MamaDino: A Hybrid Vision Model for Breast Cancer 3-Year Risk Prediction
Organizations: Fondazione Human Technopole, Milan, Italy
Abstract
Breast cancer screening programmes increasingly seek to move from one-size-fits-all intervals to risk-adapted and personalized strategies. Deep learning (DL) has enabled image-based risk models with stronger 1- to 5-year prediction than clinical models, but leading systems (e.g., Mirai) typically use convolutional backbones, high-resolution inputs (>1M pixels) and simple multi-view fusion, with limited explicit modelling of contralateral asymmetry. We hypothesised that combining complementary inductive biases (convolutional and transformer-based) with explicit contralateral asymmetry modelling would allow us to match state-of-the-art 3-year risk prediction performance on lower-resolution mammograms, suggesting that using less detail in a more structured way can recover state-of-the-art accuracy. We present MamaDino, a mammography-aware multi-view attentional DINO model. MamaDino fuses frozen self-supervised DINOv3 ViT-S features with a trainable CNN encoder at 512x512 resolution and aggregates bilateral breast information via a BilateralMixer to output a 3-year breast cancer risk score. We train on 53,883 women from OPTIMAM (UK) and evaluate on matched 3-year case-control cohorts: an in-distribution test set from four screening sites and an external out-of-distribution cohort from an unseen site. At breast level, MamaDino achieves discrimination comparable to Mirai on internal and external tests while using ~13x fewer input pixels. Adding the BilateralMixer improves patient-level discrimination to AUC 0.736 (vs Mirai 0.713) in-distribution and 0.677 (vs Mirai 0.676) out-of-distribution, with overlapping confidence intervals. Performance was evaluated across age, ethnicity, scanner manufacturer, tumour type and grade. These findings suggest that complementary inductive biases and explicit contralateral modelling can achieve Mirai-comparable discrimination at lower resolution.
Figures & tables
| In-Distribution Test Set | OOD Test Set | ||||
| Model | Resolution | AUC | (95% CI) | AUC | (95% CI) |
| DINO-only* | 512 512 | 0.621 | – | – | – |
| SEResNeXt-only* | 512 512 | 0.668 | – | – | – |
| MamaDino (resolution ablation)* | 0.638 | – | – | – | |
| MamaDino (resolution ablation)* | 0.649 | – | – | – | |
| MamaDino † (concat. ablation) | 0.731 | – | – | – | |
| Test Cohorts | In-Distribution Test | OOD Test Set | ||
|---|---|---|---|---|
| MamaDino | Mirai | MamaDino | Mirai | |
| Age (y) | ||||
| 0.746 (0.710-0.783) | 0.727 (0.690-0.764) | 0.668 (0.626-0.710) | 0.693 (0.651-0.735) | |
| 0.705 (0.650-0.760) | 0.656 (0.597-0.715) | 0.675 (0.618-0.732) | 0.657 (0.596-0.717) | |
| 0.747 (0.696-0.797) | 0.743 (0.691-0.794) | 0.691 (0.624-0.758) | 0.672 (0.606-0.737) | |
| Ethnicity | ||||
| AUC (95% CI) | AUC (95% CI) | ||||
|---|---|---|---|---|---|
| Subgroup | MamaDino | Mirai | |||
| In-Distribution Test Set | |||||
| Largest lesion diameter | |||||
| 10 mm | 112 | 0.687 [0.63, 0.74] | 0.665 [0.61, 0.72] | +0.023 [-0.034, 0.079] | 0.43 |
| 11–20 mm | 156 | 0.745 [0.71, 0.78] | 0.698 [0.65, 0.74] | +0.047 [0.000, 0.094] | 0.051 |
| 21–30 mm | 82 | 0.724 [0.66, 0.78] | 0.736 [0.68, 0.79] | -0.012 [-0.071, 0.048] | 0.70 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Ethnicity | Number of Women |
|---|---|
| White | 30,682 |
| Asian | 3,808 |
| Black | 1,851 |
| Mixed | 601 |
| Other | 1,107 |
| Missing | 15,834 |
| Test Cohorts | In-Distribution Test Set | OOD Test Set | ||
|---|---|---|---|---|
| Control (n=1050) | Case (n=525) | Control (n=1504) | Case (n=376) | |
| Age (y) | ||||
| 532 (50.7%) | 266 (50.7%) | 728 (48.5%) | 182 (48.4%) | |
| 260 (24.8%) | 130 (24.8%) | 419 (27.8%) | 105 (27.9%) | |
| 258 (24.6%) | 129 (24.6%) | 357 (23.8%) | 89 (23.7%) | |
| Median* | 59 (54–64) | 59 (54–64) | 60 (55–64) | 60 (55–64) |
Breast Cancer
Breast cancer research develops methods to detect and characterize tumors, estimate risk, and guide treatment using imaging, pathology, and clinical data.
See all 51 papers