Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/dataset.json

Name

Dynamic Contrast-enhanced breast MRI dataset

Link

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11427917/

Indexing

Keywords: DCIS, upgrade, invasive ductal carcinoma, breast MRI, DCE MRI, deep learning, CNN-LSTM, time-dependent modeling

Content: BR, MR, OI

RadLex: RID45839, RID10312, RID4265, RID49531

Author(s)

John D. Mayfield

Dana Ataya

Mahmoud Abdalah

Olya Stringfield

Marilyn M. Bui

Natarajan Raghunand

Bethany Niell

Issam El Naqa

Organization(s)

University of South Florida College of Medicine

H. Lee Moffitt Cancer Center and Research Institute

Contact

Corresponding author: John D. Mayfield (email reported in article)

Funding

NIH/NCI R01CA249016; Moffitt Cancer Center Support Grant P30-CA076292. Additional related disclosures reference R01CA248016-01A1.

Ethical review

HIPAA-compliant; IRB approved with waiver of informed consent.

Comments

Retrospective single-institution cohort of women with core-biopsy–proven DCIS who underwent preoperative DCE breast MRI (2012–2022). 154 cases; 25 upgraded to invasive cancer at surgery and 129 not upgraded. Data available by request from corresponding author; model and weights shared via GitHub for 2 years post-publication.

Date

Published: 2024-06-20

Created: 2012-01-01

References

[1] Mayfield JD, Ataya D, Abdalah M, Stringfield O, Bui MM, Raghunand N, Niell B, El Naqa I. "Presurgical Upgrade Prediction of DCIS to Invasive Ductal Carcinoma Using Time-dependent Deep Learning Models with DCE MRI". Radiology: Artificial Intelligence. 2024-06-20. doi:10.1148/ryai.230348. PMID: 38900042. PMCID: PMC11427917.

Dataset

Motivation

To evaluate whether time-dependent deep learning on DCE MRI predicts preoperative DCIS upgrade to invasive cancer without requiring lesion segmentation.

Sampling

Consecutive selection from retrospective cohort meeting inclusion/exclusion criteria (2012–2022).

Partitioning scheme

Stratified 10-fold cross-validation yielding an 80:10:10 split with a randomized 10% holdout test set; holdout data never used in training/validation.

Missing information

No external independent test set; exact per-partition counts not itemized.

Relationships between instances

Single-institution cohort; each case corresponds to a woman with biopsy-proven DCIS and a preoperative DCE MRI exam; outcome defined by surgical pathology (invasive vs DCIS; microinvasion counted as invasive).

Noise

Potential dynamic range loss converting MR images into .avi frames noted by authors.

External data

No external/open-source datasets were used; single-institution PACS source.

Confidentiality

Data were de-identified; HIPAA-compliant; IRB waiver of consent.

Re-identification

Images de-identified; cropping removed mediastinum and contralateral breast and excluded laterality markers/annotations.

Sensitive data

Contains health information (breast MRI and pathology outcomes).