Company Spotlight: Cleerly LABS (v2.0) - AI for Coronary Plaque Analysis

Maya: The software might give the appearance of healthy tissue in situations where pathology that is near or smaller than the scanning resolution may be present. That is a direct warning from the FDA clearance document we are looking at today, regarding how AI handles tiny details on cardiac CT scans.

Sam: That is a massive clinical caveat for a cardiology tool. Welcome to AI in Medicine - Smart Summaries. Today we are unpacking the FDA 510(k) summary for Cleerly LABS (v2.0), published by the U.S. Food and Drug Administration on March 7, 2025.

Maya: Before we dive in, a quick reminder up front: our voices are AI-generated. This is a summary of a publicly available FDA clearance document, and we are not reading marketing material from the company.

Sam: Let us start with the basics of this clearance. What is Cleerly LABS (v2.0)? The document says the submitter is Cleerly, Inc., located at 1099 18th St Suite 2860, Denver CO 80202.

Maya: It is a web-based software application. It is designed to be used by trained medical professionals as an interactive tool for viewing and analyzing cardiac computed tomography, or CT, data.

Sam: And this is specifically for patients who underwent Coronary Computed Tomography Angiography, which the document abbreviates as CCTA, for the evaluation of CAD, or suspected CAD.

Maya: Exactly. Coronary artery disease. The indication for use says it helps determine the presence and extent of coronary plaques, which it specifically notes as atherosclerosis, and stenosis.

Sam: So we are looking at a tool to help doctors find blockages and plaque in the heart's arteries. AI is the big buzzword right now, but the FDA summary is actually very specific about where the AI is used in this software, and where it is not.

Maya: I really appreciate that distinction in the document. Section 4, the Device Description, states that the software uses both machine learning and simple rule-based mathematical calculation components on the backend.

Sam: Right, it says it applies deep learning methodology to identify high quality images, segment and label coronary arteries, and segment lumen and vessel walls.

Maya: So the AI is doing the tedious drawing and labeling. Segmenting a 3D vessel wall from a CT scan is incredibly time-consuming for a human. But here is the critical part: the actual measurements, the plaque, the stenosis, and the vessel measurements, those outputs are mathematical calculations. The document explicitly says they are not machine-learning based.

Sam: Wait, really? That is fascinating. The deep learning just preps the canvas. It identifies the anatomy, but the final numbers come from standard geometry or math, not a black-box AI guessing the stenosis percentage.

Maya: Yes, and the human is still very much in the loop. The software presents the 2D and 3D images to the user for review and manual editing.

Sam: What exactly can the user edit once the AI does its segmentation?

Maya: The summary lists several things. The user can edit the suggested segmentation. They can adjust plaque thresholds, and they can demarcate stenosis, stents, and chronic total occlusions, which are referred to as CTOs. They can also select dominance and indicate coronary anomalies.

Sam: So the final output is a hybrid. It says it is based on the combination of user-editable segmentation and user-placed stenosis, stent, and CTO markers.

Maya: Yes. The document is very clear that this software does not perform any functions that could not be accomplished by a trained user with manual tracing methods or other commercially available software.

Sam: It literally says that? Then what is the point of a hospital buying it if a human could just do it manually?

Maya: Efficiency. The very next sentence says it represents a more robust semiautomatic software intended to enhance the performance of time-intensive, potentially error-prone, manual tasks. It is all about improving efficiency for medical professionals in the assessment of coronary artery disease.

Sam: That makes sense for the boardroom. Once the user finishes editing, the software creates something called the CORONARY Report.

Maya: Which uses the data to generate a visually interactive and comprehensive report that details the atherosclerosis and stenosis findings of the patient.

Sam: But there is a clear warning here. It says, This report is not intended to be the final report, i.e., physician report, used in patient diagnosis and treatment.

Maya: Right. It provides the ability to send the text report page to the user's PACS system, which is the hospital image archiving system, but it is an aid, not the final diagnostic word.

Sam: This brings us back to that warning you mentioned at the very beginning of the episode. The one about interpolated data.

Maya: Yes, in the Indications for Use section, the FDA includes a specific warning. It says, Users should be aware that certain views make use of interpolated data. This is data that is created by the software based on the original data set.

Sam: And the consequence of that is stark. Interpolated data may give the appearance of healthy tissue in situations where pathology that is near or smaller than the scanning resolution may be present.

Maya: If a piece of plaque is smaller than the resolution of the CT scanner, the software's interpolation might smooth it over, making the vessel look perfectly healthy. The document does not offer any medical explanation on how to spot that, or how often it happens. It just states the limitation and says the software should only be used by people appropriately trained in its functions, capabilities and limitations.

Sam: Let us talk about the testing. How did they prove this works to the FDA? What is the real numbers behind the clearance?

Maya: This is a 510(k) clearance, submission number K242338. That means they don't necessarily need a massive new clinical trial if they can prove it is substantially equivalent to an already cleared device on the market.

Sam: The predicate device. And in this case, the predicate device is Cleerly LABS v2.0, cleared under submission number K202280.

Maya: Yes, it is effectively an update to their own existing product. The new submission has the exact same trade name, Cleerly LABS (v2.0).

Sam: What did they actually change if it has the exact same name?

Maya: Section 6 breaks this down. The subject device has an identical intended use, similar operating principles, and technological characteristics. The modifications are to the product labeling, changes to the workflow, and minor incremental technological enhancements/changes.

Sam: Did they change the AI model?

Maya: No. They explicitly state, The modifications do not introduce any changes to the previously cleared existing algorithms or simple mathematical calculations.

Sam: They include a detailed table comparing the features of the new submission versus the old one. Let's look at that. Most of it is exactly the same. Both are Client-Server Google Chrome applications. Both take DICOM 3.0 Compliant or higher CT Images.

Maya: Both have 2D and 3D imaging, Multiplanar Reformat, which is MPR, Segmentation of Region of Interest, Plaque Composition Overlay, Hounsfield Unit, or HU, Distance and Volumetric Measurements, Remodeling Index, Stenosis, and Coronary Anatomical Findings.

Sam: So what is actually different in that table?

Maya: 2 things stand out. First, the product codes. The predicate only had the product code LLZ. The new device has QIH as the primary product code, and LLZ as the subsequent product code.

Sam: Why add QIH?

Maya: The document explains this is an administrative change. QIH is a classification for medical image management and processing systems that use artificial intelligence or machine learning. The document says this change was introduced because the core functions are supported by artificial intelligence, nonadaptive machine learning algorithms for CCTA processing and analysis.

Sam: Got it. So the FDA is just updating the regulatory bucket it sits in. What was the second difference in the table?

Maya: At the very bottom of the table, a feature called Cleerly Core Lab Image Analysis. For the new device, this is listed as Yes. For the predicate device, it says Not defined.

Sam: The document does not really elaborate on what the Core Lab Image Analysis feature actually does, does it?

Maya: No, it just lists it in the table as a new feature. We don't get any technical explanation of it in this summary, and we have to stop there.

Sam: Alright, so if they didn't change the algorithm, what kind of performance testing did they have to submit to the FDA for this new clearance?

Maya: Section 8 is very brief. Under Clinical Data, it says, No clinical testing was conducted to demonstrate safety or effectiveness as the device's non-clinical testing was sufficient to support the intended use of the device.

Sam: Wow. Just 1 sentence. So absolutely no new patient data was analyzed to get this clearance?

Maya: Correct. Instead, they relied on software evaluation activities. In Section 7, Performance Data, it says they tested in multiple pre-production environments using simulated data and in production for release verification.

Sam: Simulated data. So they ran test data through the system to make sure the software did not crash or output math errors.

Maya: Yes, they call it regression testing, which was completed to confirm that the changes in technological characteristics did not affect the overall performance or functionality of the device.

Sam: The document lists several technical standards they met. ISO 14971, 2019-12 for medical device risk management. AAMI TIR 57:2016 for medical device security risk management.

Maya: And ANSI AAMI IEC 62304:2005/A1:2016 for software life cycle processes. They also comply with DICOM, which stands for Digital Imaging and Communications in Medicine.

Sam: That is a lot of acronyms for software testing. But essentially, it proved the software still does exactly what the older version did.

Maya: Exactly. The conclusion section states that the software continues to perform as intended and is as safe and effective as the predicate device.

Sam: There is some standard FDA boilerplate at the beginning of the document too. The usual reminders about quality systems, good manufacturing practices, and the Unique Device Identification System rule.

Maya: Yes, it specifically mentions 21 CFR Part 820 for Quality Systems, including 21 CFR 820.30 for Design controls, 21 CFR 820.90 for Nonconforming product, and 21 CFR 820.100 for Corrective and preventive action.

Sam: The timeline of the submission is interesting. The 510(k) summary is dated August 7, 2024. It was received by the FDA on February 10, 2025, and the clearance letter is dated March 7, 2025.

Maya: And the summary document itself says it is provided per the requirements of section 21 CFR 807.92 on March 5th, 2025. The letter is signed for Jessica Lamb, Ph.D., Assistant Director, Imaging Software Team.

Sam: Oh, and I have to mention my favorite part of any government document: the Paperwork Reduction Act statement.

Maya: I knew you were going to bring that up.

Sam: It says the burden time for this collection of information is estimated to average 79 hours per response. 79 hours to review instructions, gather data, and complete the form.

Maya: Regulatory compliance is incredibly time consuming.

Sam: Let us step back. For a clinician on the ward, or a hospital IT director looking to buy this, what is the main takeaway from this document?

Maya: For the clinician, it is that this tool is designed to save you time. It uses deep learning to do the heavy lifting of tracing and segmenting coronary arteries, but you remain the ultimate diagnostic authority. The AI does not calculate the stenosis, the math does, based on the markers you approve. And you absolutely must remember that interpolated data could mask tiny pathologies.

Sam: And for the boardroom? The health system leaders buying it?

Maya: It highlights how an AI company maintains its regulatory footprint. By securing this 510(k) with an updated QIH product code, they are staying current with the FDA's AI categorizations, even without changing their core algorithm. It shows a software life cycle heavily reliant on standard regression testing and simulated data, rather than needing fresh clinical trials for every software update.

Sam: It really demystifies how software gets updated in the medical device world. You fix the workflow, you update the label, you run simulated data, and you prove substantial equivalence to a device marketed prior to May 28, 1976, or one reclassified since.

Maya: Exactly. It is a pragmatic look at the lifecycle of a medical AI product.

Sam: Alright, let's recap. We just reviewed the FDA 510(k) summary for Cleerly LABS (v2.0). It is a web-based, AI-powered post-processing tool for cardiac CT scans.

Maya: It uses deep learning to segment vessels and identify images, but relies on standard math for the actual measurements of plaque and stenosis.

Sam: It was cleared without new clinical testing, based on substantial equivalence to an earlier version of itself, relying instead on simulated data and regression testing.

Maya: And crucially, it carries a warning that interpolated data might make diseased tissue look healthy if the pathology is near or smaller than the scanner's resolution.

Sam: As always, a link to the full FDA document is in the show notes. Please read it for yourself.

Maya: And remember, this podcast is for informational purposes only. We are AI-generated voices, and this is not medical advice. Thanks for listening.