Public data homogenization for AI model development in breast cancer

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Oncology Imaging

Background:

• Developing trustworthy AI for clinical use necessitates robust data. Publicly available datasets, like those from The Cancer Imaging Archive (TCIA), offer potential but require harmonization due to heterogeneity.
• Breast cancer research benefits from accessible imaging and clinical data, yet existing archives are difficult to filter for specific AI model development needs.

Purpose of the Study:

• To develop a homogenized dataset for breast cancer, integrating clinical and imaging data for AI model development.
• To create a searchable platform for accessing and filtering breast cancer imaging and clinical data.

Main Methods:

• Acquired and harmonized five datasets from TCIA.
• Developed a common data model and an extract-transform-load process for clinical data homogenization.
• Extracted and made searchable Digital Imaging and COmmunications in Medicine (DICOM) information from magnetic resonance imaging (MRI) data.

Main Results:

• Created a homogenized dataset comprising data from 2,035 breast cancer subjects.
• Developed the RV-Cherry-Picker platform for unified searching and downloading of clinical and imaging datasets.
• Enabled selection of specific imaging series (e.g., dynamic contrast-enhanced) for AI model development.

Conclusions:

• RV-Cherry-Picker provides access to the largest, publicly available, homogenized imaging and clinical dataset for breast cancer AI development.
• The presented methodology facilitates the creation of merged public datasets for AI model development, exemplified by breast cancer MRI data.
• The platform offers unified access and detailed selection of breast MRI data, significantly aiding AI model development.