h2thez3/breast-mri-sagitall-landmarks

Breast MRI sagittal landmarks

Breast-region masks and nipple landmarks for 35 sagittal, unilateral breast DCE-MRI exams.

This is a subset of the sagittal exams from the TCIA QIN-BREAST collection: <https://www.cancerimagingarchive.net/collection/qin-breast/>. The images are redistributed unchanged from that collection under its CC BY 3.0 licence; the masks are new.

Sagittal breast MRI is under-represented in public landmark data. Models trained on axial bilateral exams are known to fail on it, and there was no annotated sagittal set to measure that against. This is one.

Contents

images/<case>.nii.gz                  pre-contrast volume
labels/breast_region/<case>.nii.gz    binary mask, whole breast INCLUDING the nipple
labels/nipples/<case>.nii.gz          binary mask, the nipple landmark
metadata.csv                          one row per case

All three volumes for a case share one grid, so a label overlays its image with no resampling. Every exam is sagittal, unilateral, 3.0 T Philips, 192x192x20 at 1.33x1.33x5.0 mm.

The nipple is INSIDE the breast region, not cut out of it — the breast-region mask covers the whole breast including the nipple, and the nipple mask marks the landmark within it. Overlap is 100% on 29 of 35 cases and 67–85% on the remainder, where a traced nipple protrudes slightly past the traced skin boundary.

How it was made

DICOM converted with mri_dicom_pipeliner; acquisition plane and laterality derived from the NIfTI affine; annotation in 3D Slicer with the SegmentationReview extension.

• —astra_reviewed — 25 case(s); produced by an AI agent, then reviewed and corrected by a human
• —user — 10 case(s); annotated from scratch by a human

Both annotator groups were reviewed in Slicer by a human before acceptance.

Known limitations

Nipple size is not consistent between annotators. Median nipple volume is 10,182 mm³ for the manually-drawn subset and 1,600 mm³ for the agent-assisted one — roughly a 6× difference. The agent's markers cluster tightly, which reads as a fixed-size marker rather than a traced structure. Both are internally consistent, but they are different targets for the same anatomy. Evaluating on nipple CENTROIDS rather than nipple volume is insensitive to this. Breast-region volumes agree well between the two groups.

These are research annotations without clinical validation. 2 case(s) sit outside a 279,000–2,634,757 mm³ breast-volume plausibility band or carry more nipple components than the laterality implies; metadata.csv flags them.

One exam is a coverage judgement. QIN-BREAST-01-0043 has slices deliberately left empty where breast tissue could not be distinguished from chest wall and ghosting artifact.

Licence and citation

Derived from TCIA QIN-BREAST, released under CC BY 3.0, which permits sharing and adaptation including commercial use with attribution. These annotations are released under the same licence. If you use this dataset, cite the source collection:

Li, X., Abramson, R. G., Arlinghaus, L. R., Chakravarthy, A. B., Abramson, V. G., Sanders, M., & Yankeelov, T. E. (2016). Data From QIN-BREAST (Version 2) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/K9/TCIA.2016.21JUEBH0

and abide by the TCIA Data Usage Policy.