Company Spotlight: Median Technologies eyonis LCS - Risk Stratifying Lung Cancer Screening

Sam: Lung cancer screening programs catch early disease, but they are also famous for turning up a lot of false positives that require stressful, unnecessary follow-ups. What if an AI could not just find the nodules, but actually score their risk of malignancy right on the scan?

Maya: Today we are looking at a document that tackles exactly that. We are unpacking the FDA 510(k) summary for eyonis LCS 1.1, published by the U.S. Food and Drug Administration on February 6, 2026, for the manufacturer Median Technologies.

Sam: And before we get fully into the weeds, we should remind you that our voices are AI-generated, and this episode is a summary of a publicly available document, not a piece of company marketing.

Maya: Exactly. This is a Company Spotlight episode based strictly on the FDA's clearance summary. That means we are looking at the exact numbers the agency reviewed, how the software works, and what the limitations are.

Sam: So let us start with the basics. What exactly is eyonis LCS 1.1? The document calls it a Radiological Computer Assisted Detection And Diagnosis Software, which falls under Product Code QDQ.

Maya: Right, and the key there is the combination of detection and diagnosis. In regulatory terms, the document describes it as an end-to-end CADe/CADx Software as a Medical Device. CADe means computer-aided detection, basically finding the needle in the haystack. CADx is computer-aided diagnosis, which is characterizing what that needle actually is. It is designed to work on low-dose Chest CT images.

Sam: And the target audience here is quite specific, right? The summary mentions high-risk patients eligible for participation in lung cancer screening programs based on USPSTF criteria.

Maya: Yes. It is looking for pulmonary parenchymal nodules that are between 4-30 mm in diameter. And it specifically targets solid and part-solid nodules. It generates a result report that flags nodules as probably benign, suspicious, and very suspicious.

Sam: But it does not just put a box around them. It actually scores them. The document says it provides a 10-point scale score, along with an associated malignancy rate observed in their reference population.

Maya: That is the critical piece for a radiologist's workflow. Alongside that 10-point score, the report provides the slice number, a full snapshot, a close-up snapshot, the long, short, and average diameters, and the total volume of the nodule.

Sam: Okay, but before we get into how well it actually works, we always have to look at the predicate device. With the 510(k) pathway, a company has to prove their device is substantially equivalent to something already on the market. What is the predicate here?

Maya: The predicate device is Transpara, manufactured by ScreenPoint Medical B.V. That device was cleared under the 510(k) Number K241831.

Sam: Wait, really? I usually associate Transpara with mammography screening. How does that work as a predicate for a lung cancer tool?

Maya: It is a great question. The FDA summary explicitly notes that there are differences. The indication for use differs in the target organ, the disease-specific findings, the type of medical images, and the patient population. But legally, they are both Class II software devices intended as a concurrent reading aid for physicians interpreting screening images. They both identify findings and assess the level of suspicion of malignancy.

Sam: So the FDA is looking at the fundamental scientific technology. They both use deep learning modules trained on large databases of cancer and normal patients. The core difference in output is that Transpara gives a continuous 100-point score, whereas Median LCS, or eyonis LCS, gives a simplified 10-point score.

Maya: Exactly. The FDA concluded that these differences do not raise new questions regarding safety and effectiveness, which is the legal standard for substantial equivalence.

Sam: Alright, let us talk about the proof. How did Median Technologies actually test this? The summary outlines two main testing phases: a standalone performance test and a clinical performance test.

Maya: Let us start with the standalone test. This was a retrospective cohort study evaluating the software across 1,147 patients. Out of those, there were 342 cancers and 805 non-cancer cases. These came from 7 different datasets, including academic sites in Europe and the United States, as well as private American data providers.

Sam: The document notes this dataset was enriched. It says it was enriched for cancer prevalence, nodules size, and spiculation, leading to a 1:2 cancer to benign ratio. Why do they do that? That does not sound like the normal prevalence in a screening population.

Maya: You are absolutely right. In a real-world screening population, the cancer rate is much, much lower. But in standalone testing for an algorithm, you need enough positive cases to statistically power the endpoints. If they used a natural prevalence, they would need tens of thousands of scans just to get enough cancers to prove the software works. The document says the non-cancerous population is very heterogeneous, so they needed adequate sampling there too.

Sam: Got it. And the patients were all high-risk individuals aged 50 to 80 with a history of smoking. So, how did the algorithm perform on its own, without a human?

Maya: The results are solid. They achieved a patient-level AUROC of 0.904. For context, their acceptance criterion was anything above 0.800.

Sam: And AUROC stands for Area Under the Receiver Operating Characteristic curve, basically a measure of how well the model distinguishes between cancer and non-cancer. A 0.904 is definitely strong. But what about sensitivity and specificity? The document mentions a Clinical Operating Threshold, or COT.

Maya: Right, the COT was pre-determined as the threshold between LCS scores 3 and 4. At that threshold, the software had a sensitivity of 84.50% and a specificity of 80.25%. Both of those easily beat their acceptance criteria of 70%.

Sam: They also mentioned a localization capability, right? An AULROC metric?

Maya: Yes, the AULROC was 0.869. That confirms the algorithm is not just guessing there is cancer somewhere in the scan, it is actually pointing to the correct location.

Sam: Okay, so the software works on a computer in a lab. But what happens when you put it in front of actual doctors? That brings us to the clinical performance testing, the multi-reader, multi-case retrospective study.

Maya: This is where the rubber meets the road for health-system leaders. They used a Paired-Split Plot design with 16 clinical readers. These were all US Board Certified radiologists from multiple regions, with a mean of 13.31 years of experience. The range was broad, from 2 to 32 years of experience.

Sam: And they looked at 480 total patient images. The document says it was conducted sequentially to simulate clinical practice. They interpreted each patient twice: once unaided as the control arm, and once aided by eyonis LCS as the test arm, in random order.

Maya: The primary objective was to see if the accuracy of the readers improved when they had the AI report. And it did. The aided AUC was 0.8434, while the unaided AUC was 0.8276. That is a difference of 0.0158, which met their statistical requirement for superiority.

Sam: A bump from 0.8276 to 0.8434 might sound small to a layperson. But the secondary endpoints really tell the story of what this does on the ward. Let us look at specificity, which is the ability to correctly identify the benign cases and avoid false alarms.

Maya: This is the most crucial number for clinical workflow. The aided specificity was 53.59%, compared to the unaided specificity of 49.45%. That is an absolute increase of 4.14. It is statistically significant.

Sam: So radiologists were better at ruling out cancer when they had the AI. But you always worry that if you increase specificity, you might accidentally drop your sensitivity and miss actual cancers. Did that happen?

Maya: No, and that is a key finding. The aided sensitivity was 93.75%, and the unaided was 92.50%. So there was a numerical increase of 1.25, but the summary notes this was non-significant for superiority. However, it met the non-inferiority objective. As the document puts it, the specificity is significantly improved while patient safety is preserved because the sensitivity is not different.

Sam: There is also a really interesting data point about inter-reader agreement. Anyone who works in imaging knows that two radiologists can look at the same scan and disagree. Did the AI help standardize their reads?

Maya: It did. The intra-class correlation coefficient, which measures agreement per patient score, went from 0.707 unaided to 0.830 aided. The agreement on patient management, meaning the decision on whether to change follow-up frequency, also increased significantly.

Sam: That is huge for a hospital system. Standardizing care pathways and reducing unnecessary recalls is exactly what you want from a screening program. But let us talk about how this actually fits into the IT infrastructure. Hospital IT teams are always wary of new software.

Maya: Median Technologies designed this to be quite flexible. The document states it is a Container-based Image Processing chain with no associated viewer. It can be deployed in any environment that supports Kubernetes, either cloud-based or on premises. It just runs in the background and drops a DICOM result report straight into the PACS system.

Sam: And there is a licensing component mentioned too, called eyonis trust. It is a license server that authorizes or revokes customer access, tracks usage, and ensures compliance with contracts. So they have clearly thought about the commercial deployment side.

Maya: Yes, but we also have to look closely at what the software does not do. The indications for use are very specific about what is excluded. The document states exactly: pure ground glass, mediastinal lesions and masses - including but not limited to hilar masses - are excluded.

Sam: Does the summary explain why pure ground glass nodules are excluded?

Maya: It does not. The FDA summary simply lists them as excluded from the intended population and the indications for use. If a patient only has pure ground glass cancer, they are not in the target population for this specific tool. We can only report what is in the document, and the clinical rationale for that exclusion is not provided here.

Sam: Fair enough. Another limitation to keep in mind is that the clinical performance testing was a retrospective reader study. It was conducted sequentially to simulate clinical practice, but it was still an enriched, retrospective environment, not a live, prospective trial.

Maya: That is standard for this type of 510(k) clearance, but it is an important context for buyers. Also, the document repeats several times that eyonis LCS is not intended to replace the clinical judgment of the interpreting physician. It is intended to be used in a concurrent read mode, meaning the AI results are displayed alongside the original CT images.

Sam: So, here is a crisp recap: Median Technologies has cleared a Class II SaMD that detects, measures, and scores the malignancy risk of solid and part-solid lung nodules. By giving radiologists this concurrent reading aid, their retrospective study showed they could boost specificity from 49.45% to 53.59% without hurting sensitivity.

Maya: Exactly. For a health system, the promise here is standardizing radiologist reads and potentially reducing the recall rate for false positives in high-risk screening populations, provided the nodules fall within the 4-30 mm range and are not pure ground glass.

Sam: That is all we have for this Company Spotlight on Median Technologies and eyonis LCS 1.1. You can find a link to the full FDA 510(k) summary in our show notes. We always recommend reading the source document yourself.

Maya: And a final reminder: this podcast is for informational purposes only, and is not medical advice. Thanks for joining us on Smart Summaries.