A mammogram looks normal. The radiologist signs it off. Five years later, the patient is diagnosed with breast cancer. This scenario plays out thousands of times every year. But an algorithm developed at MIT can now read what humans cannot, predicting breast cancer risk from a single mammogram up to five years before any visible signs appear.
Mirai: Seeing What Radiologists Cannot
The algorithm is called Mirai. Developed by researchers at MIT's Computer Science and Artificial Intelligence Laboratory, it was trained on over 200,000 mammograms from Massachusetts General Hospital. Unlike traditional risk models that rely on family history, age, and genetics, Mirai works directly from the image itself. It reads patterns in breast tissue that are invisible to the human eye.
26%
of interval cancers flagged by Mirai's top 10% risk band
Source: RSNA / UK Breast Screening study, 2025
A study across the UK breast screening programme examined 134,217 women, with 524 interval cancers. The results were striking. Ranked by Mirai's risk score, the top 10% of women accounted for around 26% of those interval cancers, the ones that appear between routine screenings and are typically caught late. These are the cancers that slip through the cracks of standard screening programmes. The AI flagged the women most at risk years before the cancer became clinically apparent.
Beyond Mammography
Breast cancer screening is just the beginning. AI diagnostic tools are expanding across medical imaging at pace.
Brain MRIs in seconds. Researchers at the University of Michigan built a system that interprets brain MRI scans in seconds, a process that typically takes radiologists 20-30 minutes per scan. The system identifies abnormalities across stroke, tumour, and neurodegenerative conditions with accuracy comparable to specialist neuroradiologists.
Biological age from chest X-rays. A team at Osaka Metropolitan University demonstrated that AI can estimate a patient's biological age from a standard chest X-ray, then flag individuals whose biological age significantly exceeds their chronological age. These patients carry higher risk for cardiovascular disease, respiratory illness, and early mortality. The information was always in the image. Nobody could see it until now.
5 years
advance prediction window for breast cancer risk from a single mammogram
Source: MIT CSAIL / Massachusetts General Hospital
The Human-AI Collaboration Model
None of this replaces radiologists. That framing misses the point entirely.
The model that works is triage and augmentation. AI scans every image first, flagging cases that warrant closer attention and prioritising the queue by risk level. Radiologists then focus their expertise where it matters most, on complex cases, on ambiguous findings, on the conversations with patients that require human judgement and empathy.
Risk scoring lets a screening service point its scarce specialist time where the danger is concentrated. In the UK study, the higher the risk band, the larger the share of interval cancers it captured: the top 5% of women accounted for around 15% of interval cancers, the top 10% for 26%, and the top 20% for 42%. That is a queue you can prioritise.
Source: RSNA / UK Breast Screening study, 2025
This matters because radiologist shortages are real and worsening. The Royal College of Radiologists reported a 30% shortfall in the UK workforce. NZ faces similar pressures. AI doesn't solve the shortage, but it makes the existing workforce dramatically more effective.
What This Means for Screening Programmes
Population-level screening is expensive and resource-constrained. Most countries screen women over 50 every two to three years. The intervals exist because the system cannot process more volume with existing staff.
AI changes this calculation. If every mammogram can be risk-scored automatically, screening programmes can move from fixed intervals to risk-adaptive scheduling. High-risk patients get screened more frequently. Low-risk patients get reassurance backed by data, not just averages.
42%
of interval cancers captured within Mirai's top 20% risk band
Source: RSNA / UK Breast Screening study, 2025
The Careful Part
This is health, not just technology. The stakes are different from optimising a supply chain or automating a workflow.
False positives cause anxiety, unnecessary biopsies, and real psychological harm. False negatives cost lives. Every deployment of AI in medical imaging must be validated against diverse populations, monitored for drift, and governed with clinical oversight that prioritises patient welfare above throughput.
The early results are extraordinary. But rolling these systems out responsibly, across different populations, healthcare systems, and regulatory environments, requires the same care as any medical intervention.
The science is moving fast. The governance needs to keep pace.