Company Spotlight: Ibex Galen Second Read - Catching Missed Prostate Cancer

Sam: When 12 pathologists were tested on a set of prostate biopsies they had initially called benign, using a new AI safety net bumped their sensitivity from 0% up to 36.3%, catching cancers that had completely slipped through the cracks.

Maya: That is the core finding we are looking at today from the U.S. Food and Drug Administration's public 510(k) summary for a device called Galen Second Read, made by a company called Ibex Medical Analytics and cleared on January 24, 2025.

Sam: And before we dive in, a quick reminder that our voices are AI-generated. This is a summary of a publicly available FDA document, not promotional material from the company, and we are independent.

Maya: Exactly. Today is a Company Spotlight episode. We are looking strictly at what the FDA cleared, how the testing was done, and what the actual numbers say about this product's performance.

Sam: So let us set the scene. Pathology labs are dealing with massive volumes of slides, specifically prostate core needle biopsies. These are tedious to review, and the stakes for missing a small focus of cancer are incredibly high.

Maya: Which is exactly why this product exists. The FDA summary classifies Galen Second Read as a "software algorithm device to assist users in digital pathology". It is designed to look at hematoxylin and eosin stained formalin-fixed paraffin embedded tissue.

Sam: Those are the standard H and E slides every lab uses.

Maya: Right. But the indication for use here is highly specific. It is intended to analyze scanned histopathology whole slide images to identify cases initially diagnosed as benign for further review by a pathologist. It is not replacing the first read. It waits until the pathologist says a slide is clear, and then it double-checks their work.

Sam: I love that workflow. It operates entirely in the background. But how does it actually do the analyzing?

Maya: The document states it uses a "deterministic deep convolutional neural network". It is a cloud-hosted software. First, it calculates a tissue ratio to make sure there is enough tissue in the image, and an out-of-focus ratio to ensure the image is not blurry.

Sam: A basic quality control check before it even tries to look for cancer.

Maya: Exactly. If those parameters are within predetermined ranges, it calculates a slide-level score for the likelihood to contain prostate adenocarcinoma. If the slide is classified with a high likelihood based on a predetermined threshold, it flags the slide.

Sam: And what does the pathologist actually see when they get that flag?

Maya: They get an alert in the Galen Second Read user interface. When they open the patient case, they see an AdC score and an AdC heatmap that highlights the specific tissue areas in the whole slide image that are likely to contain cancer. The document notes that the heatmap opacity can be controlled and toggled on and off so it does not obstruct the pathologist's reexamination.

Sam: That toggle feature sounds crucial. You do not want the AI coloring over the very cells you need to look at to make a diagnosis. Now, we should mention that 510(k) clearance is based on a manufacturer proving their product is substantially equivalent to a device already on the market.

Maya: Yes, and the predicate device here is Paige Prostate. But the summary outlines some key differences in how they operate.

Sam: Right, I caught that in the comparison table. Paige Prostate is used during the initial diagnostic review, whereas Galen Second Read is explicitly for male patients initially diagnosed as benign. Also, their outputs are different.

Maya: Correct. Paige Prostate provides a single X and Y coordinate on the image with the highest likelihood of having cancer. Galen provides a heatmap showing the areas of concern.

Sam: So if I am a health system leader looking at this, I am immediately asking about the IT requirements. What does it take to run this?

Maya: The summary lists specific interoperable components. Right now, it is intended to be used with slide images digitized with the Philips Ultra Fast Scanner. For the computer environment, it requires a minimum of 8.0 GB of RAM, a 1 GHz CPU, Google Chrome version 120 or later, and because it is cloud-hosted, it needs internet access with at least 20 Mbps download speed.

Sam: That 20 Mbps requirement is something hospital IT teams will need to flag. Moving large whole slide images to the cloud and back requires rock-solid infrastructure. If your lab network drops, your automated second read drops.

Maya: That is a very practical point. Now, let us look at the performance validation. They ran both analytical and clinical studies.

Sam: I want to hear the real numbers. How accurate is this thing?

Maya: They started with precision testing to see if the device gives consistent results across different scanners and operators. For repeatability, meaning within the same scanner and operator, the overall average of correct calls for positive slides was 97.4%, and for negative slides it was 88.6%.

Sam: And what about reproducibility? If I scan the same slide on a different Philips scanner with a different technician?

Maya: It holds up well. Between different scanners and operators, the overall average of correct calls was 96.6% for positive slides and 86.0% for negative slides.

Sam: Okay, so the algorithm is stable. But the real test is whether it actually helps doctors find cancer they missed without burying them in false alarms.

Maya: Exactly, and that brings us to the two retrospective clinical studies. The first study looked specifically at cases that had been missed by the standard of care. They collected slides from 347 cases across three sites, 2 in the US and 1 outside the US. All of these cases were initially diagnosed as benign.

Sam: So this is the exact scenario the product is built for. What did it find?

Maya: At the slide level, the device showed a sensitivity of 81.0% and a specificity of 91.6% against the ground truth, which was determined by two independent expert pathologists.

Sam: Wait, let us break that down. A True Positive here means the slide actually had cancer according to the experts, and the device flagged it. A False Positive means the slide was truly negative, but the device flagged it anyway.

Maya: Exactly. But things get interesting when you look at the case level rather than just the slide level. Remember, one patient case usually involves multiple slides from different biopsy cores. At the case level, the sensitivity was 80.8%, but the specificity dropped to 46.9%.

Sam: 46.9% specificity at the case level? That means more than half the time when the device flagged a case as suspicious, it turned out to be a false alarm for that patient.

Maya: Yes, and the FDA summary directly addresses this. It notes that there was a decrease in specificity compared to the standard of care, which was 100% specific in this cohort because they had all been called benign initially. The document states this decrease can be managed by mitigation measures, such as the use of additional stains to confirm if the slide or case is positive.

Sam: That is a huge takeaway for lab directors. Yes, you get a powerful safety net, but you have to budget time and resources for those additional stains to rule out the false positives. It is not just a software cost; it is a workflow cost.

Maya: That is exactly the trade-off. Let us move to the second clinical study, which looked at 12 pathologists across 4 sites. They reviewed 772 cases, which included 376 negative cases and 396 positive cases.

Sam: And how did they structure this to make it fair?

Maya: They used two arms. Arm A was the standard of care, reading slides digitally. Arm B was the Galen Second Read workflow. Crucially, there was a washout period of two weeks between the arms to minimize recall bias.

Sam: So they forget what they saw two weeks ago. Smart. What were the combined results for all 12 pathologists?

Maya: With the standard of care, their combined sensitivity was 90.5%. When using Galen, it jumped to 93.9%. That is a statistically significant improvement of 3.5%.

Sam: An extra 3.5% of cancers found across a massive health system is a lot of patients getting the right diagnosis. But again, what happened to the specificity?

Maya: It dropped. Specificity without the device was 91.1%. With Galen, it was 87.9%. That is a difference of negative 3.2%.

Sam: So we are seeing a consistent pattern. The AI pushes the human toward higher sensitivity, catching more cancer, but at the cost of calling a few more benign things suspicious.

Maya: Yes, and the individual variability is fascinating. The summary provides a breakdown by pathologist. For example, Pathologist 12 saw a massive 7.6% improvement in sensitivity, but an 11.8% drop in specificity. Meanwhile, Pathologist 4 had a 0.0% change in both sensitivity and specificity.

Sam: That is wild. Pathologist 4 completely ignored the AI, or they were just perfectly in sync with it. It shows that AI does not standardize humans; humans still interact with AI in their own unique ways.

Maya: But the most critical table in the entire document is Table 9. This isolates just the slides that the pathologists initially assessed as benign. This is the exact intended use population.

Sam: Right, the ones that slipped through the initial net. This is where we got that 36.3% number from the beginning of the episode.

Maya: Exactly. For slides initially diagnosed as benign, the sensitivity without the device was 0%. With the Galen device, the sensitivity of the pathologists was 36.3%. The specificity was 96.5%, meaning a 3.5% decrease from the 100% specificity of the standard of care.

Sam: Finding over a third of the cancers that a human just called benign is a remarkable safety net. But as we wrap up, we need to talk about what this document does not show.

Maya: Right. The summary notes that all the clinical data was based on retrospectively collected samples. It does not show how this performs in a prospective, real-time clinical environment.

Sam: It also does not explain why certain false positives happen, just that they do and that labs should use stains to mitigate them. And, as we mentioned, it is tightly coupled to the Philips Ultra Fast Scanner right now.

Maya: Finally, the FDA makes it explicitly clear in the indications for use: Galen Second Read outputs are not intended to be used on a standalone basis for diagnosis, to rule out prostatic AdC, or to preclude pathological assessment according to the standard of care.

Sam: It is an assistant, not the primary doctor. The final call still belongs to the pathologist.

Maya: To recap, the FDA 510(k) summary for Ibex's Galen Second Read outlines a cloud-based AI that acts as an automated double-check for prostate biopsies diagnosed as benign. It demonstrably increases cancer detection sensitivity by catching missed cases, though labs must be prepared to manage a corresponding slight decrease in specificity.

Sam: You can find a link to the full FDA 510(k) summary in our show notes. And as always, this podcast is for informational purposes only and is not medical advice. Thanks for listening.