Radiology AI glossary

The terms you need to read a study, a clearance or a radiology AI offer.

AUC (area under the ROC curve): Measure of a test's ability to separate positive from negative exams across all possible thresholds: 1 is perfect, 0.5 is chance.
Sensitivity: Share of truly positive exams (for example with a fracture) that the test flags.
Specificity: Share of truly negative exams that the test leaves without an alert.
Accuracy: Share of correct answers, positive and negative combined, over all exams.
Head-to-head study: Study that compares several software products on the same exams, against the same reference standard.
Reader study: Study that measures the performance of radiologists reading the same exams with and without the AI.
Standalone evaluation: Measure of the AI's performance on its own, without a radiologist.
Reference standard: Diagnosis held as correct in a study, often the consensus of expert radiologists.
CE marking (MDR): Conformity of a medical device with EU Regulation 2017/745; from Class IIa upwards, a notified body certifies it.
Class IIa: Intermediate risk class of the EU medical device regulation, the class of most software that provides information used for diagnostic decisions (Rule 11).
Notified body: Body designated by a member state to assess the conformity of medical devices. Milvue's is BSI (number 2797).
510(k): FDA pathway that clears a device for the US market by substantial equivalence to a device already legally marketed. Each clearance has a number starting with K, searchable in the FDA public database.
DICOM: Standard for exchanging and storing medical images and the information that goes with them.
PACS: System that archives and displays medical images.
RIS: Radiology information system: appointments, exams, reports.
IHE: Initiative that defines integration profiles between healthcare systems and tests them at Connectathons.
Triage (worklist): Sorting of the exams to read: the AI moves those where it detects an urgent finding to the top of the worklist.
Cobb angle: Angle measured on a spine radiograph to quantify scoliosis.
EOS: Low-dose radiography system that images the whole body standing, front and side, and allows 3D measurements.
Bone age: Estimate of skeletal maturity, usually from a hand and wrist radiograph.
Speech recognition: Automatic transcription of speech into text; in radiology it must also recognise punctuation and layout commands.
Word error rate: Number of words wrongly recognised (substituted, deleted or inserted), divided by the number of words in a reference transcript.
Report template: Report model specific to a radiologist or a department, for a type of exam.
VLM (vision-language model): AI that links an image and text, for example to suggest a description of an exam.