# AI lung cancer detection on CT scans: how it works, the evidence and the limits

> AI lung cancer detection finds and scores nodules on low-dose CT and predicts future risk. How it works, the key studies, the data behind it and its limits.

- URL: https://computese.com/ai-tools-diagnosing-lung-cancer/
- Author: Duong Quan Nguyen, CEO, Computese
- Published: 2024-06-30
- Updated: 2026-09-25
- Topics: AI & automation

## In short
- AI lung cancer detection software reads a low-dose chest CT with the radiologist: it finds and measures nodules, scores how likely each is to be cancer and, in research models such as Sybil, predicts a person's risk over six years.
- Screening is for people with a heavy smoking history: in the US, adults aged 50 to 80 with 20 pack-years who smoke or quit within 15 years (USPSTF, 2021). England and the Canadian provinces set their own criteria.
- Retrospective results are strong, such as a 94.4% AUC for Google's 2019 model on NLST scans. In a 2026 randomized trial, AI-assisted reading doubled follow-up CT recommendations and saved no reading time.
- Models learned mostly from NLST, whose participants were all heavy smokers and 91% white. Sybil predicted future cancers poorly in Korean never- or light-smokers in one 2025 study and well in another cohort in 2026.
- FDA-cleared lung CT tools are aids to a radiologist, each for a labelled use, and some are not indicated for screening at all. The clinician who reports the scan remains responsible for the call.

AI lung cancer detection is software that reads a low-dose chest CT scan alongside a radiologist. It finds and measures lung nodules, scores how likely each one is to be cancer and, in research models such as MIT and Mass General's Sybil, estimates a person's risk of lung cancer over the next six years from a single scan.

The software does not make the diagnosis: a radiologist reads the scan and writes the report, and only tissue from a biopsy or surgery confirms cancer. This guide covers who lung cancer screening is for in the US, England and Canada, how the AI works, the studies behind it from Google's 2019 model to randomized evidence from 2026, the data the models learned from, what the FDA has cleared and where the technology still falls short. It is general health information, not medical advice. AI in [screening mammography](https://computese.com/ai-in-breast-cancer-detection-and-treatment/), in [drug discovery](https://computese.com/the-role-of-ai-in-accelerating-drug-discovery/) and in [consumer health apps](https://computese.com/how-ai-is-transforming-health-and-wellness/) is covered in separate guides, and the materials-discovery half of the original article now sits in our roundup of [AI scientific discoveries](https://computese.com/ai-enabled-scientific-discoveries/).

## Who lung cancer screening is for

Screening means scanning people who have no symptoms, so it is offered only where the benefit clearly outweighs the harm: to people with a long smoking history. The criteria differ by country, as of September 2026:

| Where                                | Who is offered screening                                                                                                                                                             | How often                                         |
| ------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------- |
| United States (USPSTF, March 2021)   | Adults aged 50 to 80 with at least 20 pack-years who smoke now or quit within the past 15 years                                                                                      | Every year, stopping 15 years after quitting      |
| England (NHS lung health checks)     | People aged 55 to 74 who have ever smoked and are registered with a GP, in areas where checks run; those found at higher risk are offered a CT scan                                  | Every 2 years for people at high risk (2023 plan) |
| Canada (national guideline, 2016)    | Adults aged 55 to 74 with at least 30 pack-years who smoke now or quit less than 15 years ago                                                                                        | Every year, up to three consecutive times         |
| Canadian provinces (as of Feb. 2026) | Varies: Ontario takes ages 55 to 80 with 20 or more years of daily smoking and a PLCOm2012 risk of at least 2%; British Columbia ages 55 to 74 with a six-year risk of at least 1.5% | Set by each program                               |

A pack-year is one pack of 20 cigarettes a day for a year. The [USPSTF](https://www.uspreventiveservicestaskforce.org/uspstf/recommendation/lung-cancer-screening) lowered its thresholds from 55 years and 30 pack-years in 2021. In England, the [NHS](https://www.nhs.uk/tests-and-treatments/lung-cancer-screening/) invites people through their GP record to a lung health check, and the [2023 national rollout](https://www.gov.uk/government/news/new-lung-cancer-screening-roll-out-to-detect-cancer-sooner) aims to cover everyone eligible by March 2030. In Canada, the [Canadian Task Force on Preventive Health Care](https://canadiantaskforce.ca/guidelines/published-guidelines/lung-cancer/) guideline dates from 2016, and the provinces now run their own programs: according to the [Canadian Partnership Against Cancer](https://www.partnershipagainstcancer.ca/topics/lung-cancer-screening-canada/guidelines/), every province with a program selects people with a version of the PLCOm2012 risk prediction model.

The evidence behind all of them starts with the [National Lung Screening Trial](https://pmc.ncbi.nlm.nih.gov/articles/PMC4356534/) (NLST). From 2002 to 2004 it enrolled 53,454 people at high risk at 33 US medical centres and randomly assigned them to three yearly low-dose CT scans or chest X-rays. Published in the New England Journal of Medicine in 2011, it found 20% fewer lung cancer deaths in the CT group. It also showed the price: 24.2% of CT screens were positive, and 96.4% of those positives were not cancer. The USPSTF counts false positives, overdiagnosis, radiation and incidental findings among the harms; a single low-dose scan delivered 0.65 to 2.36 mSv in the studies it reviewed, against about 2.4 mSv of natural background radiation a year in the US.

> [!NOTE]
> Screening is for people without symptoms. The NHS tells anyone with possible symptoms, such as a cough lasting more than three weeks, coughing up blood or unexplained weight loss, to see a GP rather than wait for a lung health check.

## What a radiologist looks for on a screening CT

A chest CT is a stack of thin cross-sectional images that together form a 3D volume of the lungs. The radiologist scrolls through it looking for pulmonary nodules: small, roughly round spots of tissue. Most are harmless. In the [Pan-Canadian Early Detection of Lung Cancer Study](https://pmc.ncbi.nlm.nih.gov/articles/PMC3951177/) (PanCan), 1,871 people had 7,008 nodules on their first scan, and only 102 of those nodules were malignant.

For each nodule the reader records its size, its density (solid, part-solid or ground-glass), its shape and edges, and whether it has grown since the last scan. The American College of Radiology's Lung-RADS system turns those findings into a category with a suggested next step. Categories 3 and 4 count as a positive screen, and the USPSTF notes that using Lung-RADS may lower the false-positive rate.

How a nodule is measured changes the result. The NLST called a scan positive on a nodule's maximum diameter; the Dutch-Belgian NELSON trial measured volume and added an "indeterminate" category that was not counted as positive. The USPSTF attributes most of the gap in their positive predictive values, 3.3% for NLST against 43.5% for NELSON, to those different definitions and protocols. Readers also disagree with each other: [Google's 2019 paper](https://www.nature.com/articles/s41591-019-0447-x) named inter-grader variability, together with high false-positive and false-negative rates, as the problems it set out to address.

## How AI reads a lung CT scan

Lung CT software does one or more of four jobs. They are different problems with different evidence, and a tool cleared for one is not cleared for the others.

### Finding and measuring nodules

Computer-aided detection (CADe) marks candidate nodules so the radiologist can accept or reject each one. Recent systems use deep learning, such as convolutional neural networks that read the scan as a 3D volume rather than slice by slice, trained on scans in which radiologists have marked the nodules. Most products also measure each nodule's diameter and volume, and many compare it with the previous scan to calculate growth. The FDA labels describe these tools as aids used during the reading: [Fujifilm's Synapse Lung Nodule AI](https://www.accessdata.fda.gov/cdrh_docs/pdf25/K254075.pdf), cleared in May 2026, is labelled as a "concurrent reader" that helps the radiologist detect nodules in the screening population, and not for telling benign from malignant.

![A stack of CT slices passes through a box with a gear to a monitor showing both lungs, with four nodules circled, three grey and one orange. Check and cross buttons below let the reader accept or reject each.](https://computese.com/images/blog/ai-tools-diagnosing-lung-cancer/detect.e16c92130e-1536.webp)

*Detection software proposes marks. The reader decides which ones are real nodules, and a missed mark is still the reader's miss.*

### Scoring the risk that a nodule is cancer

Computer-aided diagnosis (CADx) goes a step further and estimates how likely a nodule is to be malignant. The baseline it has to beat is statistical. The Brock model, built from the PanCan cohort and published in 2013, combines age, sex, family history, emphysema and the nodule's size, type, location, count and spiculation into a probability, and it discriminated well (AUC above 0.90) when validated on a separate British Columbia cohort. A deep learning score learns its features from the image instead. In a [2020 Thorax study](https://pmc.ncbi.nlm.nih.gov/articles/PMC7231457/) of 1,397 incidentally found nodules of 5 to 15 mm from three UK hospitals, Optellum's lung cancer prediction CNN (LCP-CNN) reached an AUC of 89.6% against 86.8% for the Brock model, and placed 24.5% of nodules below the lowest-scoring cancer, against 10.9% with Brock. Several of the authors were Optellum employees, which is common in this field and worth checking in any validation study you read.

![Two CT slices show the same lung nodule on an earlier and a later scan, larger the second time. An arrow leads on from the later scan to a dial whose orange needle points toward the high end.](https://computese.com/images/blog/ai-tools-diagnosing-lung-cancer/score.073004eea5-1536.webp)

*A risk score turns size, shape and change into one number, and that number is only as good as the population it was calibrated on.*

### Predicting future risk from one scan

[Sybil](https://pmc.ncbi.nlm.nih.gov/articles/PMC10419602/), developed by [MIT's Jameel Clinic, the Mass General Cancer Center and Chang Gung Memorial Hospital](https://news.mit.edu/2023/ai-model-can-detect-future-lung-cancer-0120) in Taiwan, answers a different question. It does not ask whether a given nodule is cancer; it estimates how likely the person is to be diagnosed with lung cancer in each of the next six years. It needs one low-dose CT, no clinical data and no radiologist annotations, and it can run in the background on a reading station. Published in the Journal of Clinical Oncology in January 2023, it predicted cancer within one year with an AUC of 0.92 on held-out NLST scans, 0.86 at Mass General and 0.94 at Chang Gung, and its six-year concordance index ranged from 0.75 to 0.81. The model and annotations are public. Its authors see it as a route to personalized screening; the USPSTF, for its part, bases eligibility on age and smoking history because it found insufficient evidence that risk-model-based screening improves outcomes.

### Ruling out normal scans

The newest use is triage: if software can reliably call a baseline scan negative, a radiologist need not read it. Two 2025 analyses of European screening-trial scans tested the idea. In 3,678 baseline scans from the [4-IN-THE-LUNG-RUN trial](https://pubmed.ncbi.nlm.nih.gov/39754864/), the AI software (Coreline's AVIEW LCS) called 31 scans (0.8%) negative that had a nodule over 100 mm³, against 407 (11.1%) for the radiologists. In 1,252 baseline scans from the [UK Lung Cancer Screening trial](https://pubmed.ncbi.nlm.nih.gov/40022836/) (UKLS), AI detected all 31 cancers, though it classed one as negative because it fell under the 100 mm³ volume threshold, and it could have removed up to 79% of the reading workload. Both estimated what would have happened; neither let the AI actually skip a scan.

## The key studies, from 2011 to 2026

| Year | Study (journal)                                      | Data                                                           | Main result                                                                                                           |
| ---- | ---------------------------------------------------- | -------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
| 2011 | NLST (New England Journal of Medicine)               | 53,454 people at 33 US centres                                 | Low-dose CT cut lung cancer deaths by 20% against chest X-ray                                                         |
| 2013 | PanCan, the Brock model (NEJM)                       | 1,871 people with 7,008 nodules, validated in British Columbia | Patient and nodule features predict malignancy with an AUC above 0.90                                                 |
| 2019 | Google (Nature Medicine)                             | 6,716 NLST cases and 1,139 independent cases                   | 94.4% AUC; with no prior scan, fewer false positives (11 points) and false negatives (5 points) than six radiologists |
| 2020 | LCP-CNN (Thorax)                                     | 1,397 incidental nodules from three UK hospitals               | AUC 89.6% against 86.8% for the Brock model                                                                           |
| 2023 | Sybil (Journal of Clinical Oncology)                 | Built on NLST; tested at NLST, Mass General and Chang Gung     | One-year AUC 0.86 to 0.94; six-year C-index 0.75 to 0.81                                                              |
| 2025 | 4-IN-THE-LUNG-RUN (European Journal of Cancer)       | 3,678 baseline scans                                           | As first reader, AI missed fewer significant nodules than radiologists (0.8% against 11.1%)                           |
| 2025 | UKLS (European Journal of Cancer)                    | 1,252 baseline scans                                           | All 31 cancers detected; up to 79% less reading                                                                       |
| 2025 | Sybil external test (Radiology)                      | 18,057 Korean health-checkup participants                      | Poor for future cancers in never- or light-smokers (AUC 0.56)                                                         |
| 2026 | Randomized trial (American Journal of Roentgenology) | 911 people at one centre in South Korea                        | More nodules reported, twice the follow-up CT recommendations, no time saved                                          |
| 2026 | Sybil in an Asian screening cohort (Chest)           | 21,087 people, over half of whom never smoked                  | Six-year AUROC 0.74 overall and 0.79 in never-smokers                                                                 |

The pattern is strong numbers on retrospective data, then more modest and mixed results in new populations and live reading. Even Google's model only matched the same radiologists when a prior scan was available.

Randomized evidence from real reading is only starting to arrive, and it is sobering. In a single-centre [randomized trial in South Korea](https://pubmed.ncbi.nlm.nih.gov/41848489/), published in the American Journal of Roentgenology in 2026, 911 people having a self-requested health checkup in 2025 were read by ten thoracic radiologists with or without a commercial AI tool built into the PACS viewer. With AI, radiologists reported Lung-RADS-positive nodules in 16.9% of scans against 10.3% without, and recommended follow-up CT twice as often (15.3% against 7.4%). Reading time did not fall (187 against 172 seconds, a difference that was not significant). Nobody in either group was diagnosed with lung cancer during about seven months of follow-up, so the trial cannot yet say whether the extra findings mattered.

## The data behind the models, and why its make-up matters

Most lung CT models learned from, or were tested on, the same few datasets. Knowing what is in them explains much of what the models can and cannot do.

- **NLST.** All 53,454 participants were current or former heavy smokers aged 55 to 74 with at least 30 pack-years. According to the trial's [baseline report](https://pmc.ncbi.nlm.nih.gov/articles/PMC2994863/), 91% were white, 4.4% Black and 1.7% Hispanic or Latino, and they were younger and better educated than the eligible US population. The Google model and Sybil were both built on NLST scans.
- **LIDC-IDRI.** The [Lung Image Database Consortium and Image Database Resource Initiative](https://pmc.ncbi.nlm.nih.gov/articles/PMC3041807/) collection holds 1,018 thoracic CT cases, each annotated by four experienced thoracic radiologists in two rounds: first blind, then seeing each other's anonymized marks. Of the 2,669 lesions that at least one reader marked as a nodule of 3 mm or more, all four agreed on only 928 (34.7%). The "right answer" a model is trained on is itself a judgment.
- **Private hospital data.** Google also used a licensed dataset from Northwestern Medicine that is not public, and did not release its training code. Sybil's model and annotations are public, and independent teams have since tested it.

Who is in the data shapes who the model works for. Sybil's developers noted that their training data held only smokers, because the NLST required it, and they tested it on a Taiwanese set that included people who never smoked. Independent tests since then disagree. In 18,057 Korean health-checkup participants, a [2025 Radiology study](https://pmc.ncbi.nlm.nih.gov/articles/PMC12405708/) found Sybil flagged cancers already visible on the scan well (AUC 0.89 in never- or light-smokers) but predicted cancers not yet visible poorly in that group (AUC 0.56, close to chance). In 21,087 people from another Korean screening cohort, a [2026 Chest study](https://pubmed.ncbi.nlm.nih.gov/42140481/) found a six-year AUROC of 0.79 in never-smokers, but weaker results for subsolid nodules, for scans with no nodule at baseline and, in the short term, for scans showing the traces of old granulomatous disease. Same model, different populations, different answers.

![A stack of mostly identical scan cards trains a gear. The gear then reads two groups of cards: a check mark sits over the group like the training cards, an orange warning triangle over the other.](https://computese.com/images/blog/ai-tools-diagnosing-lung-cancer/cohort.35f84bc210-1536.webp)

*A model is proven first on people like those it learned from. People it has never seen are where it fails quietly.*

The same question applies to who gets screened at all. When the USPSTF widened its criteria in 2021, it noted the change may partly reduce racial disparities in eligibility: in one cohort, 17% of Black people who smoke were eligible under the 2013 criteria against 31% of white people who smoke. Canadian provinces even differ on whether their eligibility model includes race: Ontario and Quebec use a version without it, British Columbia and Alberta versions with it. Our guide to [fairness in AI decision-making](https://computese.com/improving-fairness-in-ai-decision-making/) covers how to check whether a model's errors fall unevenly across groups.

How the scans were made matters too. CT scanners, reconstruction filters and slice thicknesses vary between sites. On NLST scans, Sybil's performance held when filters or slice thicknesses changed one at a time, but the best and worst combinations differed significantly at years two to four, according to a [2025 Investigative Radiology study](https://pubmed.ncbi.nlm.nih.gov/39437009/).

## What the FDA has cleared, and what clearance means

The FDA publishes a [list of AI-enabled medical devices](https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices) authorized in the US. At its update of September 4, 2026, with decisions up to June 29, 2026, it held 1,614 devices, 1,230 of them with radiology as the lead review panel. Lung CT tools sit under a few product codes: computer-aided detection for lung CT (OEB), and computer-assisted detection or diagnosis software for lesions suspicious for cancer (POK and QDQ). A sample from the list, with what each is labelled for:

| Device (maker)                            | FDA 510(k)              | Type                    | Labelled for                                                                                              |
| ----------------------------------------- | ----------------------- | ----------------------- | --------------------------------------------------------------------------------------------------------- |
| ClearRead CT (Riverain Technologies)      | K161201, September 2016 | Detection (CADe)        | Helping the radiologist detect nodules on chest CT in an asymptomatic population                          |
| Optellum Virtual Nodule Clinic (Optellum) | K202300, March 2021     | Diagnosis (CADx)        | Scoring incidentally found solid and semi-solid nodules of 5 to 30 mm; not indicated for screening        |
| RevealAI-Lung (RevealDx)                  | K251769, January 2026   | Diagnosis (CADx)        | A relative risk score for incidental nodules of 6 to 15 mm; not indicated for screening                   |
| eyonis LCS (Median Technologies)          | K251474, February 2026  | Detection and diagnosis | Detecting and scoring 4 to 30 mm solid and part-solid nodules on screening low-dose CT, read concurrently |
| Synapse Lung Nodule AI (Fujifilm)         | K254075, May 2026       | Detection (CADe)        | Helping the radiologist detect nodules in the screening population; not for benign versus malignant       |

This is a sample to show how labels differ, not a complete list and not a recommendation. The 510(k) summaries for [ClearRead CT](https://www.accessdata.fda.gov/cdrh_docs/pdf16/K161201.pdf), [Optellum Virtual Nodule Clinic](https://www.accessdata.fda.gov/cdrh_docs/pdf20/K202300.pdf), [RevealAI-Lung](https://www.accessdata.fda.gov/cdrh_docs/pdf25/K251769.pdf) and [eyonis LCS](https://www.accessdata.fda.gov/cdrh_docs/pdf25/K251474.pdf) give the full labelled use. Three things follow from them:

1. **The label decides the use.** A malignancy score cleared for nodules found on a CT done for another reason is not cleared for screening.
2. **Every label describes an aid.** Median's summary states that its output is not intended to replace the interpreting physician's judgement; the others describe support for a radiologist's or clinician's reading.
3. **Clearance is not proof of benefit.** A 510(k) clearance means the FDA found the device substantially equivalent to one already on the market for that use. It is not a finding that patients do better. Sybil was not on the list as of that update; it remains a research model.

Europe looks similar. A [2026 review in European Radiology](https://pmc.ncbi.nlm.nih.gov/articles/PMC13451236/) counted 16 CE-marked lung nodule products from 16 vendors: 14 detect and measure solid and subsolid nodules, 9 estimate malignancy risk (5 with the PanCan model, 4 with their own AI scores), and none handles endobronchial or cystic lesions. Of the 60 peer-reviewed studies behind them, 7% were prospective and none reported patient outcomes.

## Where AI still falls short

**False positives can go up, not down.** AI lowered them on retrospective data, as Google's model did, but in live reading the 2026 randomized trial doubled follow-up recommendations. The UK's [NIHR review of AI nodule software](https://www.journalslibrary.nihr.ac.uk/hta/JYTW8921) (Health Technology Assessment, May 2025) found the same tension: AI may detect more cancers and more harmless nodules alike, and so keep more people under CT surveillance unnecessarily. All 27 studies it included were at high risk of bias, 24 used retrospective data, and it found no direct evidence on clinical outcomes or cost-effectiveness. In up to half of nodules, the automatic size measurement needed manual adjustment.

**Performance travels badly.** As the Sybil results show, a model is validated for the population, scanners and protocol it was tested on, and nothing more.

**Workflow gains are not automatic.** In the [LungIMPACT randomized trial](https://www.nature.com/articles/s41591-026-04253-5) (Nature Medicine, 2026), prioritizing AI-flagged chest X-rays from UK primary care, across 93,326 X-rays, did not shorten the time to CT or to a lung cancer diagnosis. It is a different test from CT screening, but the lesson carries over: where AI sits in the pathway decides whether it changes anything.

> [!IMPORTANT]
> The clinician who reports the scan remains responsible for the reading, with or without AI. Every FDA label in the table above describes software that supports a clinician's judgement; none lets it report on its own. Using AI as a first reader to skip normal scans moves that line, which is why it is still being tested on trial data first.

## Questions to ask before a screening program adds AI

For hospitals, screening programs and the IT teams that support them (the systems around the tool, from PACS to DICOM routing, are covered in [the role of IT in healthcare](https://computese.com/the-role-of-healthcare-in-it-industry/)), these questions come before any contract:

1. **What exactly is it cleared or CE-marked for?** Screening or incidental nodules, which nodule types and sizes, and which scanners and slice thicknesses the validation covered.
2. **Was it tested on people like yours?** Ask for results by sex, smoking history and ethnicity, and for any evaluation outside the developer's own data.
3. **What happens to your follow-up rate?** Estimate the extra Lung-RADS 3 and 4 findings before go-live; each one is a repeat scan and a worried patient.
4. **How does it appear in the reading workflow?** Concurrent marks in the PACS viewer, a separate report or a triage flag, and who can override it.
5. **Who monitors it after launch?** Track detection and recall rates against your pre-AI baseline, and re-test after software, scanner or protocol changes.
6. **What is the record?** Make the radiologist's report the clinical record, and keep the AI output with it for audit.

Most of that list is ordinary discipline for any AI system, regulated or not: test on your own real cases before go-live, and keep a person approving what matters. For business workflows, that is how our [AI & automation service](https://computese.com/services/ai-automation/) works, with an evaluation set built from your real cases and passed before any model, prompt or source change goes live. For a medical device, the manufacturer's labelled use and your clinical governance set the rules.

## Key terms
- **Low-dose CT (LDCT)**: A chest CT scan taken at a reduced radiation dose (0.65 to 2.36 mSv per scan in the studies the USPSTF reviewed), used to screen for lung cancer.
- **Pulmonary nodule**: A small, roughly round spot of tissue in the lung seen on CT. Most are not cancer; size, density, shape and growth between scans set the risk.
- **Pack-year**: One pack of 20 cigarettes a day for one year. Twenty pack-years can be one pack a day for 20 years or two packs a day for 10.
- **Lung-RADS**: The American College of Radiology's Lung Imaging Reporting and Data System, which sorts screening CT findings into categories that set the next step. Categories 3 and 4 count as a positive screen.
- **CADe (computer-aided detection)**: Software that marks possible findings, such as lung nodules, for a clinician to accept or reject.
- **CADx (computer-aided diagnosis)**: Software that characterizes a finding, for example with a score for how likely a lung nodule is to be malignant.
- **Brock (PanCan) model**: A statistical model from the Pan-Canadian Early Detection of Lung Cancer Study, published in 2013, that estimates a nodule's probability of cancer from patient and nodule features.
- **AUC**: Area under the ROC curve: how well a model ranks cases with the disease above cases without it. 0.5 is chance and 1.0 is perfect. The C-index is the same idea applied over time.
- **External validation**: Testing a model on data from a different hospital, country or population than it was built on. It is the best predictor of how the model behaves in a new program.
- **False positive**: A screen called positive that turns out not to be cancer. In the NLST, 96.4% of positive low-dose CT screens were false positives.

## Common questions

### Can AI detect lung cancer on a CT scan?

AI can find and measure lung nodules on a CT scan and estimate how likely each one is to be cancer, and in some retrospective studies it matched or beat radiologists. It cannot confirm cancer: that takes tissue from a biopsy or surgery. Cleared tools are labelled as aids to the clinician who reads the scan.

### Is AI more accurate than a radiologist at lung cancer screening?

In some retrospective tests, yes: Google's 2019 model beat six radiologists when no prior scan was available and matched them when one was. Evidence from live reading is thinner and mixed; in a 2026 randomized trial, radiologists using AI reported more nodules and recommended more follow-up scans without saving time. Every cleared tool is designed to work with a radiologist, not instead of one.

### What is Sybil, and is it approved?

Sybil is a deep learning model from MIT and Mass General that predicts a person's risk of lung cancer over six years from one low-dose CT scan, published in the Journal of Clinical Oncology in 2023. The model and its annotations are public for research. It was not on the FDA's list of AI-enabled devices as of the list's September 2026 update, and external tests in Asian populations have given mixed results.

### Who should be screened for lung cancer?

In the US, the USPSTF recommends a yearly low-dose CT for adults aged 50 to 80 with a 20 pack-year smoking history who smoke now or quit within the past 15 years. England invites people aged 55 to 74 who have ever smoked to a lung health check, and Canadian provinces combine age, smoking history and a risk model. Whether screening suits you is a conversation to have with your doctor.

### Does AI reduce false positives in lung cancer screening?

It can on retrospective data: Google's model had 11 percentage points fewer false positives than radiologists reading without a prior scan. In live reading it can do the opposite, because software finds more small nodules; in a 2026 randomized trial, follow-up CT recommendations doubled with AI. How a program sets thresholds and reports findings decides which way it goes.

### Which AI lung nodule tools are FDA-cleared?

The FDA's list of AI-enabled devices includes several, among them ClearRead CT, Optellum Virtual Nodule Clinic, RevealAI-Lung, eyonis LCS and Synapse Lung Nodule AI. They are cleared for different jobs: some only detect nodules, and some score only incidentally found nodules and are not indicated for screening. A clearance is not an endorsement; the FDA database entry shows each tool's labelled use.

## Sources
1. [Lung Cancer: Screening (Final Recommendation Statement, March 9, 2021)](https://www.uspreventiveservicestaskforce.org/uspstf/recommendation/lung-cancer-screening), U.S. Preventive Services Task Force
2. [Lung cancer screening](https://www.nhs.uk/tests-and-treatments/lung-cancer-screening/), NHS
3. [New lung cancer screening roll out to detect cancer sooner](https://www.gov.uk/government/news/new-lung-cancer-screening-roll-out-to-detect-cancer-sooner), Department of Health and Social Care (GOV.UK)
4. [Lung Cancer (2016)](https://canadiantaskforce.ca/guidelines/published-guidelines/lung-cancer/), Canadian Task Force on Preventive Health Care
5. [Lung cancer screening in Canada, 2026: Guidelines](https://www.partnershipagainstcancer.ca/topics/lung-cancer-screening-canada/guidelines/), Canadian Partnership Against Cancer
6. [Reduced lung-cancer mortality with low-dose computed tomographic screening](https://pmc.ncbi.nlm.nih.gov/articles/PMC4356534/), New England Journal of Medicine (PMC)
7. [Probability of cancer in pulmonary nodules detected on first screening CT](https://pmc.ncbi.nlm.nih.gov/articles/PMC3951177/), New England Journal of Medicine (PMC)
8. [End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography](https://www.nature.com/articles/s41591-019-0447-x), Nature Medicine
9. [510(k) summary K254075: Synapse Lung Nodule AI](https://www.accessdata.fda.gov/cdrh_docs/pdf25/K254075.pdf), U.S. Food and Drug Administration
10. [External validation of a convolutional neural network artificial intelligence tool to predict malignancy in pulmonary nodules](https://pmc.ncbi.nlm.nih.gov/articles/PMC7231457/), Thorax (PMC)
11. [Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk From a Single Low-Dose Chest Computed Tomography](https://pmc.ncbi.nlm.nih.gov/articles/PMC10419602/), Journal of Clinical Oncology (PMC)
12. [MIT researchers develop an AI model that can detect future lung cancer risk](https://news.mit.edu/2023/ai-model-can-detect-future-lung-cancer-0120), MIT News
13. [Feasibility of AI as first reader in the 4-IN-THE-LUNG-RUN lung cancer screening trial](https://pubmed.ncbi.nlm.nih.gov/39754864/), European Journal of Cancer (PubMed)
14. [Histological proven AI performance in the UKLS CT lung cancer screening study: potential for workload reduction](https://pubmed.ncbi.nlm.nih.gov/40022836/), European Journal of Cancer (PubMed)
15. [External Testing of a Deep Learning Model for Lung Cancer Risk from Low-Dose Chest CT](https://pmc.ncbi.nlm.nih.gov/articles/PMC12405708/), Radiology (PMC)
16. [Artificial Intelligence-Assisted Lung Nodule Evaluation on Low-Dose Chest CT in Asymptomatic Individuals: A Prospective Randomized Controlled Trial](https://pubmed.ncbi.nlm.nih.gov/41848489/), American Journal of Roentgenology (PubMed)
17. [Performance and Utility of the Sybil Deep Learning Model for Lung Cancer Risk Prediction in Asian High- and Low-Risk Populations](https://pubmed.ncbi.nlm.nih.gov/42140481/), Chest (PubMed)
18. [Baseline characteristics of participants in the randomized National Lung Screening Trial](https://pmc.ncbi.nlm.nih.gov/articles/PMC2994863/), Journal of the National Cancer Institute (PMC)
19. [The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans](https://pmc.ncbi.nlm.nih.gov/articles/PMC3041807/), Medical Physics (PMC)
20. [Significance of Image Reconstruction Parameters for Future Lung Cancer Risk Prediction Using Low-Dose Chest CT and the Open-Access Sybil Algorithm](https://pubmed.ncbi.nlm.nih.gov/39437009/), Investigative Radiology (PubMed)
21. [List of Artificial Intelligence-Enabled Medical Devices](https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices), U.S. Food and Drug Administration
22. [510(k) summary K161201: ClearRead CT](https://www.accessdata.fda.gov/cdrh_docs/pdf16/K161201.pdf), U.S. Food and Drug Administration
23. [510(k) summary K202300: Optellum Virtual Nodule Clinic](https://www.accessdata.fda.gov/cdrh_docs/pdf20/K202300.pdf), U.S. Food and Drug Administration
24. [510(k) summary K251769: RevealAI-Lung](https://www.accessdata.fda.gov/cdrh_docs/pdf25/K251769.pdf), U.S. Food and Drug Administration
25. [510(k) summary K251474: eyonis LCS](https://www.accessdata.fda.gov/cdrh_docs/pdf25/K251474.pdf), U.S. Food and Drug Administration
26. [Commercial AI for CT lung cancer screening: product capabilities, coverage of nodule management tasks and supporting evidence](https://pmc.ncbi.nlm.nih.gov/articles/PMC13451236/), European Radiology (PMC)
27. [Software with artificial intelligence-derived algorithms for detecting and analysing lung nodules in CT scans](https://www.journalslibrary.nihr.ac.uk/hta/JYTW8921), NIHR Health Technology Assessment
28. [AI-based chest X-ray prioritization in the lung cancer diagnostic pathway: the LungIMPACT randomized controlled trial](https://www.nature.com/articles/s41591-026-04253-5), Nature Medicine