Radiology was the first clinical field where AI became ordinary rather than experimental. Systems that detect findings and prioritise worklists now run inside real hospitals, with regulatory clearance and published prospective results. That's a genuine achievement — and it comes with a specific set of problems that don't show up in accuracy numbers.
What's actually deployed
- Mammography. The most studied application. AI is used both as a reader aid — flagging suspicious regions for a radiologist — and, in prospective studies, to triage exams and reduce reading workload. Several systems have regulatory clearance; some health systems have run large prospective evaluations.
- Lung cancer screening. Nodules on chest CT are a well-defined detection task, and AI assistance is in routine use in many programmes.
- Stroke and haemorrhage notification. Systems flag suspected large-vessel occlusion or intracranial haemorrhage and alert the on-call team, compressing the time between scan and treatment decision. This is one of the clearest cases where speed, not accuracy alone, produces benefit.
- Incidental findings and triage. Pulmonary embolism, pneumothorax, fractures — models flag likely positives so they move up the queue.
- Segmentation and measurement. Tumour volumetrics, organ delineation for radiotherapy planning, cardiac functional measures — tasks that are tedious for humans and well suited to automation.
Why imaging was first
Structural reasons, not a coincidence:
- Digital-native data in standardised formats, already archived at scale
- A well-defined task — "is there a finding here" is closer to what pattern recognition does well than diagnosis is
- Ground truth exists, via biopsy, follow-up, or expert consensus
- A natural workflow slot — the model sits between image acquisition and the radiologist, who remains the decision-maker
- Regulatory precedent for software as a medical device
The number that isn't the point
AUC, sensitivity and specificity in a paper are measured under conditions that don't match a hospital. The gap between them is where most disappointment lives:
- Distribution shift. A model trained at three hospitals on one scanner vendor sees a different patient mix, different protocols and different hardware at a fourth. Performance degrades, often substantially.
- Enriched test sets. A dataset with a high prevalence of disease makes sensitivity look good; in a screening population the same model generates far more false alarms.
- Reader studies vs. practice. "Radiologist with AI beats radiologist alone" is usually measured in a lab setting where the radiologist knows they're being studied.
- Outcome evidence is scarce. Detecting more findings is not the same as patients living longer. Trials powered on mortality or interval cancer rates are far rarer than detection studies.
The failure modes that matter
- Automation bias. When the model is usually right, a human reviewer stops scrutinising it. The error rate of the combined human-plus-AI system can end up worse than the human alone for the subset the model gets wrong.
- Alert fatigue. A triage model that fires on too many normal studies gets ignored.
- Overdiagnosis. Better detection finds more indolent disease — abnormalities that would never have caused harm. More diagnoses do not automatically mean better outcomes, and this is a real concern in screening programmes.
- Silent degradation. Performance can drift with scanner software updates or protocol changes, and nobody notices for months without active monitoring.
- Equity gaps. Under-representation in training data produces measurable performance differences across demographic groups — an active and serious area of research.
- Deskilling. A long-term worry rather than a measured one: whether trainees who always read with assistance develop independent skill.
What it means for a patient
- A radiologist still reads your scan. These tools are decision support, not autopilot.
- An AI-detected finding still needs confirmation by the same standards as any other — and it may turn out to be nothing.
- You can ask whether AI was used, and most systems have a policy about disclosing that.
- Screening decisions haven't changed on account of AI. Eligibility by age and risk still governs whether the scan happens at all.
- More findings may mean more follow-up scans — a real trade-off worth knowing before you agree to screening.
How to read a claim about imaging AI
- Was it external validation on data from sites not used in training, or a single-site split?
- Was the test set prevalence realistic for the intended use?
- Was it prospective, or a retrospective analysis of archived images?
- What's the comparator — unaided radiologist, or a consensus panel?
- What are the outcomes: detection, or something a patient would care about?
- Is there post-market monitoring, and is performance published by subgroup?
Bottom line
Imaging AI is real, useful, and narrower than the headlines. It's strongest where the task is a well-defined detection problem with a time-critical output — stroke notification, nodule detection, mammography triage. Its risks aren't primarily about the model being inaccurate; they're about distribution shift, over-reliance, and the fact that detecting more abnormalities is not the same as improving outcomes. The right frame is a very good assistant in a workflow that still has a trained human at the end of it.
This article is educational and discusses workflow tools rather than providing medical advice. Patients and clinicians should follow their institution's policies.