Quick answer: “AI-assisted” is not one product or a guarantee of accuracy. A tool may triage, detect, segment, measure or reconstruct images, and its performance applies only to the cleared or validated intended use. New Dawn should name a vendor, model, version or performance number only after the booked facility confirms current deployment in writing.
What AI can mean in radiology
- Triage: prioritizing studies that may contain a time-sensitive finding.
- Detection: flagging a candidate finding for radiologist review.
- Segmentation or measurement: outlining anatomy or calculating a value.
- Reconstruction: processing acquired data to reduce noise or shorten acquisition.
- Workflow support: worklist, reporting or comparison assistance.
These functions are not interchangeable. A clearance for one body region, modality or task does not establish performance for another. “AI-assisted screening” should therefore never appear as a single undifferentiated accuracy claim.
What AI does not establish
AI does not guarantee that an abnormality will be found, that a false positive will be avoided or that outcomes will improve. Performance in a study may differ from real-world performance because of patient mix, scanner, acquisition, prevalence, workflow and software version. Sensitivity and specificity must be linked to the exact study, population, endpoint and product version.
A radiologist remains responsible for interpreting the examination. AI output is not a diagnosis by itself and does not make an otherwise inappropriate screening test appropriate.
Vendor and facility claims require written confirmation
Articles may discuss vendors such as Lunit, Aidoc, Annalise.ai, HeartFlow or locally developed systems only as examples of different product categories. They must not state that a partner facility uses a specific product unless the facility confirms:
- vendor, product name and software version;
- regulatory status and intended use in the relevant jurisdiction;
- the modality, anatomy and workflow in which it is actually deployed;
- whether the tool is used for the booked service;
- who reviews the output and how errors are handled.
Likewise, PACS platform, scanner model, image count, slice count and processing workflow are operational facts that vary by site and protocol. They must not be inferred from another hospital or a marketing page.
A performance number is publishable only when the text states the exact task, reference standard, population, study design, sample size, confidence interval when available, and important exclusions. It must not be converted into a promise of per-patient accuracy or a claim that one country's screening is more accurate.
Questions for a facility
- Is AI used for this exact booked examination?
- What product and version are deployed?
- What is its intended use and regulatory status?
- Does the radiologist review all images independently?
- Will AI output appear in the patient report?
- What happens if the software is unavailable or disagrees with the reader?
Bottom line
AI can support a radiology workflow, but the clinically meaningful unit is the entire service: appropriate indication, acquisition protocol, image quality, qualified interpretation, report and follow-up. The words “AI-assisted” do not prove that service quality, diagnostic accuracy or patient outcomes are better.
Primary reference
Blog
