16 November 2026 · 14 min
Company Spotlight: iSchemaview Rapid MLS - Automating Brain Midline Shift on CT
We review the FDA 510(k) summary for iSchemaview's Rapid MLS, an AI software designed to measure brain midline shift from non-contrast CT scans. The episode unpacks the algorithm's performance against human experts, its clinical indications for intracranial hemorrhage and TBI, and the regulatory strategy of claiming substantial equivalence to a heart measurement tool.
Key points
- Rapid MLS uses AI/ML to measure brain midline shift on non-contrast CTs, providing a color overlay and quantitative JSON output.
- The software was cleared via the 510(k) pathway by claiming substantial equivalence to a predicate device that measures heart ventricles from CT pulmonary angiography (Rapid RV/LV).
- In final performance validation on 153 cases, Rapid MLS achieved a mean absolute error of 0.7 mm, compared to the 3 experts' average pairwise error of 1.0 mm.
- Validation data included a demographic split of 44% male and 56% female, spanning ages 26-93 years, sourced from 13 independent sites.
- The device is not intended for stand-alone clinical decision-making and remains restricted to sale by or on the order of a physician.
Source: FDA 510(k) summary K243378: Rapid MLS (iSchemaview) - U.S. Food and Drug Administration, 2025
This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.
Company: iSchemaview (ischemaview.com) · Product: Rapid MLS · FDA 510(k) K243378
Transcript
Sam: What happens when an AI algorithm measures the exact midline shift of a patient's brain down to fractions of a millimeter, and actually reports a smaller margin of error than the human experts? And what does it mean when the FDA clears that brain-analyzing software based on its similarity to a device that measures the heart? Welcome to Smart Summaries.
Maya: Today we are doing a Company Spotlight on the FDA 510(k) summary for Rapid MLS, published by the FDA on May 28, 2025. The applicant is iSchemaview, Inc. As always, a quick disclaimer before we dive in: our voices are AI-generated, and this is an independent summary of a publicly available FDA document, not company marketing material.
Sam: This is a really fascinating document to unpack, Maya. We are looking at a clearance for a medical image management and processing system, specifically a software device called Rapid MLS. Let's start with the basics. What exactly does this product do, and who is it for?
Maya: According to the summary, Rapid MLS is a radiological computer-assisted image processing software device that uses machine learning. It operates as a processing module within the integrated Rapid Platform. Its specific job is to analyze non-contrast CT scans, referred to in the document as NCCT, and provide a measurement of the brain midline. Essentially, it looks for the expected brain midline and measures any shift that might have occurred.
Sam: And the document actually spells out the clinical context for this, right? It names the specific medical scenarios where measuring this shift is a critical metric.
Maya: Yes, it does. Under the clinical characteristics section, the document states that the primary users are medical professionals who use midline shift as a clinical metric to determine the extent of affect of specific indications. The three indications explicitly listed are intracranial hemorrhage, tumor, and TBI, which stands for traumatic brain injury. For clinicians on the ward, knowing the exact measurement of a midline shift is vital for understanding the severity of those conditions.
Sam: So the software takes the non-contrast CT scan, processes it, and then what does the doctor actually see? How is the result delivered?
Maya: The document describes two very specific outputs. First, it provides a visual output. This is an annotated image containing a color overlay that displays the difference between the expected and indicated brain midline, specifically at a location called the Foramen of Monro. Second, it generates a text file output containing the quantitative measurement, and it notes that this text file is in json format.
Sam: Wait, a JSON format text file? That's an interesting detail for an FDA summary to highlight. Why would a doctor care about a JSON file?
Maya: A doctor looking at a screen might not care, but from a health-system or IT perspective, it's highly relevant. JSON is a standard, lightweight data-interchange format. By outputting the quantitative measurement as structured data in a JSON file, rather than just burning a number onto a static image, the software makes it much easier for hospitals to automatically pull that exact measurement into electronic health records, research databases, or clinical workflow dashboards. It shows they are building this to integrate smoothly into complex enterprise IT environments.
Sam: That makes a lot of sense. Now, let's talk about how this device actually got its FDA clearance, because this is where the document gets really interesting. This is a 510(k) clearance, which means iSchemaview had to prove this device is substantially equivalent to an existing, legally marketed device, what the FDA calls a predicate device. I expected the predicate to be another brain CT tool. But that's not what the document says, is it?
Maya: No, it's not. The predicate device listed is the iSchemaview Rapid RV/LV, cleared under K223396. And the Rapid RV/LV software is designed to measure the maximal diameters of the right and left ventricles of the heart from a volumetric CT pulmonary angiography acquisition. So they are using a software tool that measures heart ventricles on chest scans as the legal precedent to clear a tool that measures brain midline shift on head scans.
Sam: That sounds wild to a layperson. How does the FDA accept a heart measurement tool as the foundation for a brain measurement tool? Doesn't that raise safety questions?
Maya: The summary addresses this head-on. It states that Rapid MLS does not raise new questions of safety or effectiveness compared to the previously cleared Rapid RV/LV Software. The FDA 510(k) pathway is focused on intended use and technological characteristics. The document provides a direct comparison table to prove this equivalence. Both devices share the exact same FDA product code, which is QIH, and the exact same regulation number, 21 CFR 892.2050, which covers medical image management and processing systems.
Sam: I see the table here. It really breaks down the similarities. Both use AI/ML software. Both operate via a Command Line interface. Both output Reports and DICOM Secondary Capture Series. The main differences are just the inputs and the specific measurements being made.
Maya: Exactly. The predicate uses non-gated CT Pulmonary Angiography images to measure heart diameter ratios, while the new device uses non-contrast CT images to measure brain midline shift. The applicant acknowledges there are minor differences in intended use and technical characteristics with the predicate device, but concludes that with these minor changes, the clinical use is the same with no additional risks.
Sam: It's a great reminder for medtech companies listening. Your predicate device doesn't always have to look at the exact same organ, as long as the underlying technological principles, like taking an image, finding landmarks with AI, and calculating a distance, are substantially equivalent. But let's move on to the actual performance of the AI. How did they build this, and more importantly, how did they test it?
Maya: Let's start with algorithm development. The document says development was performed using 162 cases from multiple sites. That total was split into 138 cases for training and 24 cases for validation. They selected cases covering offsets from 0 to 18.6 mm from the expected midline. And the scans came from a variety of manufacturers like GE, Toshiba, Siemens, Philips, and PNMS.
Sam: So they trained the AI on 138 cases. To me, that sounds like a relatively small number for training a machine learning model, but they specifically mention making sure the cases spanned a wide range of midline shifts, all the way up to 18.6 mm. But training data is one thing. What about the final performance testing? That's what really matters.
Maya: The final performance validation is much more rigorous. It included 153 non-contrast CT cases. And importantly, the document states this validation data was sourced and blinded independent of the development cases through the data management team to ensure generalizability of the device. Ground truth for these 153 cases was established by 3 experts.
Sam: Independent and blinded validation sets are exactly what we want to see. They also provide a breakdown of where these 153 cases came from. It says they used 13 sites in total, split between 2 OUS, which I assume means Outside the US, and 11 US sites. They did this to test for broad demographic representation and to avoid overlap bias with the development data.
Maya: Yes, and they break down the demographics further. The validation set was split 44% male and 56% female, with an age range spanning from 26 to 93 years. The samples were a mix of scans from GE, Philips, Toshiba, and Siemens scanners. Notice that PNMS is missing from the validation list, though it was in the training set, but having 4 major manufacturers represented in the test set is a solid showing for real-world hospital environments.
Sam: Okay, so they have a solid, independent test set of 153 cases graded by 3 experts. Let's talk about the results. How accurate was Rapid MLS at measuring that midline shift?
Maya: The primary endpoint analysis looked at the mean absolute error, or MAE. The summary states that based on the validation data, the Rapid MLS is non-inferior to the MAE of the experts. Specifically, the Rapid MAE is 0.7 mm, whereas the experts' average pairwise MAE is 1.0 mm.
Sam: Wait, really? The AI's mean absolute error was 0.7 mm, and the human experts were at 1.0 mm? That means the software's margin of error was actually a third of a millimeter smaller than the variations among the three human experts.
Maya: That is what the numbers in the summary show. It supports the claim that the automated measurement is highly precise. The document also provides some statistical modeling to back this up. It mentions a Passing-Bablok fit used to compare Rapid MLS to the reference midline shift. The results provided an intercept of 0.12, with a confidence interval of 0 to 0.2, and a slope of 0.95, with a confidence interval of 0.9 to 1.0.
Sam: For our listeners who might be a bit rusty on their statistics, what does a slope of 0.95 and an intercept of 0.12 actually tell us in this context?
Maya: In a Passing-Bablok regression, which is often used in method comparison studies, a slope of exactly 1.0 and an intercept of exactly 0 would mean the two methods produce perfectly identical results. So a slope of 0.95 and an intercept close to zero means the AI's measurements are highly comparable to the expert consensus across the range of shifts. The summary explicitly says this supports that the methods, meaning Rapid versus Experts, are comparable.
Sam: They also ran a difference plot, right? The document notes there was no bias and gives a paired t-test p-value of 0.1800. It's interesting they wrote it exactly like that, out to four decimal places. But a p-value of 0.1800 means there is no statistically significant difference between the AI's measurements and the experts' measurements.
Maya: Exactly. The performance validation testing demonstrated that the device provides an accurate representation of key processing parameters under a range of clinically relevant conditions. Beyond the clinical accuracy, the company also conducted software verification and validation testing, stating that the device met all design requirements and specifications in compliance with 21 CFR 820.30 requirements for design controls.
Sam: On the topic of design requirements, the summary lists several performance standards the software conforms to. Things like risk management and software lifecycle processes. But one that stood out to me was UL 2900-1 from 2017, which is the Standard for Safety for Cybersecurity Network-Connected Products. It's a reminder that this isn't just a standalone calculator; it's a network-connected medical device handling sensitive images.
Maya: That's a very important point for hospital procurement teams. Any medical image management system handling DICOM files, which the summary notes Rapid MLS complies with under NEMA PS 3.1 through 3.20, needs to have robust cybersecurity validation. The FDA expects it, and the IT departments buying the software demand it.
Sam: So we've got an AI tool that measures brain midline shift with a 0.7 mm error rate, tested on 153 independent cases across multiple scanners, and it plugs into hospital networks outputting DICOM images and JSON files. It sounds impressive, but we also have to look at the limitations. What does the document say about how this can and cannot be used?
Maya: The indications for use section contains a very clear limitation. It states that its results are not intended to be used on a stand-alone basis for clinical decision-making or otherwise preclude clinical assessment of NCCT cases. Furthermore, the prescriptive statement notes that Federal law restricts this device to sale by or on the order of a physician.
Sam: Right, so this is an assistive tool. If a patient comes in with a traumatic brain injury or an intracranial hemorrhage, the AI can rapidly generate the color overlay at the Foramen of Monro and give you a fast, precise number, but the physician is still ultimately responsible for assessing the scan and deciding on the treatment plan.
Maya: Precisely. The summary states that Rapid MLS analyzes adult cases to identify locations and measurements, but it does not evaluate information related to contract liability warranties, nor does this substantial equivalence determination mean the FDA has determined the device complies with other federal statutes. The clearance simply means that, legally and technologically, it is substantially equivalent in safety and performance to the previously cleared Rapid RV/LV software.
Sam: It's a textbook example of how AI companies are expanding their platforms. You validate the core machine learning and image processing infrastructure on one anatomy, like the heart, and once you have that clearance, you use it as a predicate to introduce modules for other anatomies, like the brain, as long as you can prove the technological characteristics and clinical risks are comparable.
Maya: It is an efficient regulatory strategy. And by providing data showing the tool's error rate is 0.7 mm compared to the experts' 1.0 mm on a blinded set of 153 scans, iSchemaview has clearly demonstrated that their software can perform this specific measurement task reliably.
Sam: To recap, we looked at the FDA 510(k) summary for iSchemaview's Rapid MLS, an AI tool cleared to measure brain midline shift on non-contrast CTs. It proved substantial equivalence to a heart measurement tool and showed a mean absolute error of 0.7 mm in independent validation testing, providing quantitative outputs to assist, but not replace, physician judgment in cases of hemorrhage, tumor, and TBI.
Maya: You can find a link to the full public FDA document in our show notes. A final reminder: this podcast is for informational purposes only and is not medical advice. Thanks for listening to Smart Summaries.