TY - GEN
T1 - BreastMR-'FM v1.0
T2 - Medical Imaging 2026: Computer-Aided Diagnosis
AU - Bhattacharya, Moinak
AU - Singh, Gagandeep
AU - Chen, Chao
AU - Prasanna, Prateek
N1 - Publisher Copyright:
© COPYRIGHT SPIE.
PY - 2026/4/2
Y1 - 2026/4/2
N2 - Breast cancer remains a leading cause of cancer-related mortality, highlighting the need for robust imaging-based biomarkers to improve diagnosis, treatment response prediction, and survival stratification. While artificial intelligence (AI) has shown promise in breast imaging, existing models are typically task-or protocol-specific, limiting generalizability. We present a breast MRI foundation model (BreastMR-FM v1.0) trained in a selfsupervised manner on 1,506 cases (7,926 scans) from the MAMA-MIA dataset, comprising heterogeneous dynamic contrast-enhanced (DCE) MRI protocols. Using a 3D SimCLR framework with a ResNet10 encoder, we learned contrast-invariant imaging representations, subsequently evaluated across multiple downstream tasks. On external validation cohorts (DUKE, ISPY2), the model achieved strong performance in HER2 mutation prediction (ACC: 0.89 ± 0.04, DUKE; 0.81 ± 0.06, ISPY2) and recurrence classification (ACC: 0.88 ± 0.04, DUKE). Tumor grade prediction showed lower accuracy (0.54 ± 0.26, ISPY2), reflecting known grading variability. For survival analysis, embeddings achieved a C-index of 0.63 ± 0.04, with model-derived risk groups significantly stratifying outcomes (P < 0.001). These findings demonstrate that a single, contrast-Agnostic MRI foundation model can generalize across imaging protocols and support diverse clinical prediction tasks, establishing a versatile backbone for precision breast cancer care. Code will be availabe soon.
AB - Breast cancer remains a leading cause of cancer-related mortality, highlighting the need for robust imaging-based biomarkers to improve diagnosis, treatment response prediction, and survival stratification. While artificial intelligence (AI) has shown promise in breast imaging, existing models are typically task-or protocol-specific, limiting generalizability. We present a breast MRI foundation model (BreastMR-FM v1.0) trained in a selfsupervised manner on 1,506 cases (7,926 scans) from the MAMA-MIA dataset, comprising heterogeneous dynamic contrast-enhanced (DCE) MRI protocols. Using a 3D SimCLR framework with a ResNet10 encoder, we learned contrast-invariant imaging representations, subsequently evaluated across multiple downstream tasks. On external validation cohorts (DUKE, ISPY2), the model achieved strong performance in HER2 mutation prediction (ACC: 0.89 ± 0.04, DUKE; 0.81 ± 0.06, ISPY2) and recurrence classification (ACC: 0.88 ± 0.04, DUKE). Tumor grade prediction showed lower accuracy (0.54 ± 0.26, ISPY2), reflecting known grading variability. For survival analysis, embeddings achieved a C-index of 0.63 ± 0.04, with model-derived risk groups significantly stratifying outcomes (P < 0.001). These findings demonstrate that a single, contrast-Agnostic MRI foundation model can generalize across imaging protocols and support diverse clinical prediction tasks, establishing a versatile backbone for precision breast cancer care. Code will be availabe soon.
KW - Breast DCE MRI
KW - Foundation models
KW - Tumor characterization
UR - https://www.scopus.com/pages/publications/105040929125
U2 - 10.1117/12.3088150
DO - 10.1117/12.3088150
M3 - Conference contribution
AN - SCOPUS:105040929125
T3 - Progress in Biomedical Optics and Imaging - Proceedings of SPIE
BT - Medical Imaging 2026
A2 - Wismuller, Axel
A2 - Deserno, Thomas Martin
PB - SPIE
Y2 - 15 February 2026 through 19 February 2026
ER -