SpurBreast Dataset

The SpurBreast dataset is a curated real-world breast MRI resource designed to study how deep learning models exploit spurious correlations. Built on top of the DUKE Breast Cancer dataset of 3D MRI scans from over 900 patients with biopsy-confirmed invasive breast cancer, SpurBreast introduces carefully constructed training and validation splits that either deliberately include or avoid non-clinical signals. These splits cover demographic and imaging-protocol features such as magnetic field strength, vertical orientation, menopause status, race and ethnicity, and surgery type.