Screening for Lung Cancer

Lung cancer screening with low-dose computed tomography (CT) reduces mortality but requires integrated, system-based implementation. Review of evolving screening strategies highlighting risk-prediction models, imaging advances, and Lung CT Screening Reporting and Data System (Lung-RADS) as a standardized framework linking findings to management. Current low-dose chest CT requires comparing with earliest CT to detect change; Lung-RADS categories guide management with increasing complexity. AI enhances nodules detection and risk stratification but requires radiologist oversight; computer-aided detection supports consistency and longitudinal tracking. Implementation demands scalable infrastructure, targeted outreach, and continuous quality improvement to maximize survival benefits and equity.

Key points

•
System-based lung cancer screening: integrated, end-to-end process with scalable infrastructure for risk, imaging, interpretation, tracking, and communication.
•
Risk-based eligibility: shift from fixed pack-years to data-driven models to improve equity and personalize screening intervals.
•
Accurate longitudinal assessment: compare current low-dose chest computed tomography (CT) with the earliest CT to detect change and guide management.
•
Lung CT Screening Reporting and Data System (Lung-RADS) and AI with oversight: use Lung-RADS for standardized pathways; AI can aid detection/risk, but radiologists must be trained and maintain oversight with robust follow-up.

Abbreviations

ACR	American College of Radiology
AI	artificial intelligence
CAD	computer-aided detection
CADx	computer-aided diagnosis
CMS	US Centers for Medicare & Medicaid Services
CNNs	convolutional neural networks
FDA	Food and Drug Administration
LCS	lung cancer screening
LDCT	low-dose chest computed tomography
Lung-RADS	Lung CT Screening Reporting and Data System
MIMPS	medical image management processing systems
MIPs	maximum intensity projections
NELSON	Nederlands–Leuvens Longkanker Screenings Onderzoek
NLST	National Lung Screening Trial
ULDCT	ultra-low-dose CT
USPSTF	US Preventive Services Task Force

Introduction

Lung cancer remains the leading cause of cancer-related death worldwide. ^, Survival is strongly stage-dependent, with 5 year survival exceeding 60% for stage I localized disease but plummeting below 40% and 10% for regional and distant disease, respectively.

The introduction of screening with low-dose chest computed tomography (LDCT) marked a turning point for lung cancer care. The National Lung Screening Trial (NLST) first showed that LDCT screening in high-risk populations reduced lung cancer mortality by 20% compared to chest radiography. ^, The Nederlands–Leuvens Longkanker Screenings Onderzoek (NELSON) study subsequently validated this finding in Europe, demonstrating mortality reduction up to 33% in women, and a prospective cohort study demonstrated a 31% decrease in lung cancer mortality in China for both men and women. ^, Yet, translating clinical trial findings into positive outcomes in the real world remains challenging. Since the introduction of lung screening with LDCT in 2015, programs in the United States have been plagued by low enrollment and notable disparities across racial, geographic, and socioeconomic groups. These difficulties have highlighted a key insight: screening for lung cancer is a systematic, system-based intervention rather than a standalone test. The coordination of several interlocking domains, such as risk prediction, image acquisition, interpretation, longitudinal tracking, and patient-centered communication, all integrated within a scalable and accountable clinical infrastructure, is essential for successful lung screening programs. ^,

Several transformative developments are reshaping lung cancer screening. While pack-year–based thresholds have come under increasing scrutiny for their narrow scope and systemic bias, risk prediction models based on imaging have emerged as potential tools for identifying high-risk individuals. ^,^,^, These developments signal a shift toward dynamic data-driven eligibility frameworks and personalized screening intervals. At the same time, imaging technology has evolved, enabling image acquisition with reduced radiation exposure while enhancing image quality through iterative and deep-learning–based reconstruction techniques. ^, Lastly, artificial intelligence (AI)-driven tools aid in nodule detection, risk stratification, and even detection of comorbidities. This review synthesizes how these developments have shaped the state of lung cancer screening in 2025 and illustrates the shift from rigid eligibility thresholds and at least annual scans to a more personalized and equitable approach.

Who Should Be Screened? Rethinking Risk Beyond Pack-Years

Lung cancer screening eligibility in the United States is determined by the US Preventive Services Task Force (USPSTF) and derived from the eligibility criteria of the NLST and NELSON trials. Since 2013, USPSTF recommends annual screening with low-dose chest computed tomography for those eligible based on age, pack-years smoked, and the time since smoking cessation. Importantly, USPSTF also recommends that screening should be discontinued once a person has not smoked for 15 years or develops a health problem that substantially limits life expectancy or the ability or willingness to have curative lung surgery. In 2021, the USPSTF lowered the pack-year threshold from 30 to 20, and the minimum age from 55 to 50 years ( Fig. 1 ). The current iteration of the national coverage decision from the US Centers for Medicare & Medicaid Services (CMS) limits eligibility to adults aged 50 to 77 years, excluding the most senior population ( Fig. 2 ).

A clinical infographic depicts the 2021 U S P S T F lung cancer screening criteria that emphasize the age range of 50 to 80 years and the 20-pack-year smoking history for annual L D C T scans. — Fig. 1

A clinical infographic details the C M S eligibility for lung cancer screening, specifying the age range of 50 to 77, a 20-pack-year history, and the mandatory shared decision-making visit. — Fig. 2

The pack-year model treats smoking 1 pack of cigarettes per day for 30 years as equivalent to smoking 2 packs of cigarettes per day for 15 years, implicitly assuming that smoking intensity and duration contribute equally to lung cancer risk. Emerging evidence indicates that smoking duration is a significantly stronger predictor of lung cancer risk than intensity, challenging the validity of this assumption. ^, Time from quitting has been questioned as a criterion since increased risk for lung cancer persists for decades after smoking cessation. ^, Studies have demonstrated that many lung cancer cases arise in individuals who would not have qualified for screening under current CMS or USPSTF guidelines. This reflects in part racial and gender disparities embedded within the pack-year framework: Black Americans and women, who often have higher lung cancer risk at lower smoking intensities, are disproportionately excluded. ^, In addition, lung cancer screening (LCS) remains focused on cigarette smokers and does not consider other types of smoking such as cigars, electronic cigarettes, or marijuana. Lastly, those who never smoked remain completely excluded from LCS. Lung cancer in never-smokers accounts for 15% to 20% of all patients with lung cancer and is more common in women, Asians, and younger patients.

Recent evidence supports alternative approaches that may both improve risk prediction and reduce disparities. One such approach involves relying on smoking duration alone—regardless of the number of cigarettes smoked per day—as an eligibility criterion. Potter and colleagues found that applying a minimum of 20 years of smoking history, rather than 20 pack-years, significantly increased screening eligibility among Black adults while maintaining similar predictive performance. Additionally, validated risk prediction models offer more individualized and data-driven alternatives to binary thresholds. The PLCOm2012 risk predictor, derived from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial, incorporates age, race or ethnicity, smoking history, personal history of cancer, family history of lung cancer, body mass index, education level, and the presence of chronic obstructive pulmonary disease. ^, Other models, including the Liverpool Lung Project risk model version 2, Liverpool Lung Project risk model version 3, the Lung Cancer Risk Assessment Tool, Lung Cancer Death Risk Assessment Tool, and the Life Years Gained From Screening-CT, vary in complexity and features but stratify risk more accurately and comprehensively than USPSTF criteria. ^,^,

AI has also started to change risk prediction by allowing models to process images directly. A notable example is Sybil, an open-access deep learning model trained to predict lung cancer risk from a single baseline LDCT scan, without the need for human input such as areas of interest, smoking history, demographics, or laboratory data. Sybil demonstrated strong predictive performance during internal and external validation up to 6 years out, even in never-smokers, suggesting that intrinsic imaging features may encode more information about cancer risk than was previously thought. Sybil may serve as an effective tool for stratifying individuals into high-risk and low-risk categories for developing lung cancer, potentially supporting the implementation of personalized screening intervals based on individual risk. ^, However, like many other AI algorithms, Sybil—despite undergoing external validation —still requires large-scale, multi-institutional studies across diverse target populations before translation into clinical practice.

Emerging hybrid approaches integrate data from electronic health records and imaging to maximize individualized cancer risk assessment. ^, These innovations signal a transition to a dynamic, continuous risk assessment framework, where eligibility and screening intervals could be updated as new data become available and are integrated into a personalized profile. Such systems would allow screening programs to account not just for an individual’s smoking history, but for who they are now and what their current imaging reveals.

Low-Dose Chest Computed Tomography and Technical Advances

The success of LCS programs relies not only on selecting the right participants but also on imaging them with protocols that maximize diagnostic accuracy while minimizing radiation exposure.

The American College of Radiology (ACR) has outlined minimum technical requirements for LDCT screening, including the use of less than 1.5 mSv effective radiation dose, thin-slice (≤1.25 mm) reconstruction, and the omission of intravenous contrast. ^, Iterative reconstruction techniques have steadily replaced traditional filtered back projection, reducing radiation dose while maintaining or even improving image quality. Maximum intensity projections (MIPs) are now routinely applied for pulmonary nodule detection, and their ability to highlight small nodules has made MIPs indispensable in high-volume screening settings. ^,

Computer-aided detection (CAD) tools assist in nodule tracking, volumetric growth analysis, and classification. ^, As “second readers,” they can highlight overlooked nodules and prompt more consistent assessment across different radiologists, thereby decreasing interreader variability in nodule detection and classification. ^, Software design to measure lung nodule volume and growth rates continues to struggle with nonsolid nodules since only slight variation in attenuation exist between nonsolid nodules and the surrounding normal lung parenchyma. Furthermore, the segmentation of solid nodules next to the pleura or pulmonary vessels tends to include structures beyond the nodule margin, resulting in overestimation of nodule size.

Another important advancement in LCS practice has been the move toward ultra-low-dose imaging. Given that screening involves repeated irradiation of asymptomatic individuals, minimizing radiation exposure remains a critical concern, driving efforts to optimize dose levels in accordance with the ALARA (As Low As Reasonably Achievable) principle. To this end, ultra-low-dose CT (ULDCT) techniques were developed, lowering the effective dose to approximately 0.15 mSv, equivalent to 1.5 chest radiograph effective radiation dose. However, the inherent increase in image noise with ULDCT compromises diagnostic accuracy, particularly in detecting and measuring pulmonary nodules, limiting widespread clinical adoption to date.

To address this limitation, iterative reconstruction techniques were introduced, applying edge-preserving maximum likelihood methods and successive forward-backward steps reconstruction to iteratively reduce uncorrelated noise and enhance image quality. ^, Building on this foundation, deep learning–based image reconstruction—especially those using convolutional neural networks (CNNs)—has emerged as a powerful tool for CT denoising. These networks are trained to map noisy input images to denoised outputs using gradient back-propagation algorithm. ^, This method can be applied directly to processed images without the need for scanner raw data, making it vendor-independent, easy to implement, and capable of rapid processing. ^,

Deep learning image reconstruction applied to ULDCT has demonstrated significantly improved nodule detection rates and measurement accuracy compared to adaptive statistical iterative reconstruction, with studies also showing that the combination of both techniques enhances radiologist-rated nodule edge delineation. Furthermore, denoised ULDCT has proven particularly valuable in detecting subsolid nodules. ^,

Lung Computed Tomography Screening Reporting and Data System 2022

LCS remains focused on lung nodule detection and characterization. Radiologists need to compare the current scan with the oldest available chest CT to accurately assess for new lung nodules, and whether pre-existing nodules have changed in size or attenuation over time. Adaptable standardized reporting systems are required to match the increasing complexity of imaging findings and clinical decision-making. The Lung CT Screening Reporting and Data System (Lung-RADS) serves this purpose by assigning findings on LDCT to categories ranging from 0 to 4 and linking each category to standardized management recommendations ( Figs. 3–6 ). The ACR continues to develop Lung-RADS with updates released in 2019 (version 1.1) and November 2022.

A lung-window chest C T shows small new solid nodules in the left lower lobe with surrounding ground-glass opacities, consistent with Lung-R A D S category 0 inflammatory findings. — Fig. 3

Three C T scans show Lung-R A D S category 1, which highlights a benign calcified granuloma and a pulmonary hamartoma with internal fat and calcium. — Fig. 4

A lung-R A D S category 2 axial unenhanced chest CT scan in lung window shows a small, benign 3 millimeter solid perifissural triangular nodule in the right lower lobe, indicated by an arrow. — Fig. 5

Four axial unenhanced chest C T scans show Lung-R A D S category 4 nodules, including a 9 millimeter solid nodule, a 12 millimeter part-solid nodule, and a nodule increasing from 6 to 9 millimeters. — Fig. 6

Notable changes in the 2022 revision include the explicit recognition of atypical pulmonary cysts as potential early indicators of malignancy. Cystic components have been reported in up to 9% of lung cancers and are more likely to be misinterpreted as benign than lung cancers presenting as solid nodules. Precursor lesions of cystic lung cancers can be unilocular thick-walled cysts, cysts with mural nodularity, or multilocular cysts, which may evolve over time with nodule growth, increased asymmetric wall thickening, cyst enlargement, or transformation into a solid nodule on imaging ^, ( Fig. 7 ). An atypical pulmonary thick-walled cyst with an enlarging cystic component is classified as Lung-RADS 3 and should be followed up with a 6 month low-dose CT. Thick-walled cysts with wall thickness of 2 mm or greater, focal nodularity, or multilocular appearance are categorized as Lung-RADS 4A, with recommended follow-up including a 3 month low-dose CT or PET/CT if a solid component of at least 8 mm is present. Progressive wall thickening or nodularity in a thick-walled cyst, increased loculation in a multilocular cyst, or new or enlarging opacities within or adjacent to a multilocular cyst warrants a Lung-RADS 4B classification, prompting tissue sampling. Distinguishing a cavitary nodule from an atypical lung cyst relies on identifying the dominant feature: if the solid component predominates, it is managed as a solid nodule; if the cystic component predominates, it is managed as an atypical lung cyst.

Four chest C T scans show Lung-R A D S category 4 atypical pulmonary cysts, including a multiloculated cyst, a cystic nodule with 3 millimeter wall thickening, and a cyst growing over one year. — Fig. 7

Another major change relates to endobronchial nodules: Lung-RADS 2022 rebrands these as “airway nodules” and refines their risk categorization. Although airway malignancies are rare in LCS and most airway findings are benign, 22% of missed cancers at initial screening were later identified as central endobronchial lesions, highlighting the clinical importance of careful airway evaluation. Nodules in the large airways (segmental or more proximal) are classified as Lung-RADS 4A with a recommended 3 month LDCT follow-up ( Fig. 8 ). Nodules in subsegmental airways often reflect mucus plugging or inflammatory conditions and represent Lung-RADS 2 findings. Airway nodules with features suggestive of secretions (eg, tubular shape, air content, absence of soft tissue, or Hounsfield units <20) support a Lung-RADS 2 classification. Persistent proximal airway nodules at 3 month follow-up CT raise concern for malignancy and are upgraded to Lung-RADS 4B, requiring diagnostic FDG-PET/CT if solid component 8 mm or greater or referral for bronchoscopy.

A lung-R A D S category 4A axial unenhanced chest C T scan in lung window shows a 5 millimeter polypoid nodule in the left mainstem bronchus, indicated by an arrow. — Fig. 8

Artificial Intelligence in Lung Cancer Screening

AI is increasingly integrated into the analysis of screening LDCT scans to enhance diagnostic consistency, operational efficiency, and clinical decision-making. A leading use case of AI in LCS is the detection and classification of pulmonary nodules. AI algorithms have shown high sensitivity and accuracy in detecting lung nodules, with deep learning models surpassing radiologists in identifying small nodules. Advanced techniques using CNNs have significantly lowered false-positive rates to approximately one per scan. Several Food and Drug Administration (FDA)-cleared deep learning algorithms have demonstrated high sensitivity in identifying nodules that are small or subsolid ^, —features that can challenge even experienced radiologists. While most AI software requires radiologist oversight to confirm the presence of true nodules, these tools can streamline workflow, potentially reducing interpretation time—as demonstrated in one study reporting a 26% decrease using vessel-suppressed imaging combined with CAD. Additionally, many commercial platforms enable direct comparison of nodules across serial scans, further enhancing diagnostic efficiency. Table 1 summarizes FDA-cleared tools for nodule detection and classification based on the ACR Data Science Institute AI Central database.

Table 1

Commercial artificial intelligence models for lung nodule detection and classification on chest computed tomographic scans, extracted from the American College of Radiology Data Science Institute AI Central database, chest section, on August 11th, 2025, in alphabetical order

From the American College of Radiology Data Science Institute AI Central database– Chest Section ( https://aicentral.acrdsi.org/ ).

Product	Manufacturer	FDA Submission No.	Date Cleared	Category
AI-Rad Companion Chest CT	Siemens Healthineers	K222360	04/06/2023	CADe
AVIEW	Coreline Soft Co., Ltd	K243689	03/19/2025	MIMPS
AVIEW Lung Nodule CAD	Coreline Soft Co., Ltd	K221592	02/24/2023	CADe
AVIEW LCS	Coreline Soft Co., Ltd	K201710	10/16/2020	MIMPS
ClearRead CT	Riverain Technologies	K221612	12/05/2022	CADe
InferRead Lung CT.AI	Beijing Infervision Technology Co., Ltd.	K240554	05/16/2025	MIMPS
Lung Nodule Assessment and Comparison Option (LNA)	Philips Medical Systems	K162484	02/23/2017	MIMPS
NinesMeasure	Nines, Inc.	K202990	02/25/2021	MIMPS
Optellum Virtual Nodule Clinic, Optellum Software, Optellum Platform	Optellum Ltd.	K202300	03/05/2021	CADx
qCT LN Quant	Qure.ai Technologies	K240740	08/16/2024	MIMPS
V5med Lung AI	V5med Inc.	K242919	03/27/2025	CADe
Veolity	MeVis Medical Solutions AG	K201501	02/23/2021	CADe