Cross-Dataset Generalization Analysis of CNN and Vision Transformer Models for Mammography Classification

Deep learning models may achieve high performance on mammography datasets but may be less reliable when applied to images from other clinical domains. This paper investigates the cross-dataset generalization capability of a convolutional neural network and a Vision Transformer on mammography classification. ResNet50 and Swin Transformer were fine-tuned on the CBIS-DDSM data and tested on the independent INbreast dataset. The robustness of the model was evaluated based on a three-stage evaluation strategy, including internal validation, external validation and lightweight domain adaptation. Both models also experienced significant drop in performance at external validation with false negative rates of 37.5% and 48.75% for ResNet50 and Swin Transformer respectively. For lightweight adaptation, 20% of INbreast reduced the false negative rate of ResNet50 to 13.75% (a 63.3% relative reduction), with Swin Transformer showing a small increase to 50%. Results show that cross-dataset performance can vary significantly for different architectures, and limited target-domain adaptation can be more beneficial for some models than others. The study points to the importance of external validation and architecture-specific robustness analysis for the reliable deployment of AI in mammography.