Company Spotlight: Viz.ai Viz Subdural+ - Quantifying Brain Scans with AI

Sam: Imagine an AI algorithm that can map out a collection of fluid in the brain with a 73% overlap with human experts, but at the same time, has a standard deviation of 13.91 when calculating the volume. What do those numbers actually mean for the emergency room?

Maya: Today we are looking at exactly that. We are unpacking the official FDA 510(k) summary for a device called Viz Subdural+, published by the U.S. Food and Drug Administration.

Sam: And before we jump into the data, a quick reminder that our voices are AI-generated, and this is a summary of a publicly available regulatory document, not promotional material.

Maya: Right. This document outlines the clearance for Viz Subdural+, which is also referred to in the text as Viz SUBDURAL PLUS. It's a software-only device made by Viz.ai, Inc.

Sam: So, what exactly is this software designed to do? The brain is obviously a high-stakes area.

Maya: The document states it is intended for the automatic labeling, visualization and quantification of collections in the subdural space. It uses non-contrast head CT images, or NCCT scans, to do this.

Sam: Just to be clear, when it says collections in the subdural space, we are talking about the area between the brain and the skull where fluids or blood can pool after trauma, right?

Maya: The summary doesn't give us an anatomy lesson, it just strictly refers to them as collections in the subdural space or subdural region. But practically speaking, yes, identifying these collections on a scan is a critical task. The document notes that this software is intended to automate the current manual process of identifying, labeling and quantifying the volume of these collections.

Sam: Automating a manual process. That is the classic pitch for hospital AI. Tracing the edges of a complex shape on dozens of CT slices by hand takes time. What exactly does the AI give the doctor instead?

Maya: It provides several specific outputs. It reports the grayscale value of the collection, the widest width of the subdural collection, and the midline shift. It outputs all of this in a DICOM format, which means it feeds right into the hospital's existing imaging systems.

Sam: Okay, let's talk about how it actually works in the hospital workflow. The summary has a whole section on device description. Does the doctor have to click a button to send the scan to the AI?

Maya: No, it's designed to be seamless. The text explains that images are automatically forwarded from the scanner to Viz.ai's Backend Server. The software then automatically analyzes the applicable scans and sends the results to a destination like a PACS server.

Sam: So a radiologist or neurosurgeon opens the patient's file on their usual viewer, and the AI's analysis is already sitting there waiting for them.

Maya: Exactly. The results come back as a summary series and a segmentation series. The summary series is basically a tabular snapshot. It lists each subdural collection, its volume, its widest width, the total volume, and the midline shift.

Sam: And the segmentation series? I assume that's the visual overlay?

Maya: Yes, it shows an RGB overlay on the original slices. And there is a really interesting technical detail here. The overlay's color intensity actually corresponds to the original Hounsfield Unit values, or HU values, of the original image.

Sam: Wait, really? So it's not just a flat, single-color highlight drawn over the scan?

Maya: Right. The document specifically points out that the overlay opacity, or intensity, corresponds to the grayscale value of the collection within the native NCCT.

Sam: That is a smart detail for clinicians. It means they can still perceive the density differences within the collection even with the AI's color overlay turned on. So, how did this actually get cleared by the FDA?

Maya: It was cleared as a Class II medical device under the 510(k) pathway. That means they had to prove it is substantially equivalent to a device that is already legally marketed.

Sam: The predicate device. What did they compare it to?

Maya: They compared it to their own previous product, the Viz HDS, under application number K232363.

Sam: Did the Viz HDS also look at subdural collections?

Maya: It did not. The document states that the Viz HDS was indicated for identifying intracranial hyperdensities and lateral ventricles. The new device, Viz Subdural+, is focused on subdural collections. Both devices, however, are cleared to measure midline shift.

Sam: So this is an expansion. They took the architecture from their older device and trained a new locked artificial intelligence machine learning algorithm specifically for the subdural space.

Maya: Exactly. They both use deep-learning convolutional neural networks, but the specific algorithm for the subject device is different because it is targeting a different physiological structure.

Sam: Alright, let's get into the performance testing. If you are a hospital leader buying this, or a doctor relying on it at two in the morning, you need to know how accurate it really is. What kind of study did they run?

Maya: They ran a retrospective study to assess standalone performance. They compared the AI's output against a ground truth that was established by trained neuroradiologists.

Sam: Retrospective is standard for these imaging clearances, but it does mean it wasn't tested live in a prospective clinical trial. How big was the data set?

Maya: For measuring the subdural collection volume and maximum thickness, they used a dataset of 203 cases. For the midline shift, they used a dataset of 151 cases.

Sam: And where did the scans come from?

Maya: The summary notes that each dataset was obtained from two clinical sites.

Sam: Just two clinical sites? That is a limitation worth noting. When you only train or test an AI on images from two hospitals, you always have to wonder how it will handle the quirks of different scanners or patient populations elsewhere.

Maya: It is a fair point. The document does mention that additional stratification was provided by different patient demographic, technical and radiographic findings to demonstrate generalizability, but it doesn't give us the specific numbers or demographics in this summary.

Sam: Okay, let's look at the actual results. You teased some numbers at the start of the show. How did it do on measuring the volume of these collections?

Maya: For the 203 cases, the Mean Absolute Error, or MAE, for volume was 7.53, with a 95% confidence interval of 5.60 to 9.45.

Sam: So on average, it's off by about 7.53. But what was that standard deviation you mentioned earlier?

Maya: The standard deviation was 13.91.

Sam: A standard deviation of 13.91 on a mean of 7.53 is huge! That tells me that while it might be close on average, there are some significant outliers where the algorithm is missing the mark by quite a bit.

Maya: That variance is exactly why the median error is a helpful metric here. The median MAE was much lower, at just 2.70. And the range from the 10th to the 90th percentile was 0.0 to 22.22.

Sam: Got it. So half the time it is within 2.70 of the human experts, but in the worst ten percent of cases, it can be off by 22.22 or more. What about the DICE score?

Maya: The mean DICE score was 73%, with a confidence interval of 68% to 77%.

Sam: For those who aren't familiar, a DICE score measures spatial overlap. So a 73% overlap between the AI's drawing and the human expert's drawing. That's solid, but it again shows that the AI and the humans aren't drawing perfectly identical borders.

Maya: Right. Edges can be ambiguous, even for human experts. The testing also looked at the maximum, or widest, thickness of the collections. On those same 203 cases, the Mean Absolute Error was 1.77.

Sam: 1.77. That seems much tighter than the overall volume measurement.

Maya: It is. The 95% confidence interval for thickness was 1.24 to 2.30, with a standard deviation of 3.84 and a median error of 0.43.

Sam: Okay, so the thickness measurement is generally very close, median of 0.43, but again a standard deviation of 3.84 means there are a few wide misses. What about midline shift? That's when pressure pushes the brain off its center axis.

Maya: For midline shift, tested on 151 cases, the mean absolute error was 1.1, with a 95% confidence interval of 0.94 to 1.27. The standard deviation was 1.03.

Sam: And the median for midline shift?

Maya: The median was 0.8, with the 10th to 90th percentile ranging from 0.17 to 2.37.

Sam: That seems like a strong performance for midline shift. Especially since the document says it had to match the performance limits of the predicate device. What was that limit?

Maya: The text states the algorithm can measure midline shift within the same performance limits as the predicate, which it defines as an MAE < 2mm.

Sam: So with an MAE of 1.1, it's comfortably under that 2mm threshold. But when we look at the whole picture, the outliers in volume, the 73% DICE score, it becomes really clear why the FDA includes specific warnings about how the tool is used.

Maya: Absolutely. The indications for use section states explicitly: The device output should be reviewed along with the patient's original images by a physician qualified to interpret brain CT images.

Sam: It is a copilot, not an autopilot. It's meant to save the doctor time by automating the manual tracing, but the doctor still has to verify the work. Are there any other major limitations called out in the summary?

Maya: There is a very specific condition about timing. The document repeats multiple times that Viz Subdural+ provides volumes from NCCT images acquired at a single time point.

Sam: A single time point. So it's not automatically looking at yesterday's scan, comparing it to today's scan, and calculating the exact rate of expansion?

Maya: Correct. It evaluates one scan in isolation. If a doctor wants to compare changes over time, they would have to look at the AI's output for scan A, and then look at the AI's output for scan B, and make that comparison themselves.

Sam: That's a really important distinction for hospital leaders evaluating these tools. It doesn't automate the longitudinal tracking. So, taking a step back, what is the big takeaway here for companies building AI and the systems buying it?

Maya: For companies building AI, this is a textbook example of expanding an existing platform. Viz.ai took an established workflow, their backend server routing images to PACS, and swapped out the algorithm to target a new clinical finding, subdural collections, proving substantial equivalence to their own prior tool.

Sam: And for the hospitals buying it, the takeaway is to read the data tables closely. A median error of 2.70 sounds great, but a standard deviation of 13.91 means your clinicians still need to be sharply reviewing those overlays, because the AI will occasionally get it wrong.

Maya: To recap: today we reviewed the FDA 510(k) summary for Viz Subdural+, an AI tool for automating the measurement of subdural collections and midline shift on non-contrast CTs. While it speeds up manual tracing, the FDA requires a qualified physician to review the outputs due to variance in the performance data.

Sam: You can find a link to the source document in the show notes. And as always, remember that this is not medical advice. Thanks for listening!