Mammography Datasets for Neural Networks-Survey

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep neural networks (DNNs) are increasingly used in mammography analysis.
• Training DNNs requires substantial, high-quality mammography datasets with annotated abnormalities.
• Open-access databases are vital resources for developing AI in breast cancer screening.

Purpose of the Study:

• To conduct a comprehensive survey of open-access mammography databases suitable for training DNNs.
• To identify databases containing images with defined abnormal areas of interest.
• To review studies utilizing these databases and their reported outcomes with DNNs.

Main Methods:

• Systematic review of mammography image databases.
• Inclusion criteria focused on open-access availability and presence of annotated abnormal regions.
• Literature search for studies employing these databases with DNNs for mammography analysis.

Main Results:

• Identified key databases: INbreast, CBIS-DDSM, OMI-DB, and MIAS.
• These databases collectively offer at least 3801 unique images with 4125 findings from 1842 patients.
• Potential to access data from up to 14,474 patients through agreements with OPTIMAM.
• Detailed the annotation process for mammography images.

Conclusions:

• Established mammography databases provide essential data for training DNNs in breast cancer detection.
• The surveyed resources offer a significant volume of annotated images for AI development.
• Understanding image annotation is crucial for maximizing the utility of these datasets.