Company Spotlight: Brainomix 360 Hyperdensity - Automating Brain Bleed Measurement

Sam: What if an artificial intelligence tool could automatically spot and measure the exact volume of a suspected brain bleed, and even came with a pre-approved plan from the FDA to get smarter over time?

Maya: That is exactly what we are looking at today with Brainomix 360 Hyperdensity, based on the public FDA 510(k) summary published on June 12, 2026.

Sam: Before we dive into the details, a quick reminder that our voices are AI-generated, and this episode is a summary of the FDA's public 510(k) summary document, not marketing material from the company.

Maya: Exactly. Today is a Company Spotlight episode. We are looking at a specific product from Brainomix Limited. The product code is QIH, which falls under radiological image processing software, and it is a Class II device.

Sam: So, right out of the gate, what is Brainomix 360 Hyperdensity actually built to do? The name sounds very technical.

Maya: The document says its indications for use are for the automatic labeling, visualization, and volumetric quantification of intracranial hyperdensities. It does this from a set of Non-Contrast CT head scans, specifically ones acquired at a single time point.

Sam: Intracranial hyperdensities. In plain English, we are usually talking about bright spots on a CT scan, which in the brain often points to bleeding. Who is the target patient here?

Maya: The summary states it is for adult populations who have undergone a non-contrast CT head scan in the context of suspected acute intracranial hemorrhage as part of their primary diagnostic workup.

Sam: Okay, so a patient comes into the hospital, they suspect a brain bleed, they get a CT scan without contrast. What happens next? How does this software actually step into the workflow?

Maya: It is described as a parallel workflow tool. The software uses a locked artificial intelligence machine learning algorithm. After the scan is performed, a copy of the study is automatically sent to the Brainomix system.

Sam: So the radiologist or neurologist does not have to click a button to send it? It just happens in the background?

Maya: Right. The document mentions an automated workflow that processes image data according to pre-configured routing preferences. It can connect to a CT scanner workstation using a DICOM network integration module. The algorithm then processes the scans, searching for segmentable intracranial hyperdensities.

Sam: And what does the output actually look like for the physician reading the scan?

Maya: The system produces a summary series and a segmentation series in DICOM format. The segmentation series puts an RGB overlay right on each slice of the input series. It also generates hyperdensities segmentation masks.

Sam: RGB overlay. So it colors in the suspect areas on the scan. Does the specific color tell the doctor anything? Like red means one thing and blue means another?

Maya: No, and the summary is very clear about that. It says the colors are only for visual differentiation between the segmented regions, and that the colors do not have a meaning on their own.

Sam: Got it. So it highlights the areas, and I assume it gives you a number for the volume since quantification is in the name.

Maya: Yes. For slices that include hyperdensities, the volume is mentioned in a color legend that is overlaid on the slice. All of this is then exported and sent to a pre-configured PACS destination, along with the original scans, for the physician to review.

Sam: That is a crucial point. It goes straight to the PACS, which is the hospital's primary viewing system. But what if the AI gets it wrong? What if the algorithm includes a bit of the skull or misses a piece of the bleed? Can the doctor edit the volume?

Maya: That is an interesting detail in the document. Users can access the system through a web browser on any machine connected to the platform, including mobile devices. And yes, a user can filter the view in the Brainomix 360 Viewer to add or remove segmented volumes from the total hyperdense volume.

Sam: Okay, that is good. They have a way to correct it.

Maya: But there is a catch. The summary explicitly states that the edited results are not sent back to PACS. They are only viewed at a local level.

Sam: Wait, really? So if a doctor goes into the web browser, fixes the AI's measurement, and gets a new volume, that corrected version does not automatically update in the hospital's main PACS system?

Maya: According to the document, that is correct. It says the edited results are only viewed at a local level. So from a clinical workflow perspective, if a physician relies on an edited volume to make a treatment decision, they would presumably need to document that corrected number manually in their report, because the PACS will still hold the original AI output.

Sam: That is exactly the kind of workflow friction health system leaders need to understand. Let us shift gears to how Brainomix actually got this cleared. In the 510(k) pathway, you have to prove your device is substantially equivalent to something already on the market. What was their predicate device?

Maya: They compared it to a device from Viz.ai, specifically Viz HDS, which was cleared under submission number K232363. The summary notes they have the same intended use. Both are software as a medical device packages that automate the manual process of identifying, labeling, and quantifying the volume of segmentable intracranial hyperdensities.

Sam: Are they identical in what they output, or did Brainomix do things differently?

Maya: There is a notable difference. The summary points out that the indications for use of the Brainomix device do not provide two of the outputs that the Viz HDS predicate does. Specifically, Brainomix does not output lateral ventricles volume or midline shift.

Sam: That is interesting. Midline shift is a major indicator of pressure in the brain, often caused by a bleed. Why would they leave those out? Does the FDA summary say?

Maya: The summary does not explain the company's business or clinical reason for leaving them out. It only states that the absence of these outputs does not raise any safety or efficacy questions. It argues that identifying and quantifying hyperdensities is a distinct radiological finding, so not having the ventricle volume or midline shift does not negatively impact the core function.

Sam: Makes sense from a regulatory standpoint. You do not have to do everything the predicate does, as long as what you do offer is equivalent in safety and effectiveness. Which brings us to the actual performance data. How did they prove this works?

Maya: They performed a clinical study comparing the algorithm's output to the ground truth established by trained radiologists. They used two primary outcome measures: mean absolute error, and DICE score.

Sam: Let us take those one by one. Mean absolute error, or MAE. What were the results?

Maya: The study demonstrated that the MAE for hyperdensities total volume, looking at the upper 95% confidence interval bounds, was less than 7.5 mL between the algorithm and the established ground truth.

Sam: Less than 7.5 mL difference on average. That gives us a sense of the margin of error when measuring the size of these bleeds. And what about the DICE score? The document defines that as describing the degree of agreement between the algorithm's measurements and the manual measurements.

Maya: Exactly. For the DICE score, the study demonstrated that the lower confidence interval bound was greater than 70%. The summary states both of these results aligned with their performance goals.

Sam: So we have the error margin for volume, and the spatial overlap score. But whenever we look at AI performance, we always want to know about the testing data. How many scans did they test this on? What kind of scanners were used?

Maya: That is where this public summary is quite limited. It states that device performance was stratified by clinical site, sex, age, race or ethnicity, slice thickness, scanner manufacturer, and the size of the estimated quantity.

Sam: Stratified, meaning they looked at the results across all those different categories to check for biases or blind spots. But what are the actual numbers? How many patients?

Maya: The document does not say. It provides zero numbers regarding the size of the test dataset. It does not state how many patients or scans were included, how many different clinical sites were involved, or whether the testing was retrospective or prospective. We only have the final performance metrics of less than 7.5 mL for error and greater than 70% for the DICE score.

Sam: That is a really important caveat for health system buyers. The FDA obviously reviewed the full data package to grant clearance, but the public summary alone does not give you the sample size to judge how robust the testing was. You would have to ask the company for the full clinical validation report.

Maya: Exactly. Now, there is one more section in this document that is arguably the most forward-looking part of the whole submission. Section 12 details a Predetermined Change Control Plan, or PCCP.

Sam: Yes, let us unpack this. This is a relatively new regulatory concept under the Food and Drug Omnibus Reform Act of 2022. What does having a PCCP actually allow Brainomix to do?

Maya: It is essentially an agreement with the FDA on how the company can update the AI algorithm in the future without having to submit a brand new 510(k) application. As long as the changes stick to this authorized plan, they are cleared to proceed.

Sam: That is a massive advantage for a software company. It means they can iterate faster. What specific modifications are they allowed to make under this plan?

Maya: The document outlines two specific modifications. The first is retraining the model weights using expanded datasets. The rationale given is to address underrepresentation and morphological variability observed in real-world use.

Sam: So if they find out the AI struggles with scans from a certain demographic, or maybe scans from a new type of CT machine, they can feed it new data to improve its accuracy.

Maya: Exactly. The stated benefit is improved segmentation accuracy and volumetric precision. But the plan also lists the risks of doing this, which it identifies as a reduction in clinical performance due to overfitting and unintended bias.

Sam: Overfitting being when the AI memorizes the training data too well and fails in the real world. What is the second modification they are allowed to make?

Maya: The second one is the adjustment of probability thresholds for specific subtype classes to optimize performance. The rationale is that this enables fine-grained performance optimization without introducing new autonomous behavior or expanding clinical claims.

Sam: Meaning they can tweak the internal dials of the algorithm to make it more or less sensitive to certain types of bleeds. And again, the risks listed for this?

Maya: The risks are a reduction in clinical performance due to over- or under-segmentation, volumetric distortion, and unintended bias. To mitigate these risks, the document states they will use the same verification and validation framework applied to the originally cleared device.

Sam: Now, when we talk about AI that gets updated with new data, some people might assume the software is learning on the fly in the hospital, getting smarter with every patient it scans. Is that what this is?

Maya: No, definitely not. The summary is extremely clear on this point. It says, the PCCP does not include provisions for implementation of adaptive algorithms that will continuously learn in the field.

Sam: So the algorithm is locked. When they want to update it, they have to do that in their own labs, test it against their acceptance criteria, lock the new version, and then push it out as a software update to the hospitals.

Maya: That is right. The document notes that Brainomix will deploy the updates to the virtual machines or dedicated servers, so the end user does not need to take action, but they will provide release notes prior to deployment so users are informed of the changes.

Sam: This really paints a picture of where medical AI regulation is heading. You are not just clearing a static piece of code anymore; you are clearing a managed lifecycle of a product. It allows companies to respond to real-world data without getting bogged down in endless regulatory filings, provided they follow their own strict testing protocols.

Maya: And the summary reiterates that these updates cannot change the device's clinical role, user workflow, or interpretation of results. It is strictly about optimizing the performance within the already cleared indications for use.

Sam: This has been a great deep dive. Let us do a quick recap. We looked at Brainomix 360 Hyperdensity, an AI tool cleared to find and measure suspected brain bleeds on non-contrast CT scans. It works automatically in the background, sending color-coded masks and volume measurements directly to the hospital's PACS.

Maya: It was cleared based on substantial equivalence to the Viz HDS device, though it does not measure lateral ventricles or midline shift. Performance data showed an error margin of less than 7.5 mL for volume and a DICE score greater than 70%, though the summary lacks details on the dataset size.

Sam: And perhaps most notably, it comes with a Predetermined Change Control Plan, giving the company a green light to retrain the model and adjust thresholds to improve accuracy over time without submitting a new 510(k).

Maya: That covers the key takeaways from the document. A reminder that you can read the full FDA 510(k) summary for yourself; the link is in the show notes.

Sam: And as always, this podcast is for informational purposes only and is not medical advice. Thanks for listening, and we will catch you on the next episode of AI in Medicine - Smart Summaries.