Company Spotlight: Sonio Suspect - Moving Fetal Ultrasound AI from the Cloud to the Machine

Maya: Welcome to a Company Spotlight episode of AI in Medicine - Smart Summaries. I am Maya.

Sam: And I am Sam. Right up front, we want to remind you that our voices are AI-generated. Today, we are bringing you a summary of a publicly available document.

Maya: Specifically, we are looking at an FDA 510(k) summary for a product called Sonio Suspect, developed by a company named Sonio. The FDA published this clearance letter on June 4, 2026. The hook for today is that Sonio took an AI tool designed to detect abnormal fetal ultrasound findings with an average sensitivity of 92.5%, and secured clearance to pull it out of the cloud and embed it directly into the ultrasound machine itself.

Sam: Which is a huge shift in how AI actually lives in the clinic. But before we get into the weeds of how they did that, let us set the ground rules. This is a Company Spotlight episode, but it is not sponsored, and it is definitely not an advert. We are basing this entire conversation solely on the FDA's public 510(k) summary document.

Maya: Exactly. If it is not in the official FDA paperwork, we are not talking about it. We do not have any inside marketing material or funding numbers here. And as always, this is not medical advice.

Sam: Alright, let us dive in. Maya, what exactly is Sonio Suspect and who is it built for?

Maya: According to the indications for use, Sonio Suspect is intended to assist interpreting physicians by automatically identifying and characterizing abnormal fetal ultrasound findings on detected views. It uses machine learning techniques to do this. The document specifically notes it is for use during and/or after fetal ultrasound examinations, covering a gestational age from 11 weeks to 41 weeks.

Sam: So, 11 to 41 weeks is a pretty broad window for maternal-fetal medicine. And it says it can be used during or after the exam. Does that mean it is operating in real time while the clinician is scanning?

Maya: The document uses the specific phrase concurrent reading aid on acquired images. It clarifies that a concurrent read means the device output is available during and/or after the fetal ultrasound examination. So it is assisting the healthcare professional right there in the workflow.

Sam: I noticed something a bit strange in the regulatory classification section. This is clearly a tool for fetal ultrasounds, obstetrics, but the regulation name listed in the document is Radiological Computer-Assisted Diagnostic Software For Lesions Suspicious Of Cancer. And the regulation number is 21 CFR 892.2060. Why does it say cancer?

Maya: That is a great catch. The document absolutely states that exact regulation name: Radiological Computer-Assisted Diagnostic Software For Lesions Suspicious Of Cancer. However, the summary does not explain why an obstetric tool falls under that label. Often, the FDA maps new AI tools to existing regulatory product codes that share similar technical characteristics, like computer vision on ultrasound images, even if the clinical specialty is different. But strictly speaking, the document leaves that discrepancy unexplained, so we just have to accept that it is classified there as a Class II device under product codes POK and QIH.

Sam: Fair enough. So it is a Class II device. Let us talk about why this particular 510(k) was submitted. Because Sonio Suspect was already cleared before this, right?

Maya: Yes, it was. The summary names the predicate device as Sonio Suspect, cleared under K243614. That earlier version is described as a cloud-based web application. In that setup, the data processing is performed remotely in the cloud, and the physician accesses it via a secure internet connection through a web browser.

Sam: So if it was already cleared, what does this new document, K261519, actually change?

Maya: This submission is a Special 510(k) for what they call the Integration Configuration. They wanted to integrate Sonio Suspect directly into third-party medical devices, like ultrasound systems. To do that, the document says the device change consists of the migration of the Sonio Suspect algorithms from the Python programming language over to C++.

Sam: That is a fascinating detail for the developers listening. They rewrote the core algorithm from Python to C++ so it could be embedded locally on the machine, rather than sending data to the cloud. What does the document say about how that integration actually works in practice?

Maya: It lays out a clear division of labor. Sonio supplies the C++ algorithm package to the third-party manufacturer and provides integration support. But, and this is crucial for the regulatory side, the third-party manufacturer is responsible for implementation of the user interface, system integration, and validation of the final device.

Sam: That makes total sense. If you are an ultrasound hardware company buying this AI, you cannot just plug it in and skip the FDA. You still have to validate how your specific screen displays those findings to the doctor.

Maya: Exactly. The document states that the algorithmic logic, model weights, decision thresholds, and clinical functionality remain completely unchanged from the cleared Python version. The differences are strictly limited to the implementation language, runtime environment, deployment model, and user interface responsibility.

Sam: Okay, but if you rewrite a massive machine learning algorithm from one language to another, you have to prove you did not break anything. Let us talk about the performance data. How did they test this?

Maya: They conducted performance bench testing. The summary says they used the exact same protocol, dataset, statistical analysis plan, and acceptance criteria applied in the previous clearance to do a direct comparison.

Sam: Did the numbers match up exactly?

Maya: They were extremely close, but there are slight variations. For the Average Performance across all trimesters, the old Web Application had a sensitivity of 93.2%. The new Integration Configuration has an average sensitivity of 92.5%.

Sam: Wait, really? The sensitivity actually dropped from 93.2% down to 92.5%? What about specificity?

Maya: Specificity also saw a tiny shift. The old version was 90.8%, and the new version is 90.4%. However, the FDA concluded that these differences do not raise new questions regarding the safety and effectiveness of the device when used as labeled, and determined it to be substantially equivalent.

Sam: That is the magic phrase in a 510(k): substantially equivalent. I want to dig into Table 2 of the document. It lists the actual fetal anatomy findings they tested the software on. It is not just one generic score. They break it down by chest, abdominal, and cephalic, or head, findings.

Maya: Yes, let us look at the chest findings first, mostly for the second and third trimesters. For Malposition of the Great Vessels, the new C++ configuration achieved a sensitivity point estimate of 0.876 and a specificity of 0.934.

Sam: And there is another heart-related finding: Absence or unusual size of at least one of the 3 vessels. The sensitivity there was very high, at 0.952, but the specificity was lower, at 0.819.

Maya: Exactly. They also looked at Disequilibrium OR absence of at least one of the two ventricles. That showed a sensitivity of 0.882 and a specificity of 0.815. Rounding out the chest, they tested for Thoracic Situs Inversus, which is when the organs in the chest are mirrored. For that, performance was strong: sensitivity 0.949 and specificity 0.947.

Sam: It is fascinating to see the breakdown, because AI is never just uniformly perfect across the board. What about the abdomen?

Maya: For Abdominal Situs Inversus across all three trimesters, the sensitivity was extremely high at 0.975, and specificity was 0.974. They also tested for Non-visibility of a single stomach bubble OR abnormally big stomach. That gave a sensitivity of 0.880 and a specificity of 0.965.

Sam: Okay, and lastly, the cephalic anomalies. The brain is notoriously difficult to image perfectly depending on fetal position.

Maya: For the brain, in the second and third trimesters, they looked at two major findings. First, Absence of the Cavum Septum Pellucidum, which hit a sensitivity of 0.913 and a specificity of 0.925. Second, Absence of the Corpus Callosum. For that, the sensitivity was an impressive 0.976, but the specificity dropped to 0.858.

Sam: So sensitivity of 0.976 for missing the Corpus Callosum means it is highly likely to flag it if the abnormality is truly there, but the specificity of 0.858 means it might throw some false alarms on normal scans. That brings up a huge question for me regarding the data. How many scans did they actually run through this bench test? What was the size of the dataset?

Maya: That is a very important point, and we have to be completely transparent here. The FDA 510(k) summary document does not state the total size of the dataset. It does not tell us how many scans or how many patients were involved in this bench test.

Sam: Wait, really? No patient numbers at all in the summary?

Maya: None in this public document. It simply says the performance was validated for subgroups. It lists those subgroups as: Indication of examination, Race, Ethnicity, BMI, Gestational Age, Maternal age, Manufacturer, Geography—noting OUS and US, which means outside the US and within the US—Finding Status, meaning finding-positive and finding-negative images, and Fetal Anatomy.

Sam: It is good that they validated across different maternal BMIs and races, because ultrasound image quality is heavily dependent on those factors. But without knowing the actual size of the dataset, listeners need to keep in mind what this document represents. This is a summary of substantial equivalence to a predicate device. It is not a prospective randomized clinical trial.

Maya: Absolutely. Which leads us directly to the limitations and warnings explicitly stated in the indications for use. The document contains a very clear, stand-alone sentence: Patient management decisions should not be made solely on the results of the Sonio Suspect analysis.

Sam: That is the classic AI caveat. The human is still the primary diagnostician. The software provides information on abnormal findings that may be useful in rendering potential diagnosis, as the document says, but it is not replacing the maternal-fetal medicine specialist.

Maya: Correct. From a health-system leader's perspective, this Special 510(k) clearance is about deployment strategy. Integrating an algorithm natively into an ultrasound machine removes the dependency on cloud connectivity and browser access during a high-stakes scan. It standardizes the real-time workflow.

Sam: And from a medtech company perspective, this document is a great case study in the regulatory mechanics of software updates. Transitioning your tech stack from a Python cloud app to a C++ embedded package requires thorough bench testing to prove the code translation did not degrade performance. Even when your average sensitivity shifts slightly from 93.2% to 92.5%, if you prove it handles the same anomaly subgroups effectively, the FDA can deem it substantially equivalent.

Maya: Let us do a quick recap. The document we explored today is the FDA 510(k) summary for Sonio Suspect, an AI tool for detecting abnormal fetal ultrasound findings between 11 and 41 weeks.

Sam: We learned that this specific clearance allows Sonio to package their algorithm in C++ for local integration into third-party ultrasound machines, moving away from a strictly cloud-based approach.

Maya: We saw the performance data, with an average sensitivity of 92.5% and specificity of 90.4%, alongside specific numbers for anomalies like abdominal situs inversus and absence of the corpus callosum. And importantly, we noted that the public summary does not disclose the size of the testing dataset.

Sam: And finally, the ultimate warning from the document: patient management decisions must not rely solely on this AI tool.

Maya: That is all for this Company Spotlight. A reminder that you can find a link to the full FDA 510(k) summary document in our show notes. Please read the primary source for yourself.

Sam: And remember, this podcast is for informational purposes only. It is not medical advice. Thanks for listening, and we will catch you next time.