You might think that an AI medical product will fail due to its model’s performance, but in reality, what often stalls a project is the one or two “seemingly reasonable” sentences describing its Intended Use.

In the language of the FDA, “Intended Use” is neither a marketing copy nor a vision statement; it is the starting point for classification logic.
Especially in the field of imaging AI, the reason CADe/CADx/CADt (Triage/Notification) are differentiated depends primarily on the “Intended Use”, the output type, and the corresponding regulatory classification and product code.

If you use the wrong verb, the common consequence isn’t just “a minor revision,” but rather: the entire product gets described by you into a different category, and your subsequent test strategy, clinical evidence, labeling, and predicate demonstration will all be completely derailed.

Below, using the reading rhythm of a Reviewer, we will dismantle the most common pitfalls and provide you with a practical set of correction templates.

I. First, let’s clarify three things: Where are the red lines between CADe, CADx, and CADt?

In the regulatory review of imaging AI, the most common “first principle” is actually quite simple: what you output implies which part of the decision-making process you are assisting the physician with.

CADe (detection) The core of CADe is “detecting/locating suspicious regions,” and the typical output is an ROI/marking (circling, bounding boxes, heatmap prompts), but it does not characterize benign/malignant features, nor does it provide a diagnostic conclusion.

CADx (diagnosis/characterization) The core of CADx is “characterization/diagnostic output,” such as malignancy probability, grading, severity, staging, etc. It is usually built upon ROIs or features, and its output type is much closer to clinical decision-making.

CADt (triage/notification) The core of CADt is “triage/prioritization.” The typical output is a case-level flag/priority score used to reorder worklists or notify users, and it typically does not perform lesion localization (does not localize) nor provide a diagnosis.

When mapping these back to regulations and product codes, you will find that the FDA’s classification language is actually very consistent:

• CADt: 21 CFR 892.2080, common product code QAS
• CADx (suspicious cancer lesions): 21 CFR 892.2060, common product code POK
• CADe/CADx combined: 21 CFR 892.2090 (detection + diagnosis), e.g., QDQ for breast-related
• There are also various CADe detection-oriented product codes (e.g., NWE/OEB/OMJ/MYN, etc.)

These are not just terms to memorize. The verbs and output descriptions you choose in your Intended Use are helping the Reviewer determine which bucket you fall into.

 

II. 5 questions that will pop up in the Reviewer’s mind while reading your “510(k) Chapter 1”:

1) Are you exactly CADe, CADx, or CADt? (The product code determines what evidence you must submit) This is the most common fatal point.
The FDA distinguishes between the three primarily based on the Intended Use and output type, not on what model you use.

• Typical Mistake: Claiming verbally to pursue CADt (triage), but writing “highlights bleeding points/marks lesion locations” in the text. Once you describe localization, the Reviewer will naturally push you towards CADe/CADe+ (or even combined) because the typical logic of CADt is a case-level flag, usually without localization or diagnosis.
• Correction Principles:
  • If you want to pursue CADt, narrow your output to “flag/priority/notify,” avoiding any implication that you are doing lesion-level marking.
  • If you want to do marking, honestly admit that you are CADe, and clearly state “no diagnosis, no triage”.
    

 

2) Trying to use “CDS” as a shield? Imaging AI is almost certainly a Device.

Many teams will say, “We are just CDS, not a medical device.” However, the reference documents you provided make it straightforward: The FDA’s 2022 final CDS guidance uses four criteria to determine non-device CDS, and if any single criterion is not met, it will be considered a device (usually pursuing Class II / 510(k)).

For imaging AI, the most fatal is criterion 1: as long as you acquire/process/analyze medical image or signal, you will almost certainly fail this criterion, thus most imaging AIs fall within device regulatory scope (e.g., QAS/POK/QDQ).

• Practical Tip: You can position your product as “assisting physicians in interpretation,” but do not imply in the Intended Use that it “replaces/makes conclusions for the physician.” What the FDA cares about is whether you are supporting HCP judgment or substituting it.

 

3) User profile written too broadly: Writing “available to GPs/Nurses” instantly escalates risk perception. When you mentioned “recent cases are restricted to trained professionals,” this direction is correct in the review context: when determining whether it supports HCP judgment, the FDA inherently considers the user (whether they are an HCP, whether they possess interpretation capabilities) as a core factor.

• Pragmatic Phrasing: Describe the user as “medical professionals with relevant interpretation training,” avoiding describing the product as “lowering the threshold so non-experts can make expert judgments.” The latter easily leads the Reviewer to associate it with misuse risks and automation bias.

 

4) Claimed scope vs. Data representativeness: If you write “all adult CTs,” prepare to be cross-examined.

Your attachment points out in the trend section: recent cleared indications often focus on detection/triage/segmentation/workflow support, rather than “standalone diagnosis.” The nuance behind this is: the broader the claim, the greater the burden of representativeness and validation.

Therefore, for the Intended Use, we recommend beautifully narrowing the scope using “clinical scenario + input type + output limitations,” rather than backing yourself into the hardest battleground with a single phrase like “applicable to all…”.

 

5) Accidentally writing “diagnostic outputs” in the text: Then you are no longer CADt. Many teams think, “We are just a reminder”; but once you write out diagnostic outputs like probability / malignancy likelihood / category / staging, the semantics shift closer to CADx (diagnostic or risk stratification outputs). This will directly change the reviewer’s intuitive expectations regarding risk and evidence.

 

III. A ready-to-use Intended Use writing framework (including Reviewer checkpoints)

The framework below isn’t “teaching you to write beautifully,” but rather “teaching you to write something classifiable, verifiable, and comparable.”

 

Step 1: Choose your core sentence pattern first (get the verb right)

• CADt (QAS) phrasing: identify suspected findings → notify/prioritize for earlier review (case-level flag / priority; typically no localization, no diagnosis)
• CADe phrasing: detect/highlight/mark suspicious regions (ROI/marking, but no diagnosis, no triage)
• CADx (POK) phrasing: characterize lesions/assess likelihood/categorization (diagnosis/characterization output) Once you get this verb level right, 80% of consistency issues will automatically disappear.
  

 

Step 2: Describe inputs/outputs as “testable” rather than “imaginable” Clearly state:

• What is the input image/signal (DICOM CT/MR/CXR…)
• Is the output a case-level flag/priority score, or a lesion-level ROI/marking, or a probability/category? Because one of the key factors the FDA uses to distinguish CAD/CADe/CADx is the “type of output.”

Step 3: State the “clinical workflow integration point” (which part are you helping with)

• CADt is typically used in the worklist/notification workflow, for re-order, alerts, and notifications.
• CADe is often described as a second reader to attract attention.
• CADx is closer to decision support after lesion identification. If you clarify the workflow, it’s easier for the Reviewer to place you in the correct risk framework, instead of using the most conservative approach to imagine you are “replacing interpretation.”
  Deliverables: What outputs must you prepare in your DHF at a minimum? The deliverables you originally listed are correct; we just renamed them into a more “DHF-aligned” version, making it easier for startup teams to adopt directly:
• DOC-001 Intended Use Statement (internal approval, version control)
• DOC-002 Indications for Use (consistent external version, used to align submission content)
• DOC-003 Clinical Workflow & Use Environment Map (Tag: triage/worklist, concurrent read, second read, etc.)
• MAT-001 Output Definition Matrix (Input → Output → Clinical Use → Validation Method; aligns “type of output” with intended use)
• TPL-001 Labeling/IFU Outline (explicitly states clinician oversight and limitations, avoiding being interpreted as autonomous diagnosis)

Bonus: The 3 most common pitfalls, and how to save them with one sentence?

Pitfall 1 (Aiming for CADt): Writing “can mark the location of brain hemorrhage”
The problem isn’t that you can’t do it, but that this sentence pushes you toward lesion-level marking; whereas the typical output of CADt is a case-level flag/priority, typically without localization or diagnosis.

• Corrected version (CADt phrasing): “This software is intended to identify imaging exams with suspected XXX findings and assist clinical personnel in prioritizing review via a notification/prioritization mechanism.”

Pitfall 2: Writing “Automatically diagnoses…” This makes the reviewer intuitively associate it with autonomous diagnosis. The reference documents you provided consistently emphasize the semantics of “aid” and “support” in distinguishing CADx/CADe, rather than “replace.”

• Corrected version (CADe phrasing): “This software is intended to highlight suspected lesion areas to aid trained professionals in detection and interpretation.”

Pitfall 3 (Segmentation/Measurement): Writing “for treatment decision-making” without stating the professional review responsibility
It’s not that this sentence can’t be used, but you need to clearly delineate responsibilities to prevent it from being understood as “the system directly makes treatment conclusions.” Changing it to “providing outputs that ‘require review/edit'” is generally safer and more in line with the semantic trends of recently cleared indications (emphasizing clinician-supervised uses).

 

Conclusion:  Intended Use is the starting point for classification, and the master switch for all your evidence In the world of imaging AI, you don’t build a product first and then look for regulations; you use the Intended Use first to “define the product as a classifiable and verifiable entity.”

The FDA’s focus on distinguishing CADe/CADx/CADt essentially comes down to the intended use and output type. If you wish, we can conduct a “Reviewer Simulation” of your Intended Use sentence by sentence:

• First, match your expected output (case-level flag vs ROI marking vs probability) to the most reasonable product code bucket (e.g., QAS/POK/QDQ, etc.).
• Then, adjust every verb and limitation clause into a phrasing that “does not alter functionality, but lowers the risk of misclassification.”

For more information, please feel free to email: Joseph.Shen@behealthventures.com