Company Spotlight: Therapixel MammoScreen 5 - AI Breast Cancer Screening Clearance

Sam: Out of 9,789 patients in this medical device test dataset, a staggering 9,461 patients did not have their race or ethnicity reported.

Maya: And yet, the product cleared the regulatory bar. Today we are unpacking the FDA 510(k) summary for MammoScreen 5, published by the U.S. Food and Drug Administration.

Sam: Before we dive into the numbers, we need to state upfront that our voices are AI-generated, and this episode is a summary of a public document, not company marketing material.

Maya: Exactly. This is a Company Spotlight episode. We are looking at a product by a company called Therapixel. The document is their 510(k) summary, specifically submission number K260714, cleared on June 26, 2026.

Sam: Right, so this is the official paperwork proving to the FDA that their device is substantially equivalent to something already on the market. Let us start with the basics. What does MammoScreen 5 actually do?

Maya: It is a concurrent reading and reporting aid for physicians who are interpreting screening mammograms. It is a software-only device that processes the images and detects findings suspicious for breast cancer.

Sam: Concurrent reading means the radiologist is looking at the AI outputs at the same time they are doing their own read, correct?

Maya: That is right. The software places graphical marks on the mammograms for soft-tissue lesions or calcifications. And for every finding it detects, it assigns what they call a level of suspicion score.

Sam: How does that scoring system work in practice? Is it just a percentage?

Maya: The document says the score goes from one to ten. It is designed so that a low score means a very low level of suspicion, and the suspicion level increases as the score goes up. But it does not just give one overall number. It gives a score for each specific finding, a score for each breast driven by the highest finding in that breast, and then an overall score for the case.

Sam: That multi-level scoring makes sense for workflow. If the overall case score is low, the radiologist knows what they are dealing with immediately. Does it tell you what kind of lesion it thinks it found?

Maya: Yes, it categorizes the lesion type as mass/asymmetry, distortion, or calcifications. It also provides the exact location, including the quadrant, the depth, and the distance from the nipple.

Sam: I see a note here about how it can actually help write the report. The summary states it can populate an initial report with those chosen findings and positions. That is a tangible workflow time-saver.

Maya: It is, but the FDA document includes a very clear warning. It emphasizes that patient management decisions should not be made solely based on analysis by the device. It is adjunctive information. The medical professional interpreting the image remains the sole decision-maker.

Sam: Which is standard for this class of device. Speaking of classes, the FDA categorizes this as a Class II device under Regulation Number 21 CFR 892.2090. The regulation name is Radiological Computer-Assisted Detection And Diagnosis Software.

Maya: Exactly. And because this is a 510(k) clearance, Therapixel had to prove substantial equivalence to a predicate device. The predicate here is MammoScreen 4, which was cleared under submission K243679.

Sam: So they are comparing version 5 to version 4. What is actually different between the two? Did they change the core AI algorithm?

Maya: Interestingly, no. The summary explicitly states that the algorithm behind MammoScreen 5 did not change or evolve compared to its predecessor. Both versions use deep learning modules trained on very large databases of biopsy-proven examples of breast cancer and normal tissue.

Sam: If the algorithm is the same, why did they need a whole new FDA clearance for version 5?

Maya: Because they introduced new supported scenarios. Version 5 corrects some bugs, but the main goal of this submission was to validate the software on new combinations of images. They added support for supplementary views, duplicated views, and new combinations of current and prior exams.

Sam: That makes sense. Radiology departments acquire images in lots of different combinations. Full-field digital mammography, which is 2D, and digital breast tomosynthesis, which is 3D. The document abbreviates those as FFDM and DBT.

Maya: Yes, and to prove MammoScreen 5 works on these new combinations, they did what is called standalone performance testing. They did not run a new clinical trial. In fact, the document notes they are relying on clinical testing from an even older version, MammoScreen 3, which was cleared under K240301. For version 5, the standalone testing was deemed sufficient.

Sam: Let us talk about that standalone testing, because this is where the numbers get really interesting. They used a dataset of 9,789 patients. But the demographic breakdown is full of contradictions.

Maya: The numbers in the summary are certainly confusing. For the age distribution, they report 2,383 patients aged 50 or younger. Then 3,091 patients between 50 and 65. And finally, 1,614 patients over 65.

Sam: Right, but the total number of patients is 9,789. The document offers no explanation for why those three age groups do not add up to that total.

Maya: And the race and ethnicity distribution is even more concerning. They list Asian as 652, White as 2,295, Black as 961, Other as 467, and Hispanic as 328.

Sam: But then it lists the number of patients where race and ethnicity were not reported as 9,461.

Maya: Exactly. The reported categories plus the 9,461 not reported vastly exceed the stated total of 9,789 patients. The math simply does not align in the FDA summary document.

Sam: We can only report the exact digits as they are written in the source, but for a hospital evaluating this tool, that is a massive red flag. Even if we assume the 9,461 figure is correct out of the total 9,789, that means almost the entire dataset lacks race and ethnicity data.

Maya: Which is a known limitation in retrospective AI validation. If you do not know the demographics of your test group, it is very hard to evaluate algorithmic bias, especially since breast tissue density and cancer presentation can vary across different populations. The summary does not provide any context for these mathematical discrepancies.

Sam: Despite that, the FDA found the device substantially equivalent based on the performance metrics. How exactly did they measure performance?

Maya: They used a metric called the Area Under the Receiver Operating Characteristic Curve, or AUCROC. And they looked at it across three levels: the mammogram level, the breast level, and the finding level. The goal was to prove non-inferiority compared to scenarios already supported by MammoScreen 4.

Sam: So they just needed to prove version 5 is not worse than version 4. Let us look at the supplementary views scenario. What were the numbers there?

Maya: When including supplementary views, the AUCROC at the mammogram level was 0.911. The reference scenario, which excluded them, was 0.894. That is a difference of 0.017, and the p-value is listed as 0.000.

Sam: So performance actually went up slightly. What about duplicated views?

Maya: At the mammogram level, including duplicated views resulted in an AUCROC of 0.920, compared to 0.907 without them. That is a positive difference of 0.013. The finding level metric also showed a slight increase, going from 0.900 up to 0.908.

Sam: Then they get into the really granular combinations of current and prior images. For example, testing a current DBT and FFDM scan with a prior DBT scan.

Maya: Yes, for that specific combination, the mammogram level AUCROC was 0.951. They compared it to a reference scenario of a current DBT and FFDM with a prior FFDM, which scored 0.955. The difference there was -0.004.

Sam: So a tiny bit lower. What about looking at a current 2D image, an FFDM, combined with a prior 3D image, a DBT?

Maya: In that scenario, the mammogram level AUCROC was 0.864. The reference, which was an FFDM with no prior at all, scored 0.869. The difference was -0.005, and the lower bound of the 95% confidence interval was -0.013. Interestingly, the document's stated acceptance criteria is a positive lower bound, but its final conclusion simply states that the standalone tests demonstrate non-inferior performance compared to the predicate.

Sam: How did they establish the absolute truth for these tests? How did they know if a lesion was actually cancer or just benign tissue?

Maya: The truthing process is detailed clearly. For positive cases, they required biopsy-proven presence of cancer. For negative cases, they verified them by imaging follow-up.

Sam: And for the benign cases in the middle?

Maya: It depended on whether a biopsy was done. If the case was not biopsied, it was verified by imaging follow-up. If it was biopsied, they required the biopsy result confirming it was benign, plus the imaging follow-up.

Sam: The document also makes a point about the independence of the test data. It states that the sources in the test group were entirely left out during training and tuning. The data used for this standalone performance testing only belonged to the test group.

Maya: That is vital for proving the model actually generalizes to new, unseen data, rather than just memorizing the images it was trained on.

Sam: I noticed the FDA letter itself includes a lot of standard regulatory requirements. It mentions that Therapixel is subject to the Quality Management System Regulation, or QMSR. It specifically cites ISO 13485 clauses for Design controls, Nonconforming product, Corrective action, and Preventative action.

Maya: Yes, those are the general controls provisions of the Act. The FDA also notes that medical devices must comply with the final Unique Device Identification System rule, meaning the device needs a unique identifier on its label and in the Global Unique Device Identification Database.

Sam: It is a good reminder that gaining 510(k) clearance is not just about AUCROC scores. It is about maintaining strict manufacturing and design controls. The FDA also explicitly states that clearance does not mean they evaluate information related to contract liability warranties.

Maya: Exactly. And for health system leaders buying this software, the document confirms that Therapixel applied a risk management process following FDA-recognized standards. They evaluated risks that could cause false positives or false negatives and concluded the residual risk is outweighed by the benefits.

Sam: Before we wrap up, I want to circle back to what this means for a radiology department. MammoScreen 5 is a software-only device. The fact that we are seeing version 5 cleared based on the clinical data of version 3, and the predicate of version 4, really highlights how AI products iterate in the market.

Maya: It changes the dynamic in the boardroom. You are not buying a static piece of hardware. You are licensing a software pipeline. As the vendor adds support for new imaging combinations or fixes bugs, they submit a special 510(k) for substantial equivalence to keep the product moving forward.

Sam: But the core value proposition remains the same. It detects lesions, scores them from one to ten, and gives you the quadrant, depth, and distance to the nipple to help draft the report. As long as you remember it is just an adjunctive tool.

Maya: Precisely. The memorable takeaway from this document is that MammoScreen 5 successfully gained clearance for new complex imaging scenarios, maintaining high performance metrics, even though the demographic reporting in the standalone testing dataset is severely lacking.

Sam: That is a perfect recap. MammoScreen 5 is cleared and on the market. We have linked the full FDA 510(k) summary in the show notes so you can look at the data yourself, especially those demographic tables.

Maya: And please remember, this podcast is for informational purposes only and is not medical advice. Thank you for listening to AI in Medicine - Smart Summaries.