Company Spotlight: Keya Medical DEEPVESSEL Plaque - Automated Coronary Plaque Phenotyping

Maya: What if a cloud-based AI could process a routine coronary CT scan and output a total plaque volume that correlates with three expert cardiologists at 0.976?

Sam: That is the exact core claim in the FDA 510(k) summary for DEEPVESSEL Plaque by Keya Medical Technology. The summary was published by the U.S. Food and Drug Administration, and we are breaking it down today.

Maya: Before we begin, a quick reminder. Our voices are AI-generated, and this is a summary of a public FDA document, not company marketing material. We have no ties to the company, and this is an independent look at exactly what is, and isn't, in the clearance.

Sam: Okay, so DEEPVESSEL Plaque. Let's start with the basics. What exactly is this product, and who is it intended for?

Maya: The document defines DEEPVESSEL Plaque as a web-based image analysis application. It is meant to analyze Computed Tomography Angiography, or CTA, DICOM data. As for the patient population, it is indicated for patients aged 22 and above with suspected or known coronary artery disease.

Sam: And the main goal here is quantifying plaque and stenosis. The summary says it provides the presence and extent of coronary plaques and stenosis for patients who undergo CCTA for evaluation.

Maya: Exactly. And it does this by applying deep learning neural networks to segment the coronary arteries, the lumen, and the vessel walls. Once it does that, it outputs two main visual reports to the clinician.

Sam: Right, I see those listed here. The Cardisight Overview Report and the Plaque Analysis Report. The Cardisight one gives a 2D visual overview of the location of the plaque and the CAD-RADS stenosis category.

Maya: And the Plaque Analysis Report goes heavy into the quantitative characteristics. It outputs the total plaque volume, calcified plaque volume, non-calcified plaque volume, low-density plaque volume, and the plaque burden. The document defines plaque burden as percent plaque volume, which is calculated as total plaque volume divided by vessel volume.

Sam: It gives these numbers on both the main vessel branch level and the patient level, right?

Maya: Yes, it does.

Sam: This brings us to a really fascinating part of the workflow. When you hear about the application of deep learning neural networks in software, you might picture an AI just spitting out a final report completely autonomously. But that is not what happens here.

Maya: No, it is not. The document explicitly states that the AI segmentations will be reviewed and edited by trained and qualified analysts, if needed.

Sam: That is a huge operational distinction. There is a human in the loop, but it is not the clinician at the hospital doing the primary editing. The document says these trained operators are located in regional data centers off-site.

Maya: This is a major point for health system leaders to understand. The location of the DICOM image data processing is entirely different from the predicate device it was compared against.

Sam: Let's talk about that predicate device. It was compared to Autoplaque 3.0, under submission K212758, which was submitted by Cedars-Sinai Medical Center.

Maya: Right. Autoplaque 3.0 is a workstation-based desktop device used at clinical sites. It receives DICOM image data by loading it directly from the local computer hard drive. DEEPVESSEL Plaque, on the other hand, is a client-server Google Chrome application. It receives the images via transmittal by a secure picture archiving and communication system, or PACS.

Sam: So with the predicate, the software and the trained operator are sitting right there at the clinical site. With DEEPVESSEL Plaque, the software is web-based, and the operator can be at a remote site.

Maya: The document points out how this changes the workflow for the end user, the clinician. With Autoplaque, if the user wants to revise the analysis, they can directly repeat or modify it on-site. With DEEPVESSEL Plaque, the user can request the remote operator to perform a re-analysis.

Sam: That is essentially AI-as-a-service. You send the scans out, the remote team and their AI process it, and they send the report back. The FDA summary notes that for both devices, the final analysis results are provided to a healthcare professional at a clinical location, so the indicated recipients are the same.

Maya: And the FDA determined these minor technological differences do not raise different questions of safety or effectiveness. Both devices are indicated as post-processing tools for viewing and analyzing cardiac CT data, and both are not intended to replace the skill and judgement of a qualified medical practitioner.

Sam: Looking deeper into the technology comparison in Table 1, there are a few features the predicate has that DEEPVESSEL Plaque does not.

Maya: Yes. The summary lists several quantitative measurements available in Autoplaque 3.0 but not in DEEPVESSEL Plaque. For example, the predicate measures Remodeling index, Area stenosis, Plaque length, and Minimum luminal area over the lesion.

Sam: The predicate also outputs Plaque composition percentages, like calcified plaque volume divided by total plaque volume. DEEPVESSEL Plaque doesn't output those specific ratios, though as we mentioned, it does output the raw volumes for those plaque types.

Maya: On the visualization side, the predicate uses Maximum intensity projection, or MIP, and Multiplanar reformatting with oblique slicing. DEEPVESSEL Plaque doesn't use those, but it does add Volume rendering based on centerlines and Ideal Model Adjustment, which the predicate lacks. The FDA noted these differences do not impact product efficacy or safety.

Sam: Okay, so how did they actually prove the safety and effectiveness? Let's get into the performance testing.

Maya: They ran a retrospective, multi-center clinical validation study to evaluate how well the software detects, quantifies, and characterizes the plaque and stenosis. They used 147 CCTA datasets collected from three independent U.S. clinical sites.

Sam: And they stress that this validation dataset was completely independent from the development dataset.

Maya: That is a critical detail for AI validation. The development data was made up of CCTA images from 1,200 patients acquired from five medical centers in China and one medical center in Europe. The pivotal validation dataset was kept entirely separate, using only those 147 cases from the three U.S. sites.

Sam: Let's look at the demographics of those 147 patients. The ages ranged from 22 to 88 years old. There were 91 men and 56 women.

Maya: For ethnicity, the breakdown is 72 White patients, which is 49.0%. There were 35 Hispanic patients, making up 23.8%, and 22 Asian patients, at 15.0%.

Sam: But there were only 4 Black patients in the entire study, which is 2.7%. There was 1 patient listed as others, which is 0.7%, and 13 patients, or 8.8%, were unknown.

Maya: We have to flag that 2.7% representation for Black patients. Having only 4 Black patients out of 147 is a limitation in the validation dataset's diversity. The summary does not provide a demographic breakdown for the 1,200 patients in the training set, so we only have these numbers for the validation cohort.

Sam: In terms of disease severity, the cohort was pretty broad. The patient-level CAD-RADS categories spanned from CAD-RADS 0 to 5.

Maya: Right, and 61 of those subjects, or 41.5%, had clinically significant stenosis, defined as a CAD-RADS of 3 or higher. The dataset also included all major plaque subtypes. 130 subjects, or 88.4%, had calcified plaque. 134 subjects, or 91.2%, had non-calcified plaque, and 133 subjects, or 90.5%, had low-density plaque.

Sam: So they had the scans, they ran them through DEEPVESSEL Plaque. What were they comparing the AI's results against?

Maya: The reference standard was established by a consensus of three expert readers. Two experts independently reviewed each case and provided measurements for the volumes and stenosis categories. Then, a third senior expert reader adjudicated any discrepancies to establish the final consensus truth.

Sam: And the vessel-level analyses were performed on a lot of anatomy. The document says they evaluated 441 coronary vessel territories.

Maya: Yes, specifically consisting of the left main plus left anterior descending artery, the left circumflex artery, and the right coronary artery for each subject. They evaluated 711 DVPlaque-detected plaques against the expert-reader ground truth.

Sam: And they made sure the software analysts and the expert readers were blinded to each other's results during the validation.

Maya: Exactly. So, what were the actual results? Let's look at Table 3, which lists the performance metrics on a per-patient basis.

Sam: For Total Plaque Volume, the Pearson Correlation Coefficient was 0.976.

Maya: At 0.976, that shows incredibly strong agreement between the software and the expert consensus.

Sam: It gets even higher for Calcified Plaque Volume, with a correlation of 0.994. Non-calcified Plaque Volume was 0.945, Vessel Volume was 0.985, and Plaque Burden was 0.978.

Maya: They also looked at the CAD-RADS Stenosis Category and used a Weighted Kappa statistic. The result was 0.810, which passed their acceptance criteria.

Sam: For the actual segmentations, drawing the lines around the anatomy, they used the Dice coefficient. Vessel Segmentation was 0.888, and Lumen Segmentation was 0.887.

Maya: The summary notes that subgroup analyses showed consistent performance across coronary vessel territories, stenosis severity categories, scanner manufacturers, and imaging parameters.

Sam: Speaking of scanner manufacturers, the CCTA studies were acquired using 64-detector row CT scanners or higher. The brands represented included GE, Toshiba, Arineta, and Siemens.

Maya: Those same scanner brands were also used in the Repeatability and Reproducibility evaluations.

Sam: Let's touch on that R&R testing. They used a smaller group for this, right?

Maya: Yes, a total of 23 subjects with a mean age of 62.2 years. They wanted to evaluate the variation of repeated analyses when using different image analysts, which tests reproducibility, and on different days, which tests repeatability. The cases spanned stenosis grades 0 through 4.

Sam: The summary mentions they used a washout-period between the days to avoid memory effects during the testing.

Maya: That makes sense. You don't want the human analyst just remembering how they edited the scan the previous time. The acceptance criterion was a percent coefficient of variation of 5.0% or less.

Sam: And it passed across the board. The percent CV due to the day and the percent CV due to the analyst were both equal to or less than 5.0% for both Total Plaque Volume and Vessel Volume.

Maya: While the performance numbers are strong, there is a specific warning in the Indications for Use that clinicians need to pay attention to.

Sam: Right. The label states that users should be aware that certain views make use of interpolated data. This is data that is created by software based on the original data set.

Maya: The warning goes on to explain why that matters. It says that 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.

Sam: So if a piece of pathology is tiny, near or below the resolution limit of the CT scanner, the software's interpolation might smooth it over and make it look like healthy tissue.

Maya: Exactly. This is why the document explicitly states that the software is not intended to replace the skill and judgement of a qualified medical practitioner and that results should be used with other clinical info.

Sam: It's a tool to support the clinician, not a final diagnostic oracle. And it should only be used by people who have been appropriately trained to use it.

Maya: Taking a step back, this clearance highlights a fascinating trend. Clinicians want comprehensive, volumetric measurements of all the plaque in the coronary arteries, not just a simple diameter reading.

Sam: But doing that manually takes far too much time. So Keya Medical is offering a solution that combines deep learning with remote human analysts to do the heavy lifting.

Maya: It removes the processing burden from the local hospital's workflow, but it does mean adopting a web-based, remote-analysis model instead of keeping everything on an on-site workstation.

Sam: To recap, we've been looking at the FDA 510(k) summary for DEEPVESSEL Plaque by Keya Medical Technology. It's a cloud-based software that uses deep learning and trained remote analysts to quantify coronary plaque and stenosis from CCTA scans.

Maya: In a 147-patient retrospective validation study, it showed a 0.976 correlation with expert readers for total plaque volume, though we did note the limited representation of Black patients in the validation cohort.

Sam: As always, a link to the full FDA summary is in the show notes. This podcast is for informational purposes only and is not medical advice. Thanks for listening.