Unpacking the FDA Clearance of the AIRA Kidney Algorithm

Sam: Did you know a new AI tool just got cleared to automatically calculate kidney volume with a mean Dice similarity coefficient of 0.9114 on the left kidney?

Maya: That is an incredibly precise metric for automatic segmentation. Today, we are looking at the official U.S. Food and Drug Administration clearance letter for the Artificial Intelligence Renal Assessment AIRA Version 1.

Sam: And just a quick reminder up front, our voices are AI generated, and this episode is a summary of a publicly available document.

Maya: Exactly. This document is a premarket notification summary. It outlines exactly how a company called Aramis Global AI Consultancies LLC brought their imaging software to market and proved to the FDA that it is safe and effective.

Sam: The document itself is dated September 15, 2026. The specific submission number is K261677. Let us start with the absolute basics. What exactly is the AIRA tool?

Maya: According to the FDA classification, it is a medical image management and processing system. The specific regulation number governing this is 21 CFR 892.2050. It falls into Regulatory Class II, with the product code QIH.

Sam: Class II means it requires special controls, which we will get into later. But practically speaking, what does this software actually do when a doctor opens it up in the hospital?

Maya: AIRA operates by taking medical imaging data, specifically CT scans, to analyze the size and shape of the kidney. It utilizes a deep learning algorithm to accurately estimate the renal parenchymal volume of the kidney.

Sam: So it measures the actual tissue volume of the organ. I noticed the document gets surprisingly specific about the software architecture. It says the output of the algorithm is a NumPy array.

Maya: Yes, that is a level of technical detail you do not always see in high level regulatory summaries. It takes that NumPy array and uses it to generate a NIfTI file format. From there, the system uses that file to calculate the kidney volumes.

Sam: And it does not just leave those numbers floating in the software. Those values are compiled into a PDF report that can be easily downloaded from the AIRA tool.

Maya: Right. That PDF report provides the user with detailed information about the kidney size of the patient. Specifically, the report includes the separate volumes of the left and right kidneys. But they also get a visual output to verify the math.

Sam: The text mentions that volume masks are produced. Those masks can be superimposed directly on the original DICOM gray scale image. That way, the doctor isn't just trusting a black box; they can visually verify the boundaries the AI drew.

Maya: And there is a crucial fail-safe built into the workflow. An editor tool is also available. If the AI gets the boundaries slightly wrong, the user can use this editor to modify the outputs where necessary and recalculate the volumes.

Sam: Though the text notes this editing step is optional. The intended users for all of this are medical professionals. It specifically lists radiologists, nephrologists, and urologists who are trained in the interpretation of medical imaging data.

Maya: Exactly. And why do these specific specialists need such precise volume measurements? According to the submission, this volume information is compared to baseline measurements to detect any changes that may indicate early kidney swelling or volume loss.

Sam: Catching volume loss early could be absolutely vital for patient outcomes. The document also specifies the target demographic. AIRA is intended for adult patients who require volumetric measurement of the kidneys for further clinical assessment.

Maya: Let us look at how Aramis proved to the FDA that this software actually works. They conducted a clinical validation called the AIRA Standalone Study. This was a retrospective study conducted to assess the performance of the algorithm.

Sam: They collected a total of 119 computed tomography cases retrospectively. These were collected under the approved Aramis Protocol 66017 Data Collection Protocol. And they gathered these cases from 3 US sites at distinct geographic locations.

Maya: Having 3 distinct geographic sites is very important for regulatory clearance. It helps prove the AI works on different types of CT scanners and across diverse patient demographics. Out of those 119 cases originally collected, the final study analyses included 113 cases.

Sam: The document does not explicitly explain why those six cases were excluded from the final analysis, but what really caught my eye was how they structured the data. The sample was enriched to have at least 25% of cases with atrophied kidneys.

Maya: That enrichment strategy is a very deliberate and smart move. Algorithms often perform flawlessly on perfectly healthy, normal anatomy. But in a real hospital, you are dealing with diseased, shrunken, or abnormal organs.

Sam: By intentionally requiring at least 25% of the data to feature atrophied kidneys, they proved the tool works exactly when the clinical team needs it most. To judge the AI, they needed a gold standard.

Maya: They used 2 radiologists to establish the ground truth volume and segmentation. And impressively, the document notes that no adjudication was required between them.

Sam: That means the 2 radiologists agreed closely enough on the boundaries of the kidneys that they never had to call in a tie-breaker. So they took this human ground truth and compared it to the software.

Maya: Specifically, AIRA Version 1.0 was used to process all cases and provide segmentation and renal parenchymal volume for each kidney. The accuracy was measured by two primary metrics: the Dice similarity coefficient and the Hausdorff distance.

Sam: Let us unpack those numbers, because this is what matters to anyone evaluating medical AI. The Dice similarity coefficient measures spatial overlap. The results were very strong.

Maya: The mean Dice similarity coefficient between the AIRA mask and the larger ground truth mask was 0.9092 for the right kidney. The 95% confidence interval for that was 0.9032 to 0.9148.

Sam: And it was even slightly better on the other side. The mean Dice similarity coefficient was 0.9114 for the left kidney, with a 95% confidence interval from 0.9035 to 0.9179.

Maya: Consistently seeing those high overlap scores is excellent for automatic segmentation. What about the Hausdorff distance? While Dice measures overall volume overlap, Hausdorff measures the maximum distance between the AI boundary line and the human boundary line.

Sam: The mean Hausdorff distance was 9.9743 mm for the right kidney. The confidence interval there was 8.4073 to 12.5967. For the left kidney, it was 9.9395 mm, with an interval of 7.8926 to 13.6073.

Maya: A maximum error of less than ten millimeters at the most extreme points is quite small when you consider the overall size of an adult kidney. But they also provided the mean 95th percentile Hausdorff distances.

Sam: Right, that filters out the absolute most extreme outliers. For the right kidney, the mean 95th percentile Hausdorff distance was 3.7734 mm, with a confidence interval of 3.5544 to 4.0104.

Maya: And for the left kidney, it was 3.9151 mm, with an interval from 3.4522 to 4.6661. Those numbers prove the precision is clinically viable.

Sam: To get this clearance, Aramis did not just show their own data. They had to prove their device is substantially equivalent to a legally marketed predicate device. They chose a device called Ceevra Reveal 3.

Maya: Yes, Ceevra Reveal 3 was manufactured by Ceevra Inc. It was cleared under submission K222676, and its decision date was 04/25/2023. The document includes a substantial equivalence table detailing the similarities and differences.

Sam: Both devices are classified under 21 CFR 892.2050 as a Medical image management and processing system. Both are Class II. Both use Web browser Google Chrome, and both target the adult patient population.

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: But there are some major differences in their scope. Ceevra Reveal 3 was cleared for analyzing a lot more anatomy. Its anatomical sites included the prostate, bladder, neurovascular bundles, kidney, vein, and arteries.

Sam: Whereas AIRA is strictly focused on the kidney. Ceevra also handled both CT or MR images, while AIRA utilizes CT images only. Both are Software as a Medical Device, meaning they do not have physical hardware components.

Maya: This is a classic regulatory strategy. By narrowing the focus to just CT scans and just the kidney, Aramis likely made their verification and validation testing much more straightforward than if they had tried to match the entire feature set of the predicate device.

Sam: And when you look at the performance testing copied directly from the 510(k) summary of K222676, the narrower focus seems to have paid off. Ceevra reported a 0.89 DSC for the kidney using CT abdomen imaging, and 0.87 DSC using MR abdomen imaging.

Maya: Exactly. With scores of 0.9092 and 0.9114, AIRA performs the specific task of CT kidney segmentation as well as the predicate device. Therefore, the FDA concluded it is substantially equivalent.

Sam: The FDA letter itself is fascinating because it outlines a mountain of regulatory obligations that come with this clearance. For medtech leaders listening, getting the AI to work is just step one.

Maya: That is so true. The letter explicitly states that the FDA issuance of a substantial equivalence determination does not mean that the FDA has made a determination that the device complies with other requirements of the Act or any Federal statutes.

Sam: For example, they are subject to the Quality Management System Regulation found in 21 CFR Part 820. The letter specifically calls out several ISO 13485 clauses they must follow.

Maya: Yes, it mandates compliance with ISO 13485 clause 7.3 for Design controls. It also lists ISO 13485 clause 8.3 for Nonconforming product, along with clauses for Corrective action and Preventative action.

Sam: It also mentions that device manufacturers have to review and approve changes to device design and production under ISO 13485 clause 7.3 and ISO 13485 clause 7.5, and document all of this in the Medical Device File.

Maya: This means every single time Aramis wants to tweak their neural network or update how that NumPy array is generated, they have a strict, documented process they must follow to stay legally compliant.

Sam: And the FDA notes that further changes might require a brand new premarket notification. They link to guidance documents, specifically one titled Deciding When to Submit a 510(k) for a Change to an Existing Device.

Maya: They also link to another guidance titled Deciding When to Submit a 510(k) for a Software Change to an Existing Device. Those documents are critical because software updates happen constantly in AI.

Sam: The FDA provides the specific document numbers for these downloads in the text: one ends in 99812 and the other in 99785. There are also strict rules about labeling and tracking.

Maya: Right. The FDA points out the final Unique Device Identification System rule. Unless an alternative applies under 21 CFR 801.20(b), the device has to bear a unique device identifier on its label and package under 21 CFR 801.20(a).

Sam: And the dates on that device label must be formatted in accordance with 21 CFR 801.18. Plus, they have to submit certain information to the Global Unique Device Identification Database.

Maya: That database requirement is governed by 21 CFR Part 830 Subpart E, specifically sections 830.300(a) and 830.320(b). It is a massive tracking infrastructure to ensure accountability in the market.

Sam: They also have to comply with medical device reporting for adverse events under 21 CFR Part 803. If the software fails and causes harm, it must be reported to the agency.

Maya: There is also a note about the regulation entitled Misbranding by reference to premarket notification under 21 CFR 807.97. Essentially, companies cannot use the fact that they got this clearance to make misleading claims in their marketing.

Sam: The FDA letter also includes a fascinating disclaimer stating the CDRH does not evaluate information related to contract liability warranties. They are basically saying they checked the science, but the legal warranties are between the company and the hospitals buying it.

Maya: Exactly. They remind the submitter that device labeling must be truthful and not misleading. This clearance is based on the predicate device pathway established prior to May 28, 1976, when the Medical Device Amendments were enacted.

Sam: It is amazing that a law from 1976 is the framework for approving a deep learning algorithm in 2026. The regulatory text mentions that existing major regulations affecting the device are found in Title 21, Parts 800 to 898.

Maya: It bridges two totally different eras of technology. If companies have questions about these regulations, the letter points them to the Division of Industry and Consumer Education.

Sam: They even provide the contact numbers for that division: 1-800-638-2041 or 301-796-7100. And the letter is signed by Jessica Lamb, Assistant Director of the Division of Radiological Imaging Devices and Electronic Products.

Maya: Let us circle back to the company that submitted this, Aramis Global AI Consultancies LLC. Interestingly, their address is listed at the Rahet Albal Business Centre, Court Tower Building, Office #39 in Dubai, UAE.

Sam: Their international phone number is +91 9288045050. But their contact person, John DeLucia, is based at John DeLucia Consulting on 34 Laurel Rd. in Guilford, Connecticut, zip code 06437.

Maya: His phone number is (603) 566-0146. It is a very common setup. A global AI company utilizes a US based regulatory consultant to handle the complex FDA submission process.

Sam: The submission date was May 21, 2026, and the date prepared on the summary is August 8, 2026. The FDA received it on August 17, 2026, before issuing the clearance in September.

Maya: What about combination products? The FDA letter mentions that some cleared products may instead be combination products, governed by 21 CFR Part 4, Subpart A for manufacturing and Subpart B for safety reporting.

Sam: That is standard boilerplate text the FDA includes in these letters. AIRA is purely a Software as a Medical Device, so it isn't a combination product. It also mentions electronic product radiation control provisions under Sections 531-542 of the Act, and 21 CFR Parts 1000-1050.

Maya: Again, probably boilerplate, since the software itself does not emit radiation, it just reads the CT scans. But it shows how comprehensive the FDA checklist is for every single applicant.

Sam: Absolutely. Before we wrap up, what do you think is the biggest takeaway for hospital buyers looking at this new tool?

Maya: For me, it is the intentional data enrichment. The fact that they ensured 25% of the 113 analyzed cases had atrophied kidneys gives me confidence. They proved it on diseased anatomy, not just perfectly healthy organs.

Sam: I agree. And for the medtech developers listening, the takeaway is the regulatory strategy. By comparing themselves to Ceevra Reveal 3, but restricting their claim to just adult kidneys and just CT scans, they secured a smooth path to substantial equivalence.

Maya: The performance metrics speak for themselves. Beating the predicate device on automatic segmentation is a high bar, and AIRA cleared it with a 0.9092 Dice score on the right and 0.9114 on the left.

Sam: To recap, today we unpacked the FDA clearance of the Artificial Intelligence Renal Assessment tool, a deep learning algorithm for measuring kidney volume from CT scans.

Maya: We explored its strong clinical performance, the rigorous ISO quality controls required, and the strategic use of a predicate device to gain market entry.

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