Company Spotlight: Circle CVI cvi42 Coronary Plaque - Automating AI Plaque Quantification

Maya: We are looking at a machine learning algorithm that can quantify calcified coronary plaque volume with a 0.99 correlation to expert human annotators.

Sam: Today we are unpacking the FDA 510(k) summary for the cvi42 Coronary Plaque Software Application, a document published by the U.S. Food and Drug Administration on October 27, 2025.

Maya: Before we dive in, a quick reminder that our voices are AI-generated, and this episode is a summary of a publicly available FDA clearance document, not company marketing material.

Sam: And this is a Company Spotlight episode. We are looking strictly at what is inside this 510(k) summary from Circle Cardiovascular Imaging Inc., based in Calgary, Canada.

Maya: Exactly. Let us start with what this product actually is. The document describes it as a Software as a Medical Device, or SaMD. It is designed to enable the analysis of CT Angiography scans of the coronary arteries.

Sam: Right, so this is for patients who underwent Coronary Computed Tomography Angiography, which the document abbreviates as CCTA, for the evaluation of coronary artery disease, or suspected CAD.

Maya: And the core function here is helping physicians visualize, evaluate, and analyze coronary vessel plaques. The summary states that it allows for manual or semi-automatic segmentation of the vessel lumen and the vessel wall.

Sam: That segmentation is the crucial part. It determines the presence and extent of both the plaques and the luminal stenoses. But how does it actually do that?

Maya: The document says the software implements an Artificial Intelligence and Machine Learning algorithm. It detects the lumen and vessel wall structures, and then uses post-processing methods to convert those structures into editable surfaces.

Sam: Editable being a key word there.

Maya: Very key. The document explicitly notes that all surfaces generated by the system are editable and users are advised to verify and correct any errors.

Sam: So it is not a black box that just spits out a final diagnosis. The summary says it is intended to be used as an aid to the existing standard of care, and it does not replace existing software applications that physicians use.

Maya: Right. It is an application module that can be integrated into an image viewing software, like Circle's already cleared cvi42 application. It does not interface directly with the scanner itself, and it is independent of the type of vendor acquisition equipment.

Sam: Which is a big deal for health systems. You do not want a software that only works with one specific brand of CT scanner.

Maya: Exactly. From a clinical workflow perspective, the analysis results are available on-screen, and they can be sent to a report or saved for future review.

Sam: Let us get into the exact measurements this software provides, because the summary includes a detailed table. What exactly is the AI measuring?

Maya: Table 1 in the document lists specific measurements. First, we have Calcified Plaque Volume in cubic millimeters, which is the volume of calcified plaque in a lesion or across all lesions in a vessel.

Sam: And I see it also measures Non-Calcified Plaque Volume and Low-Attenuation Plaque Volume, both also in cubic millimeters.

Maya: Yes, and it is interesting to note the document specifies that Low-Attenuation Plaque volume is also included within the Non-Calcified Plaque volume.

Sam: Then it combines them for a Total Plaque Volume. But it does not just give you the raw volume. It also calculates the burden as a percentage for all of those categories.

Maya: Correct. The burden is obtained by dividing the plaque volume by either the lesion vessel wall volume or the entire vessel volume, depending on what you are looking at.

Sam: There is another interesting measurement here called the Remodeling Index. The document defines this as a measure of enlargement of vessel dimensions to accommodate plaque development compared to normal vessel sections.

Maya: It says it is expressed as the ratio of the largest diameter across the lesion and the reference diameter, or the maximum vessel wall cross-sectional area compared to the reference area.

Sam: And then, of course, the classic stenosis measurements. It gives you both Diameter-Based Stenosis and Area-Based Stenosis as percentages, plus the overall Lesion Length in millimeters.

Maya: For the clinical and business leaders listening, this is what you are actually buying or prescribing: a highly structured, quantitative breakdown of a patient's coronary plaque, broken down by composition.

Sam: So how do we know the software actually gets those measurements right? Let us talk about the performance testing.

Maya: This is a 510(k) clearance, so we need to remember the standard here is substantial equivalence to legally marketed predicate devices. The document states that no clinical studies were necessary to support substantial equivalence.

Sam: Wait, really? No clinical trial data at all?

Maya: That is completely standard for this class of device. They did bench testing and validation per FDA guidance. The summary outlines validation for both the non-machine learning mathematical outputs and the machine learning derived outputs.

Sam: Let us dig into the machine learning validation, because that is where the real numbers are. How did they train and test this AI?

Maya: The algorithms were trained and tested on images acquired from major vendors of CT imaging devices. Importantly, all data used for validation were kept separate and were not used during the development of the algorithms.

Sam: And where did that data come from? The summary says the validation data was sourced from multiple sites, with 100% of the data sampled from US sources.

Maya: But notice what is not in the document. It does not tell us the raw number of scans, the number of patients, or the demographic breakdown of that US data. It only says the information was anonymized.

Sam: That is a limitation of what we can glean from this summary. We know the location, but not the scale of the dataset.

Maya: What we do have are the specific performance results. They evaluated the algorithm against a reference standard established from expert annotators.

Sam: Let us hear the actual numbers. They used the Dice Similarity Coefficient, or DSC, and the Hausdorff Distance, or HD.

Maya: Right. For the lumen, the Mean DSC was 0.76. For the wall, the Mean DSC was 0.80. Both of those passed their pre-defined acceptance criteria.

Sam: And the Hausdorff Distance, which measures the maximum distance between the predicted boundary and the expert's boundary, was a Mean HD of 0.77 mm for the lumen, and 0.87 mm for the wall.

Maya: Then they looked at how well the derived plaque volumes correlated with the experts across different plaque compositions, using the Pearson Correlation Coefficient, or PCC.

Sam: This is where we get the number from our intro. The Calcified Plaque PCC was 0.99. That is incredibly high.

Maya: It is. Total Plaque PCC was 0.97. Non-Calcified Plaque PCC was 0.93. But then there is a drop.

Sam: Right, the Low-Attenuation Plaque PCC was 0.74. Still listed as a pass, but definitely lower than the others.

Maya: The document does not explain why the correlation for low-attenuation plaque is 0.74 compared to 0.99 for calcified plaque. It just presents the results. As always, we will not guess at the technical or medical reasons why that might be.

Sam: Fair enough. The summary also notes that the machine learning algorithms demonstrated consistent and comparable performance across diverse subgroups.

Maya: Yes, it lists variations in scanner vendor, tube current, pixel resolution, slice thickness, patient age, sex, vessel location, and plaque lesion volume.

Sam: So, to get this clearance, they had to prove it was substantially equivalent to devices already on the market. Let us look at those predicate devices.

Maya: The primary predicate is actually their own previous software: the cvi42 Auto Imaging Software Application, cleared under K213998.

Sam: But there is a catch. Table 3 shows that the primary predicate did not offer segmentation of the lumen and vessel wall structures. It also did not have plaque thresholds or the Remodeling Index.

Maya: Exactly. The document says that to demonstrate that the addition of semi-automatic segmentation does not raise new questions of safety, they used a secondary predicate: Cleerly Labs, cleared under K190868.

Sam: Ah, so Cleerly Labs had the AI functionality for plaque thresholds and the remodeling index, which allows Circle to say their new module is equivalent to a mix of their old software and Cleerly's software.

Maya: That is a very common regulatory pathway. But the tables highlight a few key differences. For instance, the operating system. The new cvi42 Coronary Plaque runs on macOS and Microsoft Windows.

Sam: While the secondary predicate, Cleerly Labs, is described as a Client-Server application running on Google Chrome.

Maya: Which matters for the IT folks in a health system. Do you want software running locally on your own operating systems, or a web-based client-server setup?

Sam: There is another fascinating difference in the Indications for Use comparison in Table 2. The Cleerly Labs indication includes a specific warning about interpolated data.

Maya: Yes, the Cleerly Labs text states that users should be aware that certain views make use of interpolated data created by the software, which may give the appearance of healthy tissue where pathology might be present.

Sam: But that specific warning about interpolated data does not appear in the Indications for Use for this new cvi42 Coronary Plaque software. The summary does not elaborate on why.

Maya: It does, however, emphasize that the new software shall be used by qualified medical professionals, experienced in examining cardiovascular CT images, for the purpose of obtaining diagnostic information as part of a comprehensive diagnostic decision-making process.

Sam: And it specifies that the semi-automated machine learning algorithms are intended for an adult population.

Maya: Let us talk briefly about cybersecurity, because the FDA is paying much more attention to this now, and it gets its own section in the summary.

Sam: Right, the document states it meets cybersecurity requirements through threat modeling, software composition analysis, static analysis, and vulnerability testing.

Maya: They also conducted penetration testing to ensure there were no unidentified vulnerabilities and that risk control measures were in place. Plus, they maintain a cybersecurity risk management plan for post-market monitoring.

Sam: It is good to see that laid out so clearly in a public document.

Maya: So, looking at the big picture, what does this change for the clinical or business leaders?

Sam: For hospitals and imaging centers, it means there is another cleared option for quantifying coronary plaque on CT scans. And because it integrates into an existing DICOM viewer, it might fit smoothly into workflows that already use Circle's software.

Maya: But again, the summary reminds us that FDA clearance is based on substantial equivalence, not a determination that it complies with all other federal statutes. And the system requires the physician to review and edit those AI-generated boundaries.

Sam: My memorable takeaway is that correlation of 0.99 for calcified plaque compared to human experts. That is a striking number, even with the lower 0.74 correlation for low-attenuation plaque.

Maya: It certainly shows where the technology excels and where the performance metrics differ by plaque type. To recap, we have been discussing the FDA 510(k) summary for the cvi42 Coronary Plaque Software Application, which clears a new AI module for quantifying coronary plaque and stenosis.

Sam: You can find a link to the full FDA summary document in our show notes. Remember, this podcast is for informational purposes only and is not medical advice.

Maya: Thank you for listening.