Detecting lung nodules in LDCT images is central to LC screening workflows, as it guides participant management. However, the repetitive nature of this task and the overwhelming volume of images contribute to high intra- and inter-observer variability and a high false positive rate.51,52 Computer-aided diagnosis (CAD) systems assist radiologists by automatically identifying subtle findings, thereby mitigating human limitations like memory, distraction, and fatigue and offering objective data interpretation.53 Computer-aided detection (CADe) systems are used for detection, while computer-aided diagnosis (CADx) systems are used for diagnosis.51,53 CADe systems have been shown to reduce the rate of false-negative baseline LDCT examinations.3,54–61 Additionally, they can help detect infra-threshold nodules that do not qualify as positive according to Lung-RADS,62 but need to be monitored in subsequent LDCT examinations.29 However, only a small fraction (below 1%) of micronodules (<4mm) evolves into LC,63 indicating that the specificity of a micronodule at baseline LDCT is extremely low.

Using CADe for the computation of lung nodule volume rather than diameters has improved classification of nodules and decreased the number of indeterminate or false positive LDCT examinations.64 However, the clinical integration of CADe remains limited due to persistent concerns over high false positive rates.65,66 Researchers are addressing this issue through several strategies. CADe tools may be used as pre-screening instruments to rule out negative LDCT examinations, allowing radiologists to concentrate on more challenging and suspicious cases.67,68 Another strategy involves integrating more data into the models. For example, a ‘collaborative CAD’ system incorporating radiologists’ gaze patterns into a 3D multi-task convolutional neural network (CNN), a particular DL architecture,69 achieved a 97% classification accuracy in identifying nodules.70

In LC screening, it is crucial to distinguish between benign lung nodules, which constitute the vast majority observed in low-dose CT scans of screened subjects according to Lung-RADS v2022,62 and malignant lung nodules. This differentiation often leads to additional examinations, such as follow-up LDCTs at intervals of 1–3–6 months, FDG-PET,71 and invasive procedures, significantly increasing both the costs and potential harms associated with screening.7 Notably, malignant nodules demonstrate an increase in size, density, or both over subsequent 3 or 6-month follow-up LDCT scans, as outlined in Lung-RADS v2022.62 The calculation of volume doubling time (VDT) serves as a practical and effective method to assess nodule growth characteristics and malignancy risk.64 The Lung-RADS guidelines recommend specific management strategies for baseline LDCT-detected nodules, particularly solid non-calcified nodules ≥6mm in diameter or ≥113mm3 in volume, which helps streamline further investigations aimed at confirming malignancy and minimizing unnecessary procedures.72 Furthermore, this differentiation can be enhanced by integrating LDCT features such as nodule size and density, the number of nodules, and presence of emphysema, with pertinent subject history, as incorporated in the PanCan/Brock model,73,74 or with biomarker results such as plasma DNA methylation75 or plasma total cfDNA.13 However, the PanCan/Brock model has been developed, tested, and calibrated specifically for prevalent solid nodules ≥6mm in diameter.73,74,76 It may not be well-suited for newly appearing nodules detected at next LDCT screening rounds.76 Additionally, this model may not effectively identify malignant micronodules, potentially leading to the delayed (“pseudo-incidental”) LC diagnosis. Therefore, DL algorithms predicting LC based on baseline LDCT and radiomics77–80 may improve the characterization of these small nodules. For example, an ML approach combining epidemiological, clinical and radiomic features, extracted from the nodules present at baseline LDCT, was able to predict the nodule's malignancy risk score with an area under receiving operator curve (AUROC) of 0.93, outperforming the PanCan/Brock model and with optimal performance for both solid and sub-solid nodules.81 Still, a generative approach to enhance the characterization of indeterminate nodules from the baseline LDCT scan80 exploited a growth model based on the Wasserstein generative adversarial network framework (GP-WGAN) to predict the nodule growth patterns in the 1-year follow-up LDCT scans. By leveraging the ability of GANs to generate data similar to the original, they can simulate follow-up LDCT examinations requiring only the baseline LDCT as input. The results demonstrated that the generated follow-up nodule images, when used as input to a model for LC malignancy prediction, achieved performance comparable to using real follow-up nodule images (AUROC of 0.82±0.02 for generated nodules, compared to 0.86±0.02 for real nodules).80

For LC screening to be effective and minimize related harms, it is crucial to carefully select the at-risk population.49 Once selected, LDCT examination information allows for valuable risk stratification, enabling a tailored screening schedule.82 Several models for estimating LC risk have incorporated baseline LDCT findings.83–85 Their implementation is hindered by limited external validation and the need for manual input of LDCT findings into the model to calculate the score. Unlike traditional models, AI-based algorithms autonomously analyze the entire LDCT volume, identify lung nodules and incidental findings, and combine this information with demographic data to generate a comprehensive, automated risk score.

The Google DL model evaluates LDCT examinations to predict LC incidence. It extracts local and global features from the current, and optionally prior, LDCT examinations and estimates the likelihood of a LC diagnosis within a year.86 Despite achieving a high AUROC of 0.959 on single LDCT examination and outperforming radiologists, the model was criticized for its ‘black-box’ nature, lack of source code availability, and small validation set.87

DeepScreener is a DL algorithm designed to predict a patient's cancer status from CT scans through three tasks: nodule segmentation, nodule-level classification and patient-level classification.88,89 For each nodule, the nodule-level classifier extracts morphological, textural and other features and combines them with the nodule location to calculate a risk score. Subsequently, the patient-level classifier aggregates the risk scores of all detected nodules to generate an overall risk score for the patient and determine the label (“cancer” or “no cancer”). The model achieved an AUROC of 0.89 and a sensitivity of only 42.4%, indicating that further refinement and validation are needed.89

Sybil, a DL model designed to predict using a 0–1 score the LC risk from a single LDCT examination up to the next six years, without the need for radiologist annotations or additional data, represents a recent advancement.90 It achieved AUROC for LC prediction at one year of 0.86–0.94 in three different datasets. Interestingly, when Sybil predicts high LC risk, the used signal localizes to specific at-risk regions rather than being equally spread over the entire thorax.90

DL algorithms such as Sybil could be used to stratify the risk of LC after a baseline LDCT and could be particularly valuable in providing the LC risk in a screened subject showing infra-threshold nodules, that correspond to benign or pseudo-incident LC, or, after a negative baseline LDCT, anticipating interval or incident LC. Examples of application of the Sybil algorithm are shown in Fig. 4.

Fig. 4.

(A–E) Assessment of risk of LC in the next 1–6 years based on the analysis of baseline LDCT with the Sybil deep learning algorithm.90 (A) Prediction of a very low probability of LC after 1 year (risk score=0.0109) and 6 years (risk score=0.0831) since baseline LDCT in a 59-year-old man from NLST with a small infra-threshold (1.8mm in mean diameter) solid benign nodule in the right upper lobe (white arrow) at baseline LDCT who was alive 11 years after randomization. (B and C) Prediction of a moderate probability of LC after 1 year (risk score=0.3057) and 6 years (risk score=0.5998) since baseline LDCT in a 56-year-old woman from NLST with a small infra-threshold (3.8mm in mean diameter) (black arrow) solid nodule in the left upper lobe at baseline LDCT (B) which showed growth (9mm in mean diameter) at the annual LDCT performed two years later (C) consistent with a pseudo-incidental LC and who received a diagnosis of stage IA adenocarcinoma and was alive 11 years after randomization. (D and E) Prediction of a very low risk of LC after 1 (risk score=0.0017) and 6 years (risk score=0.0329) since baseline LDCT in a 57-year-old woman from NLST with a negative baseline LDCT (D) who showed a large (18mm in mean diameter) solid lesion (*) at the next annual LDCT (E) consistent with an incident LC who received a diagnosis of small cell carcinoma and died of LC (ICD code C349) 1559 days after randomization.

(0.11MB).

Tobacco smoking is a well-established risk factor for CVD, COPD, and LC. In LC screening cohorts, which primarily include current and former smokers, these conditions are the leading causes of death,2,3,91,92 and are often referred as the ‘Big 3 killers’. Using AI to extract comorbidity-related biomarkers from baseline LDCT images offers a valuable opportunity to enhance LC screening. AI can help optimize screening schedules – such as determining when to start, how frequently to screen, and when to stop – by refining individual risk profiles.9 Although radiologists’ visual scoring of comorbidities provides adequate predictive values25,35 (Figs. 2 and 3), AI-derived biomarkers offer greater robustness and objectivity, all without increasing the clinician's workload.93 One significant proof of this concept is a DL algorithm for the automatic quantification of coronary calcium.94 The resulting calcium scoring showed a high correlation with readings from expert radiologists and demonstrated robust test-retest accuracy.94 Beyond using AI-derived CAC as a predictor of CV events in LC screening cohorts,94–97 researchers are exploring additional approaches. For instance, a model was developed that based on the extraction of the coronary calcium and juxta-cardiac fat uses a single LDCT examination and provides a 0–1 score to estimate the probability of CVD risk.98 The model's ability to predict the risk of CVD and CV mortality equalized or surpassed that of radiologists and surpassed that of other state-of-the-art DL tools.94,98,99 Examples of its application to predict CV death in subjects with no or mild CAC are shown in Fig. 5. Other DL-derived indices include the prediction of adverse events based on the left atrial volume100 and of CV risk based on epicardial adipose tissue amount alone.101

Fig. 5.

(A and B) Assessment of risk of CV disease based on the analysis of baseline LDCT with Chao et al. deep learning algorithm.98 The algorithm attributes a moderate (score=0.351) CV risk in a 55-year-old man from NLST who did not show any coronary artery calcification at baseline LDCT (A) and who died of ischemic heart disease (ICD code I250) 2004 days after randomization. The algorithm attributes a high (score=0.700) CV risk in a 70-year-old woman from NLST with mild coronary artery calcifications (white arrow) at baseline LDCT (C) and who died of acute myocardial infarct (ICD code I219) 511 days after randomization.

(0.14MB).

COPD is typically diagnosed and evaluated through symptom assessment, spirometric testing, and tracking respiratory exacerbations.102 While lung densitometry is more reproducible than visual assessment of emphysema103 and is increasingly used for COPD assessment,29,31,104 it is notably sensitive to variations in CT scanners and acquisition/reconstruction parameters, such as slice thickness, radiation dose, and reconstruction kernel.105 A two-step DL model was developed to normalize the kernel effect for emphysema quantification in LDCT Images.105,106 This tool allows accurate emphysema quantification even when images are reconstructed using different kernels, thus improving consistency across large screening trials.105

Combining quantitative and semi-quantitative biomarkers for CVD and COPD in risk stratification after LDCT examinations is gaining attention. A logistic regression model that integrates participant demographics with LDCT measures of LC, CVD and COPD was developed to predict the 5-year risk of competing death. This approach helps identify individuals who may benefit more from preventive care for other conditions than from LC screening.107 The results suggest that a model based exclusively on quantitative LDCT measures, even when automatically derived, is suitable for calculating risk scores in a LC screening cohort and informing the post-LDCT screening process. Similarly, the predictive value of CAC visual score and of densitometry assessment of emphysema (relative area of the lung with density below −950 Hounsfield Units – RA950) in baseline LDCT along with age, gender, smoking status and pack-years were evaluated to predict the overall, LC, and CVD mortality in a screening cohort.36 Using an ML paradigm based on decision trees108 and the SHAP framework109 to assess the importance of each feature, the model interpretation revealed that RA950 was the first ranking feature for predicting overall and CVD mortality, with AUROC values of 0.70 and 0.73, respectively. The most important features for predicting LC mortality were pack-years and RA950, with an AUROC of 0.61.

COPD is frequently associated with other extra-pulmonary systemic manifestations, including osteoporosis,110 that leads to an increased risk of fractures.111 Since bone attenuation measured on routine chest CT has shown strong correlation with bone mass density (BMD) assessed by dual-energy X-ray absorptiometry (DXA) in patients with COPD,112 opportunistic DL-aided assessment of osteoporosis in LDCT scans in LC screening cohorts has emerged.

Different approaches have been proposed. A DL model was combined with geometric operations to automatically measure BMD from LDCT scans achieving a good agreement with quantitative CT.113 AI-RAD Companion was evaluated as an end-to-end solution to derive a LDCT biomarker for osteoporosis in LC screening whose score moderately correlated with WHO T-scores allowing to stratify participants into normal, osteopenia, and osteoporosis categories.114 Additionally, the combination of ML with radiomics texture analysis of automatically detected vertebral body achieved an AUROC of 0.90 and 0.72 on internal and external validation cohorts, respectively,115 establishing that osteoporosis can be part of the evaluation of LDCT for LC screening with impacts on morbidity, mortality, and the overall efficacy of LC screening.

Recently, several studies have demonstrated the capability of DL algorithms to help classify the ILD detected in full dose thin-section CT116–119 and, more importantly, to predict the progression of the disease and the mortality due to this condition.120–122 Validation of these algorithms in the LDCT examinations performed for LC screening is still required.