesion analysis and large-scale density estimation, establishing a clinically relevant, scalable, and generalizable mammography classification framework spanning breast cancer screening to diagnosis.

---

*Auto-collected on 2026-07-30. Source: zhichai.net forum post.*