Cross-Sensor and Cross-Population Generalization of Deep Learning Models for Digital Mammography: A Controlled...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Mammography
• Computer-Aided Diagnosis

Background:

• Digital mammography AI models are increasingly used across diverse hospital settings.
• Variability in X-ray detector technology and patient demographics poses challenges for model generalizability.
• The performance of advanced mammography foundation models across different platforms remains under-tested.

Purpose of the Study:

• To evaluate the transferability and performance of state-of-the-art deep learning models for digital mammography.
• To assess whether models trained on one sensor and population maintain accuracy on others.
• To determine the necessity of local validation and recalibration for clinical deployment.

Main Methods:

• Fine-tuning five deep learning architectures (ResNet-50, DINOv2-B14, Rad-DINO, Mammo-CLIP B5, Mammo-FM) on the CBIS-DDSM dataset.
• Ablating a density-aware focal loss function to assess its impact on model performance.
• Evaluating model transfer to three external sensor cohorts (CMMD, DMID, MIAS) with statistical significance testing.

Main Results:

• ResNet-50 outperformed foundation models on the training dataset (CBIS-DDSM).
• Transferring models to external datasets resulted in significant AUC degradation (0.165-0.320 points).
• Sensitivity at 95% specificity decreased substantially upon transfer, highlighting performance collapse.

Conclusions:

• Mammography-specific pretraining did not guarantee superior in-distribution performance or robust transferability.
• No evaluated architecture achieved sufficient performance for defensible deployment without local validation.
• Sensor-stratified, population-stratified external validation and local recalibration are essential prerequisites for clinical use.