Image recognition systems now match specialists on several narrow tasks.
Reported accuracy in published studies is frequently above ninety percent.
Accuracy on a dataset, however, is not performance in a clinic.
Hospital images are taken on older machines, at odd angles, on patients who move.
Models trained in one country routinely lose several points of accuracy in another.
The deeper issue is what the number means for a patient rather than a dataset.
A screening tool applied to a low- population generates mostly .
Each false positive costs a , three weeks of fear and a small physical risk.
Deployment therefore depends on , not only on the model.
The most successful uses so far are not at all.
Systems that a waiting list, flag urgent scans or draft a report save time without making the final call.
Responsibility remains with a clinician who signs.
is the quiet danger in that arrangement.
A tired doctor at the end of a shift agrees with the machine more often than the evidence warrants.
Any serious evaluation measures the doctor-plus-system pair, which is the thing that actually treats the patient.