AI and breast cancer — frequently asked questions
Straight answers on AI in breast cancer care: accuracy, approvals, safety, bias and what AI can — or can't — do.
AI and breast cancer — frequently asked questions
Straight answers on AI in breast cancer care: accuracy, approvals, safety, bias and what AI can — or can't — do.
On average, studies show that AI-assisted reading finds more cancers than a radiologist alone — a Swedish trial reported around 20% more. But the best results come from human–AI collaboration, not either one alone. AI never replaces the radiologist; it supports them.
In the United States, several AI systems are FDA-cleared for breast imaging. In Europe, AI devices must carry the CE mark under the Medical Devices Regulation and increasingly satisfy the European AI Act. Approval varies by country — ask your centre which systems it uses and what evidence supports them.
No. AI is a decision-support tool. A radiologist or pathologist examines every case and makes the final decision. If AI flags something, a human reviews it. If they disagree, the case is re-read. Your doctor remains the decision-maker.
The model converts the image into numerical features and compares them with patterns learned from hundreds of thousands of annotated mammograms. It highlights regions it judges suspicious and assigns a risk score. Radiologists see those highlights and decide what happens next.
Risk models estimate the probability of developing breast cancer in the coming years by combining imaging, genetics and lifestyle. It is a probability, not a certainty, and it helps personalise screening — without replacing medical advice.
Patient data is protected by strict laws — GDPR in Europe, HIPAA in the United States. Systems use encryption, access controls, de-identification and audit logs. Ask your centre how it protects imaging data and who it is shared with.
AI can be less reliable for populations under-represented in its data; it can produce false positives and false negatives; it may not generalise to other equipment or regions; and it cannot consider a patient's full context. That is why human review is always in place.
Not automatically. Models trained mostly on one population can underperform for others — by age, breast density, ethnicity or socio-economic background. Regulators now require subgroup performance reporting, and diverse training data is the first line of defence.
It already is in many programmes. AI-supported screening is rolling out across several national screening services. The pace varies by region, funding and regulation — but the direction is clear.
Look for evidence: peer-reviewed studies, independent validation, regulatory authorisation documents. Be wary of marketing numbers with no study behind them. The machine learning and regulation sections of this site explain how to read the evidence.