OMAMA-DB: the Oregon-Massachusetts Mammography Database.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Public datasets for training AI in breast cancer screening are often limited in size and quality.
• Developing reliable AI systems for mammography requires extensive, high-quality data.

Purpose of the Study:

• Introduce OMAMA-DB, a comprehensive public dataset of 2D mammograms and 3D tomosynthesis volumes.
• Provide a resource to overcome limitations in existing datasets for AI model development in breast cancer screening.

Main Methods:

• Curated 231,080 images from 967,991 initial images using multi-stage filtering.
• Applied histogram filtering and variational autoencoder for outlier detection in 2D images.
• Generated pathology-based cancer labels and automated lesion annotations, with expert validation via a web tool.

Main Results:

• OMAMA-DB contains 231,080 images, including 7351 2D and 374 3D cancer cases.
• Fine-tuned MedGemma achieved high performance (accuracy 0.989, sensitivity 0.997, F1 0.989) on a balanced subset.
• Automated methods significantly outperformed human accuracy in classifying real vs. synthetic mammograms.

Conclusions:

• OMAMA-DB offers a valuable resource for medical imaging research in mammography.
• Fine-tuned foundation models show strong performance, emphasizing the need for real clinical data.
• Open availability of data, models, and parameters supports further research and development.