When an AI vendor describes their radiology tool, they will typically choose one of two labels: clinical decision support (CDS) or diagnostic AI. These are not interchangeable marketing terms. They reflect different regulatory classifications, different liability frameworks, and different assumptions about the relationship between the AI output and the radiologist's judgment. Getting clear on the distinction matters before you sign a contract, configure a workflow, or explain to your department what the tool is and is not doing.
The FDA's Software as a Medical Device Framework
The regulatory starting point is the FDA's Software as a Medical Device (SaMD) framework, which in the US is informed by the International Medical Device Regulators Forum guidance that the FDA adopted in its 2017 discussion paper and subsequent 2019 proposed regulatory framework for AI/ML-based SaMD.
Under the 21st Century Cures Act, Congress created a category of software that is specifically excluded from FDA device regulation: software intended for administrative support, general wellness, certain electronic health records functions, and, critically for this discussion, clinical decision support software that meets specific criteria. Section 3060 of the 21st Century Cures Act defines non-device CDS as software that: displays, analyzes, or prints medical information; supports or provides recommendations to a health care professional; is not intended to replace the clinical judgment of a health care professional; and requires independent review of the basis for the recommendation by the health care professional receiving it.
The operative phrase is "requires independent review of the basis for the recommendation." If the AI outputs a finding or recommendation in a way that the clinician is expected to verify by reviewing the underlying data, that structure supports classification as non-device CDS. If the AI output is intended to stand on its own or to be acted upon without the clinician reviewing the basis, it moves toward the device end of the spectrum.
What This Means for Radiology AI Products
A radiology AI tool that surfaces a triage priority score and requires the radiologist to read the study and form their own opinion before the report is signed is structured as CDS. The AI output is an input to the radiologist's workflow, not a replacement for their review. This is the architecture of pre-reading: the AI pre-reads and prioritizes; the radiologist reads and reports. The radiologist's review is independent of the AI's pre-read output.
A tool that generates a diagnostic output intended to be acted upon without radiologist review of the underlying images is structured differently. Automated screening tools that generate a positive/negative classification for release without radiologist sign-off on each study are closer to the diagnostic device end of the spectrum, and the regulatory and liability implications are different.
Most radiology AI products in the current market are positioned as CDS to take advantage of the non-device exclusion under the 21st Century Cures Act, which reduces the regulatory burden significantly compared to going through 510(k) clearance or De Novo classification as a device. That positioning is legitimate for tools that genuinely maintain the independent review structure. It becomes problematic if the actual workflow design undermines the independence of the review.
The Liability Dimension
The radiologist's professional and legal liability for a signed report does not change based on how the AI output is labeled. Whether the AI is described as CDS or diagnostic AI, the radiologist who signs the report is attesting to its accuracy based on their independent review. The AI's role in the process is background context for workflow efficiency, not a defense against a missed finding claim.
This has a practical implication that is sometimes glossed over in vendor conversations: the CDS classification is a regulatory construct that affects the vendor's obligations, not the clinician's. If an AI pre-read tool is used in a workflow where the radiologist regularly accepts AI findings without independent verification because the volume pressure makes thorough review impractical, the CDS label does not create a liability shield for the radiologist. The radiologist signed the report. Their standard of care applies.
We are not saying that AI pre-reads increase radiologist liability. The point is that the CDS vs. diagnostic AI distinction primarily affects vendor regulation, not clinician accountability. A radiologist should not choose a workflow based on an assumption that the CDS label changes their standard of care. It does not.
What "Cleared" and "Authorized" Actually Mean
AI vendors in radiology sometimes describe their products as FDA-cleared or FDA-authorized. These are specific terms with specific meanings. 510(k) clearance means the FDA has reviewed the device and determined it is substantially equivalent to a legally marketed predicate device. De Novo classification is used for novel, lower-risk devices with no predicate. Premarket approval (PMA) is the highest regulatory standard, required for Class III devices posing significant patient risk.
A product that is described as FDA-cleared has gone through 510(k) review for a specific intended use. A product described as CDS under the 21st Century Cures Act non-device exclusion has not gone through FDA device review at all. Both products may be legitimate and well-validated. But they represent fundamentally different regulatory pathways, and a department purchasing a radiology AI tool should understand which one applies to the specific product and intended use.
The nuance that gets missed most often: FDA clearance or authorization applies to a specific intended use as described in the submission. A device cleared for detecting pneumothorax on chest X-ray has not been cleared for detecting pneumothorax on chest CT, even if the algorithms are similar. Using a cleared device outside its cleared intended use shifts the regulatory and liability picture. This is worth asking your vendor specifically: what is the cleared or authorized intended use, and does our planned workflow match that use?
Implications for AI Draft Reports
AI-generated report drafts sit in an interesting position in this framework. A draft that the radiologist edits and signs is clearly CDS in its structure: the radiologist is reviewing and modifying the output before any clinical action occurs. The independent review criterion is satisfied because the radiologist is doing active editing work, not just approving a completed output.
Where it gets complicated is if draft acceptance rates become very high for a specific study type and the editing activity becomes nominal. If a radiologist is accepting 95% of AI drafts for normal chest X-rays with changes only to patient name and study date, are they performing independent review of the basis for the AI's assessment? Technically yes, because they are still reading the images and reviewing the draft. Practically, the review is light relative to what it would be for a complex study. The CDS structure is preserved, but the workflow intensity of the review has shifted.
This is not necessarily a problem. A radiologist who reads a clearly normal chest X-ray, reviews a structurally accurate draft, and signs with minor edits is doing their job efficiently. The concern is if acceptance rates become a substitute for image review. A department that monitors AI draft acceptance rates as a workflow metric without also monitoring radiologist image review time is measuring the wrong thing.
The Practical Summary for a Department Making Purchasing Decisions
Before signing with any radiology AI vendor, ask for clear answers to four questions. First, how does the FDA classify this product, and has the vendor obtained any clearance or authorization, or is it positioned as non-device CDS under the 21st Century Cures Act? Second, what is the specific intended use as described in any regulatory submission, and does it match your planned deployment? Third, how does the product's workflow design preserve independent radiologist review of the basis for AI outputs? Fourth, what data does the vendor have on the accuracy of AI outputs in the specific study types and clinical settings relevant to your department?
The answers will not all be clean. Some vendors have done the regulatory work and can answer precisely. Others are in earlier stages and will give you vaguer responses. The quality of the answer is as informative as the content: a vendor that has thought carefully about their regulatory positioning understands their product's limitations and is more likely to be accurate about what it can and cannot do in your environment.