What can AI do in lung cancer screening with low-dose CT?

AI lung cancer screening uses deep learning to read low-dose CT scans, flag suspicious nodules and estimate a person’s risk of developing lung cancer within six years. Models such as Sybil perform very well for cancers already visible on the scan and less well at predicting future cancers, so they support radiologists rather than replace them.

Low-dose CT (LDCT) screening reduces deaths in high-risk smokers: the National Lung Screening Trial reported a 20% reduction in lung cancer mortality, and the NELSON trial reported a reduction of 24% to 33%. The cost is a large number of indeterminate nodules, many of which are not cancer, leading to repeat scans, anxiety and sometimes unnecessary invasive procedures.

This page explains how AI models are used on LDCT, what the Sybil studies showed in the United States, Taiwan and an Asian health-checkup cohort, where performance drops, and what patients considering screening or treatment in India, China or elsewhere should ask.

Key facts about the AI lung cancer screening evidence

The main published and registered evidence, in brief:

  • Sybil (Journal of Clinical Oncology, 2023): trained on National Lung Screening Trial scans; needs one LDCT and no clinical data or radiologist annotations; predicts risk up to six years ahead
  • Sybil one-year AUC: 0.92 on NLST, 0.86 at Massachusetts General Hospital and 0.94 in a Taiwan validation set
  • External test (Radiology, 2025): 18,057 people in an Asian health-checkup cohort; one-year AUC 0.91 and six-year AUC 0.74
  • Heavy smokers (2,848 people): AUC 0.94 for cancers visible on the baseline scan and 0.70 for future cancers
  • Never or light smokers (9,943 people): AUC 0.89 for visible cancers and 0.56 for future cancers
  • Planned head-to-head study (ClinicalTrials.gov NCT07727122, registered July 2026): three imaging-only AI models to be compared prospectively for nodule malignancy risk
  • Screening context: NLST reported 20% lower lung cancer mortality and NELSON 24% to 33%

Key takeaways: AI lung cancer screening and Sybil

  • AI can read LDCT quickly and score nodule or lung cancer risk from the image alone.
  • Sybil’s accuracy for cancers already visible on the scan was high (AUC 0.94 in heavy smokers), but predicting future cancers was weaker (AUC 0.70, and 0.56 in never or light smokers).
  • LDCT screening itself reduces lung cancer deaths; the open question is whether adding AI improves outcomes, cuts false positives or finds more early cancers.
  • Head-to-head prospective comparisons of AI models are only now being designed.
  • Nodule findings need radiologist review, follow-up imaging and often a multidisciplinary discussion; an AI score is one input.

How does AI read a low-dose CT scan for lung cancer?

LDCT produces a three-dimensional stack of chest images. AI systems for lung screening generally do one of two jobs. Detection software finds and measures nodules and marks them for the radiologist. Risk models such as Sybil look at the whole scan, highlight regions of concern and output a probability of developing lung cancer over the following one to six years. Sybil’s attention maps show which part of the lung drove the score, which helps a radiologist judge whether it makes sense.

Sybil was developed on National Lung Screening Trial scans and does not need smoking history or radiologist annotations. It is designed to run in the background on a radiology reading station. Other tools predict whether a specific nodule is malignant, aiming to cut follow-up scans for benign nodules and to speed biopsy for suspicious ones.

What did the Sybil studies show in the United States and Taiwan?

The first Sybil paper, in the Journal of Clinical Oncology in 2023, reported one-year area-under-the-curve (AUC) values of 0.92 on additional National Lung Screening Trial participants, 0.86 on scans from Massachusetts General Hospital and 0.94 on scans from Chang Gung Memorial Hospital in Taiwan, which included people with a range of smoking histories, including never-smokers. AUC measures how well a model separates people who develop cancer from those who do not, with 1.0 perfect and 0.5 no better than chance.

The authors concluded that Sybil can accurately predict an individual’s future lung cancer risk from a single LDCT and that further study is needed to understand its clinical applications. In other words, accuracy on retrospective datasets was strong, but the paper did not show that using Sybil changes patient outcomes.

Does Sybil work in Asian populations and in never-smokers?

A 2025 external test in Radiology evaluated Sybil on 18,057 people in an Asian health-checkup cohort, including 2,848 heavy smokers and 9,943 never or light smokers. Across the full sample, Sybil achieved an AUC of 0.91 for one-year risk and 0.74 for six-year risk. Among heavy smokers, the one-year AUC was 0.94 for cancers visible on the baseline scan and the six-year AUC was 0.70 for future cancers. Among never or light smokers, the AUC was 0.89 for visible cancers but only 0.56 for future cancers.

The pattern is important. Sybil is very good at recognising cancers already present on the scan, which overlaps with what radiologists do. Its power to foresee cancers that appear years later is moderate in heavy smokers and weak in never or light smokers, a group that matters because standard screening criteria do not cover them. The subgroups also contained few cancers (24 and 41), so the estimates are uncertain, and results from a health-checkup cohort should not be assumed to apply to screening programmes elsewhere.

Key Sybil performance numbers at a glance

Reported results from the two main Sybil publications:

  • Journal of Clinical Oncology 2023, one-year AUC: 0.92 (NLST), 0.86 (Massachusetts General Hospital), 0.94 (Taiwan)
  • Asian health-checkup cohort (18,057 people): one-year AUC 0.91; six-year AUC 0.74
  • Heavy smokers: 0.94 for visible cancers and 0.70 for future cancers (24 cancers, 0.8%)
  • Never or light smokers: 0.89 for visible cancers and 0.56 for future cancers (41 cancers, 0.4%)
  • Interpretation: strong for visible cancers, moderate to weak for future cancers

AI lung nodule detection and malignancy prediction: where does the evidence stand?

Beyond Sybil, many AI products detect nodules and predict their malignancy on chest CT, but head-to-head evidence is thin. A prospective multicentre study registered in July 2026 (NCT07727122) states that comparisons of AI models in real-world populations with a pathological gold standard are lacking, and it will compare three imaging-only models, including products from Shukun Technology and United Imaging, on AUC for malignancy prediction.

Registered research also explores AI for new nodules after a treated cancer. The NIMBLE study uses AI and machine-learning radiomics, alongside blood biomarkers, to distinguish benign disease from malignancy and, where malignant, metastatic recurrence from a new primary lung cancer, with the aim of tailoring surveillance intensity. Such uses are still research.

False positives, nodules and overdiagnosis in AI-assisted lung screening

LDCT screening finds many small nodules, and most are not cancer. Registered study protocols describe a high prevalence of indeterminate nodules and substantial false positives that lead to unnecessary invasive procedures and anxiety, while some early cancers are still missed. AI could help by sorting nodules into those needing urgent work-up and those suited to routine follow-up, but it can also add false alarms if it is too sensitive.

In a prospective study of 28,899 high-risk participants in Hebei Province, China, between 2013 and 2019, 3,697 people had a lung nodule (12.7%) and 421 (1.5%) had suspected lung cancer. Those numbers illustrate the scale of nodule follow-up that screening programmes must manage, and why accuracy in sorting nodules matters.

Limits of the AI lung screening evidence

AI lung screening is promising, but it is not yet proven to improve outcomes.

  • Most accuracy data are retrospective; prospective evidence on outcomes is limited
  • Predicting future cancer is much weaker than recognising visible cancer, especially in never or light smokers
  • Cancer counts in some validation subgroups were small (24 and 41 cancers), so estimates are uncertain
  • Different AI products have not yet been compared head to head in completed prospective multicentre studies

Who could ask about AI-assisted lung cancer screening?

Screening eligibility is set by national guidelines. These are questions to raise with a doctor, not recommendations:

  • Are you a current or former heavy smoker in the age range your country’s guidelines cover for low-dose CT?
  • Do you have a persistent cough, coughing up blood, unexplained weight loss or chest pain? Symptoms need diagnostic assessment, not just screening.
  • Were you told that a nodule was found on a scan done for another reason?
  • Do you have a family history of lung cancer or exposure to asbestos, radon or heavy air pollution?
  • Are you a never-smoker with risk factors who has been told you do not qualify for screening?
  • Do you want a second opinion on a CT report or on whether a nodule needs biopsy?

How does AI lung screening compare with radiologist reading and Lung-RADS?

Radiologists grade screening results with a structured system called Lung-RADS, which uses nodule size, density and growth to decide the next step. The 2025 Radiology study also compared Sybil’s sensitivity and specificity with Lung-RADS, although we have not reproduced those figures here. Because Sybil can run silently in the background of a reading workstation, it can sit alongside this process rather than replace it.

In practice, AI is best viewed as a second reader and a prioritiser. It can draw attention to a subtle nodule, estimate how worrying a nodule looks and rank scans for review. Final decisions on follow-up CT, PET-CT, biopsy or surgery rest with radiologists, pulmonologists, thoracic surgeons and the tumour board.

AI lung screening evidence and access by region

What has been published or registered, with gaps stated:

  • United States and Taiwan: Sybil validated retrospectively on NLST, Massachusetts General Hospital and Chang Gung Memorial Hospital scans
  • Asian health-checkup cohort: external test of Sybil published in Radiology in 2025 (18,057 people)
  • Registered prospective comparison (NCT07727122, July 2026): three imaging-only models, including LungDoc from Shukun Technology and a United Imaging model
  • India and China routine screening: we have not verified which hospitals use AI for lung screening or which products hold local approval

Is AI lung cancer screening available in India or China?

AI software for nodule detection is used in hospitals in many countries, but we have not verified which hospitals or screening programmes in India or China use AI for routine lung screening, which products they use or whether those products hold local regulatory approval. We also have not verified whether Sybil is deployed in either country.

For patients treated in India or China, the practical steps are the same: ask whether your CT was read by a thoracic radiologist, whether nodule size and growth were measured by volume, and whether a pulmonologist or tumour board reviewed a suspicious nodule.

What happens after a suspicious lung nodule is found?

A suspicious nodule starts a work-up. Depending on size, appearance and growth, the next step may be a repeat CT after a few months, a PET-CT scan, a needle biopsy or bronchoscopy, or surgery for a nodule with a high chance of being cancer. If lung cancer is confirmed, pathology and biomarker testing, such as EGFR, ALK and PD-L1, guide treatment, which may include surgery, radiotherapy, targeted therapy, immunotherapy, chemotherapy or participation in clinical trials.

AI findings do not change that pathway, but they may change how quickly a nodule is flagged and how confidently a radiologist can categorise it. A second opinion on the scan and the pathology is reasonable before major treatment, particularly if findings are borderline.

What happens next for AI lung cancer screening research?

Evidence is moving from retrospective accuracy to prospective, head-to-head and outcome studies. A multicentre comparison registered in 2026 will test three imaging-only models against a pathological gold standard. Other work aims to personalise screening intervals, extend risk prediction to never-smokers and combine imaging with blood biomarkers.

The most useful future results will show whether AI-assisted screening finds more early cancers, reduces unnecessary procedures and lowers deaths compared with LDCT read by radiologists alone. Until then, AI is a promising assistant. Tools such as AI clinical trial matching may also help patients find studies of new screening and nodule-management approaches.

Questions to ask your doctor about AI and lung cancer screening

  • Does my screening programme or imaging centre use AI software, and does a radiologist review every scan?
  • Do I qualify for low-dose CT screening under current guidelines, and how often should I be scanned?
  • If a nodule was found, how big is it, how is growth measured and when is the next review?
  • Would a PET-CT, biopsy or specialist review change the plan?
  • If cancer is confirmed, which biomarker tests, such as EGFR, ALK and PD-L1, should be done?
  • Can my CT scans be reviewed by a thoracic radiologist for a second opinion?

Sources

Where the figures in this guide come from:

Journal of Clinical Oncology: Mikhael et al., Sybil, a validated deep learning model to predict future lung cancer risk from a single low-dose chest CT (2023)

Radiology: External Testing of a Deep Learning Model for Lung Cancer Risk from Low-Dose Chest CT (2025)

EurekAlert: Artificial intelligence tool developed to predict risk of lung cancer (Sybil news release)

MIT: AI model that can detect future lung cancer risk (Sybil)

ClinicalTrials.gov: Multicenter Prospective Validation of AI Models for Malignancy Risk Prediction in Pulmonary Nodules (NCT07727122)

ClinicalTrials.gov: Nodule IMmunophenotyping Biomarker for Lung Cancer Early Diagnosis, NIMBLE (NCT05432739)

PMC: Epidemiological characteristics and risk factors of lung cancer in a population with incidental pulmonary nodules in China (10,560 cases)

PubMed: Sybil, a validated deep learning model to predict future lung cancer risk (record 36634294)

Frequently asked questions

Can AI detect lung cancer on a low-dose CT scan?

Yes, AI can flag suspicious nodules and score risk on low-dose CT. Sybil achieved a one-year AUC of 0.92 on NLST scans and 0.94 in a Taiwan validation set. Accuracy is much lower for predicting cancers that appear years later, and results come from retrospective data. A radiologist still reviews the scan and decides on follow-up.

What is Sybil in lung cancer screening?

Sybil is an open-source deep learning model, trained on National Lung Screening Trial scans, that predicts the risk of developing lung cancer within one to six years from a single low-dose CT. It needs no clinical data or radiologist annotations and can run in the background on a reading station. Its clinical impact on outcomes has not yet been shown.

Does AI lung cancer screening work for never-smokers?

Only partly. In an Asian health-checkup cohort, Sybil scored an AUC of 0.89 for cancers already visible on the scan in never or light smokers, but only 0.56 for cancers that appeared later, little better than chance. Only 41 cancers were in that subgroup, so estimates are uncertain. Never-smokers with symptoms need diagnostic assessment.

Does AI reduce false positives in lung cancer screening?

It may, but this is not yet proven. LDCT produces many indeterminate nodules and false positives, and AI aims to sort them better. A prospective multicentre study registered in July 2026 will compare three AI models for nodule malignancy risk, because head-to-head real-world comparisons with a pathological gold standard have been lacking.

Is AI lung cancer screening available in India?

We have not verified which hospitals or screening programmes in India use AI for lung screening, or whether specific products hold local approval. If you are in India, ask your imaging centre whether AI is used and whether a thoracic radiologist reviews every scan.

What should I do if my CT scan shows a lung nodule?

Follow your doctor’s plan, which may include a repeat CT, PET-CT, biopsy or specialist review depending on the nodule’s size, appearance and growth. Most small nodules are not cancer, but they need proper assessment. If cancer is confirmed, ask for biomarker testing and consider a second opinion before treatment.

Does low-dose CT screening reduce lung cancer deaths?

In high-risk smokers, yes. The National Lung Screening Trial reported a 20% reduction in lung cancer deaths with low-dose CT, and the NELSON trial reported a reduction of 24% to 33%.

The cost is a large number of indeterminate nodules, many of which are not cancer, leading to repeat scans, anxiety and sometimes unnecessary invasive procedures. Whether adding AI improves outcomes, cuts false positives or finds more early cancers is still an open question. Screening eligibility is set by national guidelines, so ask your doctor whether you qualify.

How does AI fit with radiologists and Lung-RADS?

Radiologists grade screening results with Lung-RADS, a structured system that uses nodule size, density and growth to decide the next step. Because Sybil can run silently in the background of a reading workstation, it can sit alongside this process rather than replace it.

AI is best seen as a second reader and a prioritiser. It can draw attention to a subtle nodule and rank scans for review. Final decisions on follow-up CT, PET-CT, biopsy or surgery rest with radiologists, pulmonologists, thoracic surgeons and the tumour board.

This page is general information, not medical advice. OncoVOIS explains and helps you prepare; your oncologist decides about your care.