Blog · ppj.es
Mirai: The AI That Outperforms Breast Density in Predicting Breast Cancer at 5 Years
Published on 6 August 2026
🇪🇸 Leer este artículo en español
A deep learning model applied to screening mammograms (Mirai) predicts breast cancer risk at 5 years with AUROC 0.71, compared to 0.53 for BI-RADS breast density [PMID: 42554691]. The difference is not merely statistical; it changes the clinical decision about which women to offer supplemental screening (MRI/ultrasound).
What was done
Retrospective multicenter cohort study (Mass General Brigham, 5 sites) with 123,091 mammograms from 67,019 women (mean age 58 years, follow-up through 2023 to capture 5-year outcome) [DOI: 10.1001/jamanetworkopen.2026.10559]. All mammograms were 2D full-field digital (Hologic). Mirai — a model trained on 210k prior mammograms, externally validated, open-source code — was applied without retraining. The output is a numerical score for absolute 5-year risk, with no heatmaps or lesion localization (black box) [PMID: 42554691].
Breast density was classified according to BI-RADS 5th ed. (non-dense = fatty + scattered fibroglandular; dense = heterogeneously + extremely dense) — the binary criterion used by the FDA and >30 US states to decide coverage for supplemental screening [DOI: 10.1001/jamanetworkopen.2026.10559].
Key results
| Metric | Mirai (DL) | BI-RADS density | p |
|---|---|---|---|
| AUROC cancer at 5 years | 0.71 (95%CI 0.70-0.72) | 0.53 (0.52-0.54) | <0.001 [PMID: 42554691] |
| AUROC dense breasts | 0.70 | 0.53 | — [PMID: 42554691] |
| AUROC non-dense breasts | 0.72 | 0.53 | — [PMID: 42554691] |
| Cancer incidence by DL risk | |||
| Low risk (<1.7%) | 1.0% (450/44,502) | — | <0.001 [DOI: 10.1001/jamanetworkopen.2026.10559] |
| Intermediate (1.7-3.0%) | 2.7% (1347/50,683) | — | [DOI: 10.1001/jamanetworkopen.2026.10559] |
| High risk (>3.0%) | 6.2% (1736/27,906) | — | [DOI: 10.1001/jamanetworkopen.2026.10559] |
| False negatives (BI-RADS 1-2 + cancer ≤1 year) | |||
| Overall FN rate | 1.1/1000 (135/123k) | — | [PMID: 42554691] |
| FN dense vs non-dense breasts | 1.7 vs 0.6/1000 | <0.001 [PMID: 42554691] | |
| FN high-risk DL | 2.1/1000 | — | <0.001 [PMID: 42554691] |
Practical key: Adding density to the DL model does not improve AUROC (0.70 vs 0.71, p=0.08) [DOI: 10.1001/jamanetworkopen.2026.10559]. AI already captures the prognostic information of density — and more.
Subgroups: real equity
- White women: DL 0.70 vs density 0.53 [PMID: 42554691]
- Black women: DL 0.72 vs density 0.57 [PMID: 42554691]
- Asian women: DL 0.69 vs density 0.54 [PMID: 42554691]
- Hispanic women: DL 0.69 vs density 0.56 [PMID: 42554691]
- Invasive: DL 0.71 vs 0.53 | DCIS: DL 0.70 vs 0.56 [DOI: 10.1001/jamanetworkopen.2026.10559]
The model works equally well in dense and non-dense breasts [PMID: 42554691], and across races/ethnicities [DOI: 10.1001/jamanetworkopen.2026.10559]. Density fails due to subjectivity (interobserver variability) and because it is binary: it treats the same a woman with 51% dense tissue as one with 90% [PMID: 42554691].
The clinical case that illustrates it (Fig. 3 of the paper)
40-year-old woman, history of prior left breast cancer. Screening mammogram: **scattered fibroglandular tissue + postsurgical changes
— This analysis was generated by ANGIE (Always Next to Guide, Inspire and Empower), an artificial intelligence system with SOUL profiles, designed by Dr. Javier Pumares Pérez.
Disclaimer: this article is educational and informational in nature and reflects the personal opinion of the author. It does not constitute medical advice nor replace the assessment of a healthcare professional. If you have a health concern, consult your physician.