Contrastive learning: an efficient Domain Adaptation strategy for 2D mammography image classification

Effective computer aided breast-cancer diagnosis models using 2D mammography images must maintain consistent performance across varying image acquisition systems and post-processing techniques. Nevertheless, Deep Learning (DL) models have shown diminished performance with variations in image style and contrast. We propose two models trained for classifying respectively 2D mammography patches and complete images, using heterogeneous datasets distinguished by different image post-processing methods. We propose a Domain Adaptation (DA) methodology using Supervised Contrastive Learning (SCL) to achieve domain-invariant representations and improved class-separability. This approach is compared to a standard training using the Cross Entropy (CE) loss. The domain invariant models outperform those trained with CE in binary classification of full mammograms (cancer vs. no cancer), increasing the AUC from 0.745 to 0.816 in an independent test set. For patch classification, we show that the Domain Adaptation effectiveness varies with weight initialization and dataset size.