A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Artificial intelligence (AI) research in breast cancer Magnetic Resonance Imaging (MRI) is hindered by a scarcity of expert-labeled segmentation data.
• Existing public datasets often lack sufficient expert annotations for robust deep learning model development.

Purpose of the Study:

• To introduce a comprehensive, multicenter dataset of breast cancer MRI scans with expert-verified segmentations.
• To facilitate the development and validation of advanced AI models for breast cancer diagnostics and treatment prediction.

Main Methods:

• Compiled 1506 pre-treatment T1-weighted dynamic contrast-enhanced MRI cases from four The Cancer Imaging Archive (TCIA) collections.
• Utilized a deep learning model for initial segmentation, followed by expert correction and verification by 16 experienced breast cancer specialists.
• Integrated 49 harmonized clinical and demographic variables and provided pre-trained weights for a baseline nnU-Net model.

Main Results:

• Established a large-scale, publicly available dataset of breast cancer MRI with expert annotations for primary tumors and non-mass-enhanced regions.
• The dataset includes imaging data, clinical variables, and pre-trained AI model weights, addressing a critical resource gap.
• The rigorous annotation process involved 16 experts, ensuring high-quality data for AI research.

Conclusions:

• This novel dataset significantly enhances resources for AI-driven breast cancer research using MRI.
• It enables the development, validation, and benchmarking of deep learning models for improved breast cancer diagnostics and personalized care.
• The resource is expected to accelerate progress in predicting treatment response and tailoring patient management.