The Cancer Imaging Archive

BCBM-RadioGenomics | MRI Dataset of Metastatic Breast Cancer to the Brain with Expert-reviewed Segmentations and Tumor-derived Radiomic Features 

DOI: 10.7937/rrse-w278 | Data Citation Required | 9 Views | Image Collection

Summary

This dataset consists of 268 T1-post contrast Magnetic Resonance Images (MRIs) in NIfTI format from 165 patients with histologically confirmed metastatic breast cancer to the brain. Some patients underwent repeat Gamma Knife treatments (usually for recurrence), resulting in 268 total studies. Each MRI has a unique identifier to the patient and their treatment session. For example, if a patient had their first treatment this was labeled PatientID-0, and their second treatment (maybe some months later for a new metastasis) labeled as PatientID-1.

The data were obtained on 1.5T and 3T scanners for stereotactic radiosurgery (SRS) treatment for metastatic breast cancer to the brain.  Each scan has isotropic spacing varying from 0.5mm to 1.0mm (although a fixed value in each scan). A brain extraction tool was applied to remove the skull and face (HD-BET). We also performed intensity inhomogeneity correction using N4 bias correction. 

The dataset includes segmented masks for tumor, postoperative tumor bed/resection cavity, and neighboring structures of interest as well as a comprehensive catalog of 107 tumor-derived radiomic features (for each segmentation). The segmentations were created and reviewed by at least one neurosurgeon and one radiation oncologist during their initial treatment.  We also provide clinical demographics (patient age and gender) and genetic status for progesterone receptor (PR), estrogen receptor (ER), and Human Epidermal Growth Factor Receptor 2 (HER2). 

This collection provides a platform for diverse applications ranging from the development of machine learning models in computer vision and radiogenomics to educational uses for neurosurgery, radiation oncology, and neuroradiology trainees.  The comprehensiveness of this data makes it a prime resource for investigations on any individual component (raw images, genetics, radiomics, etc.)–whether serving as a validation dataset or as an exploratory one. Experts in computer vision may use the expert-labeled segmentations here to augment or train a tumor segmentation algorithm de novo. For more radiogenomic investigations, researchers may decide they do not agree with the radiomic features provided and want to try a different set of radiomic features (i.e. a wavelet transformed or scatter transformed) and re-train for a classification goal (genetic status prediction, outcome prediction, etc.). Our data provides the raw segmentation information in which researchers can derive their own radiomic features using their algorithms of choice. 

Note: There are minor inconsistencies in the segmentations/label names. The segmentations/label names were written years ago by several technicians who used varying abbreviations, misspellings, and acronyms.

Data Access

Please contact the helpdesk to request access to the Images of the head, not skull-stripped data files (165 Participants, 11.7 GB).

Version 1: Updated 2025/01/28

Citations & Data Usage Policy

Data Citation Required: Users must abide by the TCIA Data Usage Policy and Restrictions. Attribution must include the following citation, including the Digital Object Identifier: