Mammography datasets and artificial intelligence for breast cancer detection: challenges, bias, and future directions

Breast cancer is still one of the main causes of death in women from cancer worldwide, and early detection is crucial. Mammography is the primary imaging modality for breast cancer screening and recent advances in artificial intelligence (AI) have significantly enhanced automated breast cancer detection. This review takes a data-centered approach, examining key characteristics of public datasets, the contributions of mammography data to AI for breast cancer detection, and the implications for clinical generalizability. This review, however, takes a data-centered approach, critically analyzing characteristics of publicly available mammography datasets, and the role of mammography data to the development, performance and clinical generalization of AI models for the detection of breast cancer. The datasets are compared on the following aspects: dataset size, quality of the annotations, imaging modality, clinical metadata, dataset accessibility and demographic diversity. Analysis reveals that the quality, diversity and representativeness of the data used are critical for the performance and generalizability of AI models, ranging from the traditional machine learning era to the current era of deep learning, transformer-based systems, multimodal learning, explainable AI and privacy-preserving learning.

The results show that, apart from the development of AI model architectures, it is a necessity to have standardized, diverse and clinically representative datasets to make further progress in AI-assisted mammography. The review emphasizes the importance of establishing a comprehensive benchmark dataset for mammography that is large and diverse to enhance reproducibility, to fairly evaluate models, and to develop strong, clinically viable AI systems for breast cancer detection.