15 November 2026 · 14 min

Company Spotlight: Qure.ai qXR-Detect - Triage on routine chest X-rays

Radiology & ImagingFDA Clearances

This episode examines the FDA 510(k) clearance summary for Qure.ai's qXR-Detect, a computer-assisted detection tool for chest X-rays. We explore its clinical performance data, how it impacts reader sensitivity and false positive rates, and a unique regulatory plan that allows future algorithm architecture changes.

Key points

Source: FDA 510(k) summary K251934: qXR-Detect (Qure.Ai Technologies) - U.S. Food and Drug Administration, 2026

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: Qure.Ai Technologies · Product: qXR-Detect · FDA 510(k) K251934

Your company here. This podcast is looking for its first sponsors: reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsorship options and rates →

Transcript

Maya: Welcome to AI in Medicine - Smart Summaries. Today we are looking at an AI tool for chest X-rays that actually dropped the false positive rate per image from 0.4182 to 0.3300 while increasing sensitivity. That is according to the FDA 510(k) summary K251934 for qXR-Detect, published by the U.S. Food and Drug Administration.

Sam: And just a quick reminder before we get started. Maya and I are AI-generated voices. We are here to give you a plain English summary of that public FDA document, not to speak for the company or offer a sales pitch.

Maya: Exactly. This is a Company Spotlight episode. We are diving into the exact testing numbers, the device capabilities, and the limitations outlined in the public clearance summary for qXR-Detect, which is manufactured by Qure.ai Technologies. We are going to look at what this tool actually does, how it was tested, and what this clearance means for health systems evaluating imaging AI.

Sam: So let us start with the basics. What exactly is qXR-Detect and who is it built for?

Maya: The document defines qXR-Detect as a computer-assisted detection software device, or CADe. It analyzes DICOM chest X-ray images and highlights suspicious regions of interest. It is indicated for adults only, and it is explicitly intended to be used as a concurrent reading aid by trained professionals. The summary specifically lists ER physicians, family medicine practitioners, and radiologists as the intended users.

Sam: Concurrent reading aid means the clinician looks at the image and the AI output at the same time, right?

Maya: Yes. The document is very clear that qXR-Detect is an adjunct tool. It says it is not intended to replace a clinician's review of the radiograph or their clinical judgment. Users must not use the generated output as the primary interpretation.

Sam: Okay, so what exactly is it looking for on these X-rays? Does it just flag an image as abnormal, or is it more specific?

Maya: It is quite specific. It categorizes any suspicious region of interest into 6 categories. Those are lung, pleura, bone, Mediastinum & Hila & Heart, hardware, and a catch-all category called other.

Sam: That is a broad net. How did they train the algorithm to recognize all those different categories?

Maya: The summary notes that the underlying algorithm was trained on a large data set of chest X-ray scans. Those scans came from the US, EU, India, Vietnam, and Brazil, and contained various abnormal regions of interest. Though I should note, the document does not break down the exact percentages of how many scans came from each of those regions.

Sam: Still, that is a massive training set. But training is one thing. How did they actually prove to the FDA that it works? Since this is a 510(k) clearance, they had to compare it to something already on the market, right?

Maya: That is right. A 510(k) clearance is based on proving substantial equivalence to a predicate device. In this case, the predicate is Chest CAD, manufactured by Imagen Technologies, Inc., which was cleared under K210666. Both use deep learning to analyze chest radiographs and highlight suspicious regions across similar categories. Both are intended as concurrent reading aids.

Sam: So to prove it is substantially equivalent, they have to run performance testing. Let us get into the real numbers. What did the standalone testing look like?

Maya: They ran non-clinical standalone performance testing against a ground truth. For this dataset, they pulled scans mostly from across the US, spanning 40 states and 5 regions. The gender distribution was approximately 50-50 male-to-female, and the age distribution went from 22 years to over 85 years. They had 3 ground truthers annotate the scans for those suspicious categories.

Sam: And how did the AI perform against those 3 human truthers?

Maya: They measured performance using Area Under the Curve, or AUC. For the lung category, the AUC was 0.893. Pleura was higher at 0.95. Mediastinum and Hila was 0.891, bone was 0.879, hardware was the highest at 0.958, and the other category was 0.915.

Sam: Those are solid numbers for detecting the presence of a finding. But does it actually show the clinician where the problem is on the image?

Maya: Yes, they also tested localization accuracy. For this, they used a metric called wAFROC, which the document notes was mostly above 0.8 across categories, with an aggregate score of 0.839. The document states this indicates the device demonstrates good performance in detecting and localizing the regions of interest.

Sam: Okay, so the standalone test proves the software works in a vacuum against an established baseline. But medicine does not happen in a vacuum. A concurrent reading aid only matters if it actually helps the human reader do a better job. Did they test that?

Maya: They did. And this is the most critical part of the document for clinical leaders. They ran a multireader multicase clinical performance study, often called an MRMC study. They used 301 samples and 18 readers, which included radiologists, ER physicians, and family medicine practitioners.

Sam: How exactly do you set up a study like that? Do the doctors just look at a bunch of images with the AI turned on and then look at a bunch with it turned off?

Maya: It has to be strictly controlled to avoid bias. The document states it was a fully-crossed concurrent design. The readers read scans in 2 sessions. The scans were randomized in blocks. If a scan was read with AI assistance in the first session, it was read unaided in the second session, and vice versa. Crucially, they enforced a washout period of 28 days between the 2 sessions.

Sam: Ah, the 28 days washout period is to make sure the doctors forget the specific X-rays from the first round, so they are not just relying on their memory when they read them the second time.

Maya: Exactly. The primary metric was to see if the AI improved the readers' detection and localization performance, measured by the area under the weighted alternative free-response receiver operating characteristic curve, which the document abbreviates as wAFROC-FOM.

Sam: That is quite a mouthful. What did the actual data show? Did the 18 readers get better with the AI?

Maya: They did. The overall wAFROC significantly improved from 0.6894 in the unaided reads to 0.7505 in the reads aided by qXR-Detect. That is an improvement of 0.0611, with a significant P value of < 0.001. Furthermore, 17/18 readers showed improvement in reads aided by the software across all categories.

Sam: Wait, 17/18 readers? So almost every single doctor improved when using the tool. Did they break that down by what they were looking at?

Maya: Yes, and the results were particularly strong in the lungs. All 18 readers demonstrated improvement in detecting and localizing suspicious lung regions of interest.

Sam: That is impressive. But here is my concern with AI in radiology or the emergency department. Usually, when you give an algorithm high sensitivity, it starts flagging everything. It sees shadows and calls them abnormalities. This creates alert fatigue, and doctors start ignoring it. Did the summary address the false positive rate?

Maya: That is exactly what makes these numbers interesting. The summary specifically notes that the number of false positives per image actually went down. In the unaided reads, the false positives per image was 0.4182. In the reads aided by qXR-Detect, it dropped to 0.3300. The summary explicitly states this indicates that the use of the device does not cause the users to overcall conditions.

Sam: Let me make sure I have this straight. The overall sensitivity of the readers went from 0.8896 unaided, up to 0.9338 when aided. So they caught more real problems. But at the same time, their false positive rate per image dropped from 0.4182 to 0.3300. They became both more sensitive and more precise.

Maya: That is exactly what the clinical performance testing data in the document shows. Overall specificity also increased from 0.5556 unaided to 0.6219 aided.

Sam: Okay, that is a compelling clinical narrative. It explains why they secured the clearance. But I see there is a massive section in this document about something called a Predetermined Change Control Plan, or PCCP. What is that?

Maya: The PCCP is fascinating from a regulatory and business perspective. Historically, if a company changed how its medical AI algorithm worked, it often had to submit a whole new 510(k) to the FDA. Under section 515C of the Act, a new notification is not required if the change is consistent with an established PCCP. It basically pre-clears specific future modifications, as long as the company follows strict testing rules.

Sam: So they told the FDA in advance exactly how they plan to update the software, and the FDA said, as long as you follow this specific plan, you do not need to ask us again. What changes did they pre-clear for qXR-Detect?

Maya: The document outlines 2 modifications. The first is Modification 1, or M1, which is data-driven retraining. The plan says they can retrain the existing model using newly acquired, high-quality adult chest radiograph datasets to improve device performance without altering the underlying model architecture.

Sam: That makes sense. If they get a batch of better data from new hospitals, they can feed it into the model to make it smarter. The document mentions they have to test it against a locked test-set to confirm performance improves without introducing regression. What is the second modification?

Maya: Modification 2, or M2, is a major architectural modification. The document states it involves a controlled architectural change through modular replacement of the existing Convolutional Neural Network encoder, or CNN, with a Vision Transformer-based encoder, while maintaining the remaining components and workflow.

Sam: Wait, they are pre-cleared to swap out the entire foundational architecture of the image analysis from a CNN to a Vision Transformer? That is a huge structural change in how the AI processes visual data. Vision Transformers are the newer generation of image analysis models.

Maya: Yes. The rationale listed in the document is to enhance feature extraction and global context understanding to further improve diagnostic performance. The trigger for this change is if a superior Vision encoder becomes available.

Sam: But how does the FDA ensure that swapping the engine of the AI does not break the tool or make it unsafe?

Maya: The document lists strict risk mitigation and testing methods. For M2, the candidate models trained on the Vision Transformer architecture must undergo rigorous evaluation against the baseline CNN model. They have to assess performance metrics like AUC, box level sensitivity, and specificity. The update is only released if it meets predefined success criteria. Plus, the document emphasizes that before deployment, each update is reviewed to ensure there are no new outputs, indications, or workflow changes.

Sam: And do they just push the update quietly in the background?

Maya: No. The update procedure requires users to be notified through formal communication channels, including revised labeling, updated user manuals, and detailed release notes. The modification is implemented through a controlled, manual deployment process.

Sam: This PCCP framework is a game changer for hospital IT teams. Usually, buying medical AI means you are buying a frozen model that gets outdated quickly. This document shows Qure.ai has built a regulatory runway to continuously deploy better versions of the model, and even upgrade the core architecture, without getting bogged down in repeated FDA submissions.

Maya: It certainly offers a clear pathway for continuous improvement. But as buyers or clinicians, we also need to be clear about what this document does not tell us. For instance, while we know the standalone testing included 40 states, there is no breakdown of race, ethnicity, or specific comorbidities in the testing dataset, which is a common limitation in these high-level summaries.

Sam: Right, and it does not detail what exactly falls into that catch-all other category of abnormal findings. We also do not see any data here on the latency of the cloud deployment, like how many seconds it takes for the image to go to the cloud and come back to the PACS system. In a busy emergency department, workflow speed is everything.

Maya: That is a great point. The summary confirms de-identified X-rays are sent via HTTPS and results are forwarded to the PACS or other radiology systems, but it offers no performance metrics on processing speed.

Sam: So, Maya, for the hospital executives or radiology chiefs looking at this, what is the single biggest takeaway from this 510(k) summary?

Maya: The most memorable takeaway is the clinical study outcome. This device demonstrated that an AI reading aid can increase reader sensitivity from 0.8896 to 0.9338, while simultaneously pushing the false positive rate down from 0.4182 to 0.3300 per image. It is rare to see both sensitivity and precision improve together, and doing so via a model that has clearance to evolve its architecture over time makes this a notable clearance in the chest radiograph space.

Sam: That is a strong place to wrap up. We have covered the indications for use, the MRMC study numbers, and the mechanics of the Predetermined Change Control Plan for qXR-Detect.

Maya: You can find a link to the full FDA 510(k) summary in our show notes. Remember, this podcast is for informational purposes and is not medical advice. Thanks for listening to AI in Medicine - Smart Summaries.