Skip to main content
Back

DoSReMC Fixes Breast Cancer AI That Fails at New Hospitals Without Retraining

A new fine-tuning recipe touches only batch norm and classifier layers to keep breast cancer models accurate when hospitals swap scanners and imaging protocols.

Pro

Paper, Full Access
  • DoSReMC adapts breast cancer CNNs to new hospitals by fine-tuning only batch norm and classifier layers.
  • Batch normalization layers identified as the main source of domain shift in mammography CNNs.
  • Adds gradient-reversal domain adversarial training to further align features across scanners.
  • Evaluated on HCTP (new Hacettepe dataset), VinDr-Mammo, and CSAW-CC full-field digital mammography sets.
  • Adversarial variant hits 0.83 ROC-AUC and 0.85 PR-AUC on unseen CSAW data.
  • Selective fine-tuning matches or beats full retraining while drastically cutting compute cost.

Mammography classifiers tend to look brilliant on the data they were trained on and mediocre everywhere else. A new framework called DoSReMC pinpoints why that happens inside convolutional networks, and proposes a lightweight fix that skips full retraining while restoring most of the lost accuracy on unseen hospitals' data.

The generalization tax on medical CNNs

Deep learning models for breast cancer screening have a well-known reliability problem. Numerous systems have been built for automatic recognition of malignancies in mammograms, but their performance often declines when applied to data from a different source, largely because of domain shift: the variation in data distributions between where a model was trained and where it is deployed. Different scanner vendors, exposure settings, and patient populations produce subtly different pixel statistics, and a model tuned to one hospital's look often flounders at the next.

The authors, working with radiologists at Hacettepe University in Türkiye, zero in on a specific culprit: the batch normalization layers that sit between convolutions in essentially every modern CNN. Their experiments show that BN layers are a primary source of domain dependence, performing well when training and testing occur within the same source but significantly impairing generalization under shift. Since BN stores running means and variances from the training set, those statistics silently encode scanner-specific biases into the model.