AI in imaging: where it helps the radiologist and where it still fails
Triage, prioritisation, detection — three different functions with three different risks. What a radiology service working with algorithms looks like.
18 February 2025 · 8 min
Key points
- Most systems in practice prioritise or flag; they do not diagnose on their own.
- A negative algorithm output does not cancel the clinician’s read.
- Clinical benefit is measured in time to treatment, not only in sensitivity on a test set.
Three different functions
There are three things an algorithm can do in an imaging service and they are often conflated in pitches. Detection marks a lesion on the image. Prioritisation moves a study up the worklist. Triage decides, before reading, whether a case looks urgent. Each fails differently and each must be measured differently.
Prioritisation in stroke
Stroke is the best-documented case, because the therapeutic window is hours, not weeks. Systems that detect a large-vessel occlusion on CT angiography and alert the on-call team shorten time to assessment — provided the hospital already has a flow ready for the alert. Technology does not create the flow; it accelerates it, or exposes its absence.
Screening: the mammography lessons
Screening programmes are the most sensitive ground, because decisions are made at scale on healthy people. Here not only sensitivity matters but recall rate and predictive value: a system that raises detection at the cost of many more women recalled for further tests can do more harm than good. That is why guidance asks for prospective studies with double reading and clinical outcomes, not only comparisons on datasets.
Where the errors appear
The three most common sources of error are not “the AI is bad” but the environment. Images of a different quality than the training set, different scanners, changed protocols. Lesions the model does not know because the label never existed in the dataset. And shortcut learning, where the model picks up a contextual marker — a drainage tube, a film marker — instead of the sign of disease.
Then there is the effect on the reader: a radiologist receiving forty alerts per shift starts ignoring them. Alert fatigue is a measurable clinical outcome, not a comfort issue.
What to ask the vendor
On which population and which scanner model was it validated? What is performance in my subgroups — age, sex, comorbidity? What happens with poor-quality images? How is the output recorded in the record and who is legally accountable? Is there post-deployment monitoring, and who reads it?
If the vendor answers all of that with a brochure, you do not have an answer — you have a promise.