When AI Upgrades Chest CTs: Inside the FDA Clearance of qCT Lung

Sam: Imagine giving a radiologist a tool and seeing their accuracy in identifying lung nodule texture jump from 46.17% to 60.79% just by turning it on.

Maya: That is exactly the kind of performance bump we are looking at today with a newly cleared document from the U.S. Food and Drug Administration, specifically the 510(k) summary for qCT Lung v1.0 by Qure.ai, published on September 14, 2026.

Sam: And before we dive into the exact numbers and what they mean for the boardroom and the reading room, we need to remind everyone that our voices are AI-generated, and this is a summary of a public document.

Maya: Exactly. We are here to unpack the facts exactly as they appear in the filing. So let us start with what this device actually is. qCT Lung is an artificial intelligence-based device that processes non-contrast CT scans. It is built to assist radiologists and pulmonologists specifically with the adult screening population.

Sam: And it is not just finding any random spot on the lung. The document is very specific. It is looking at pulmonary nodules of size 4-30 mm. I find that size window interesting because it is highly specific to the clinical workflow of lung cancer screening. But it goes way beyond just putting a box around a finding, right?

Maya: Right. Detection is only the first step. The FDA clearance highlights three core pillars for this software: detection, characterization, and quantification of the region of interest. When they say characterization, they are talking about three very specific clinical features. First is texture, which the document breaks down into solid, part-solid, and ground glass. Second is spiculation status, which it labels as present or absent. And third is calcification status, also present or absent.

Sam: And then there is the quantification piece, which includes the average diameter in millimeters and the volume in cubic millimeters. It even computes the long axis and the short axis on an axial slice. It is basically automating the most tedious parts of a radiologist's job when they are staring at a chest CT.

Maya: But the document is crystal clear on its boundaries. This is intended to be used as a concurrent reader. The FDA explicitly states it provides adjunctive information and is not intended to be used as a replacement for a complete interpreting physician's review or clinical judgment. It is also not to be used without the original CT series.

Sam: Which makes total sense. You cannot just look at the AI output in a vacuum; you need the original imaging. Speaking of workflow, how does this actually get in front of the doctor? For the health system leaders listening, integration is always the biggest hurdle. Buying the software is easy, but getting it into the clinical workflow is where these projects usually die.

Maya: The filing actually details the deployment options. qCT Lung can be deployed on a secure cloud server or an on-premises server. It interacts with the picture archiving and communication system, or PACS, to receive de-identified scans. Then it presents the results in one of two ways. The first is PACS-based mode. It returns a new series, a secondary capture, to the originating PACS with the segmentation burnt on the series, or as a GSPS overlay if the PACS supports it.

Sam: I love that they included the GSPS overlay option. It means the radiologist can toggle the AI findings on and off rather than having a separate image with permanent markings burned into it. What is the second mode?

Maya: The second is a Web-based mode. This is on Qure's own web application where the secondary capture and visualisations are shown along with output labels for review. Another crucial feature mentioned in the workflow section is that the device segments the nodules initially, but the user can verify the correctness of the segmentation and labels. They can manually correct them using editing tools provided by the application.

Sam: That manual correction capability is massive for clinician adoption. It keeps the physician in the driver's seat. Now, I want to pivot to something that I found completely fascinating from a regulatory strategy perspective. When a company submits a 510(k) to the FDA, they have to prove their device is substantially equivalent to a legally marketed predicate device.

Maya: Yes, they have to show it is just as safe and effective as something already out there.

Sam: Right. So this is a lung CT AI. But the primary predicate device listed in this document is Lunit INSIGHT DBT v1.1. That is a software used for digital breast tomosynthesis. It is a breast cancer screening tool. Maya, how on earth is a breast imaging AI the predicate for a lung nodule AI?

Maya: It is a great question, and it really shows how the FDA categorizes these software functions. The filing explains that both devices are classified under 21 CFR 892.2090. The regulation name is Radiological Computer-Assisted Detection And Diagnosis Software. The product code is QDQ. Even though the anatomical focus is different, breast versus lung, and the modality is different, digital breast tomosynthesis versus non-contrast CT scan, the fundamental technological basis is the same.

Sam: So the FDA is looking at what the software fundamentally does. Both are artificial intelligence-based deep learning algorithms intended to be used concurrently by interpreting physicians to detect and characterize suspicious lesions. It is all about the mechanism of action, not just the body part.

Maya: Exactly. The document notes that both devices share the same intended purpose as a concurrent diagnostic support tool. And to bridge the anatomical gap, the sponsor also included two reference devices: ClearRead CT by Riverain Technologies and Lung Nodule Assessment and Comparison Option by Philips Medical Systems. Both of those process thoracic CT scans and fall under different regulation numbers, but they help prove that the specific lung-related features are well understood.

Sam: It is a really smart regulatory strategy. Okay, let us get into the actual performance of the algorithm. The document details two major testing phases: a standalone study and a clinical performance testing multi-reader multi-case study. Let us start with the standalone. This is where they test the AI against a known ground truth, right?

Maya: That is correct. The standalone study used a total of 670 scans. That was broken down into 462 scans with at least one nodule and 208 without nodules. The reference standard, or ground truth, was established by three American Board of Radiology-certified thoracic radiologists, each with more than 10 years of experience.

Sam: And the results are quite striking. For pure detection, the scan-level AUROC was 0.912. That is the Area Under the Receiver Operating Characteristic curve, a standard measure of performance. A 0.912 is a very robust signal.

Maya: It is. But the characterization metrics are where it gets really granular. For detecting calcification, the AUROC was 0.959. For spiculation, it was 0.851. And for texture, the Macro AUC was 0.934. They even broke the texture down further: AUC Solid was 0.917, AUC Part-solid was 0.935, and AUC GGN, which stands for ground glass nodule, was 0.951.

Sam: Wait, really? A 0.951 for ground glass nodules? That is historically one of the trickier things to nail down consistently. Those numbers show it is highly capable in a vacuum. But what about the quantification? Measuring these things is where inter-reader variability usually goes off the charts.

Maya: The document provides the Bland-Altman analysis for both average diameter and volume. For the average diameter, the mean bias was −0.36 mm. For volume, the mean bias was +2.4%.

Sam: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Maya: And now, back to the document.

Sam: So on average, it slightly underestimates the diameter by just under a third of a millimeter, and slightly overestimates volume by 2.4%. Clinically speaking, a bias of −0.36 mm is practically negligible when you are talking about tracking nodule growth over time. But we should also look at the spread, right? The Limits of Agreement?

Maya: Yes, the filing includes those. For the average diameter, the lower limit of agreement was -3.48 mm and the upper limit was 2.77 mm. For volume, it is a bit wider as a percentage: the lower limit was −41.7% and the upper limit was +79.8%. They also report the MANE, which is the mean absolute normalized error. For average diameter, the MANE was 0.093, and for volume, it was 0.18.

Sam: That volume variance is wider, but volume is notoriously difficult to calculate perfectly, especially on irregular, spiculated nodules. The document concludes that these outputs are accurate and uniform across a wide range of potential sources of measurement error. But an algorithm performing well against a static dataset is one thing. How does it actually change a human doctor's performance? That is the multi-reader multi-case study.

Maya: Exactly. The MRMC study is where we see the real-world impact. This study included 300 cases: 200 diseased scans and 100 non-diseased scans. They had 16 readers, a mix of radiologists and pulmonologists with varying levels of clinical experience. They compared their performance unaided versus aided by qCT Lung.

Sam: And the improvements are substantial. Let us look at detection first. They used wAFROC, which is the weighted alternative free-response receiver operating characteristic. The unaided wAFROC was 0.655. With the AI, the aided wAFROC jumped to 0.763. That is a difference of 0.108.

Maya: That is a solid gain for simply detecting the nodules. But the characterization accuracy is where the AI really acts as a force multiplier. For calcification accuracy, the readers unaided were at 51.85%. Aided, they hit 66.11%. That is a difference of 14.26%.

Sam: It is exactly the same story for the other features. Texture accuracy went from 46.17% unaided to 60.79% aided, a difference of 14.62%. Spiculation accuracy climbed from 45.64% unaided to 60.35% aided, which is a 14.71% difference.

Maya: What is striking to me as a former clinician is how consistent that improvement is across the board. The differences are 14.26%, 14.62%, and 14.71% in accuracy for calcification, texture, and spiculation. The document concludes that this assistance significantly improves performance across all primary and secondary endpoints, with consistent benefits observed across reader specialties and experience levels.

Sam: If you are a health system leader looking at this, you are not just buying a tool that points out a spot. You are buying a tool that fundamentally upgrades the baseline accuracy of your entire reading room, regardless of whether a senior radiologist or a newer pulmonologist is looking at the scan. It enhances diagnostic consistency, which is the holy grail for a department.

Maya: And there is another major component to this clearance that boardroom executives and AI developers need to pay close attention to. Section 11 of the summary details a Predetermined Change Control Plan, or PCCP.

Sam: This is a relatively new FDA framework, right? It essentially allows a manufacturer to pre-clear specific future modifications to the AI without having to submit a brand new 510(k) every single time they update the algorithm.

Maya: Exactly. The filing states that modifications will be made in accordance with the FDA's guidelines on Predetermined Change Control Plans. Specifically, it outlines a modification called M1, which stands for Data-Driven Retraining. The rationale is to improve the algorithm performance. They are allowed to retrain it with newly acquired, high-quality adult CT datasets including scans from lung cancer screening population.

Sam: But there are guardrails. The document explicitly says the algorithm modification will not include any adaptive algorithms that can continuously learn from the environment. The convolutional neural network architecture has to remain unchanged. They basically have to train it, test it, validate it, and then lock it before deploying the change.

Maya: Yes, and the FDA requires them to map out the risks and mitigations for this retraining. Under the risk assessment for this data-driven retraining, the document lists potential overfitting and unintended bias if new data are imbalanced, as well as a risk of minor performance variability.

Sam: And their mitigation strategy is testing the modified model against an independent validation dataset. When they do lock a modification, they have to communicate it to all users through updated labelling via the Instructions for use. This PCCP changes the commercial lifecycle of the software. Instead of the version being frozen in time for years while awaiting new clearances, it can actually get iteratively better at its specific job based on more real-world data.

Maya: It is a massive operational advantage for the manufacturer. But we must also talk about the limitations stated in the document. We already mentioned it is not a replacement for physician judgment. Furthermore, it explicitly states that the device is not intended for clinical diagnosis of any disease. It does not replace the role of physician or of other testing in the standard of care for lung abnormalities.

Sam: That is the critical legal and clinical distinction. It is a computer-assisted detection and diagnosis software, sure, but the diagnosis part is about characterizing the lesion, not diagnosing the patient with lung cancer. It provides adjunctive information. Period.

Maya: Correct. And one more technical note on the output: the device computes specific measurements for each segmented region, like the long axis, which it defines as the longest diameter on an axial slice in millimeters, and the short axis, the longest diameter perpendicular to the long axis on the same slice. It is highly structured data.

Sam: Which is exactly what you need to feed downstream clinical registries or reporting templates without human data entry errors. The structured output is almost as valuable as the detection itself. If I am taking one memorable thing away from this document, it is that consistent jump in characterization accuracy. Moving readers from an unaided accuracy like 45.64% for spiculation up to 60.35% aided is a meaningful clinical shift.

Maya: My main takeaway is the regulatory pathway. Using a breast tomosynthesis software as the primary predicate for a lung CT tool perfectly illustrates how the FDA evaluates the core technology—the convolutional neural network and its function as a concurrent reader—rather than getting distracted by the anatomy.

Sam: It is a smart strategy that other digital health companies should definitely study. To recap, we have been looking at the FDA clearance for Qure.ai's qCT Lung v1.0. It is a class II device cleared to help detect, characterize, and quantify pulmonary nodules on non-contrast CT scans for adult screening populations.

Maya: It proved its effectiveness in a 670-scan standalone study and a 300-case multi-reader study, demonstrating significant improvements in physician accuracy. It also features a Predetermined Change Control Plan allowing for future data-driven retraining.

Sam: As always, you can find a link to the full FDA 510(k) summary document in the show notes so you can read the exact numbers for yourself.

Maya: And remember, we are analyzing a regulatory document. This podcast is for informational purposes only and is not medical advice. Always rely on clinical judgment and standard of care.