Despite advances in lung cancer treatment, especially targeted therapies for genetic mutations, the disease remains one of deadliest in the world. Lung cancer is often diagnosed at an advanced stage, as most patients do not develop symptoms until late in the disease. In 2022, lung cancer caused 1.23 million deaths globally, accounting for 18.7% of all cancer fatalities.1
As with most cancers, earlier diagnosis leads to better outcomes. To drive earlier detection, low-dose computed tomography (LDCT) scanning has become the most widely used lung cancer screening tool in use today.
Since LDCT was introduced in the US for high-risk individuals in 2011, the proportion of cases diagnosed at a localized stage has steadily increased. By 2016, for the first time, localized diagnoses surpassed those diagnosed at the regional stage.2 Expanded screening could increase early-stage lung cancer diagnoses by up to 70%, as well as lead to the detection of more cases of carcinoma in situ and other precursor lesions.3,5 Notably, chest radiography (x-rays), with or without sputum cytologic examination, showed no improvement in patient outcomes.6
Trials showed that LDCT screening reduces lung cancer mortality in smokers by 20% to 24%. However, it comes with challenges, both clinically and in the context of risk assessment.
From a clinical perspective, implementation of LDCT lung cancer screening requires resources, systems, and processes to be put in place to manage the findings – particularly uncertain results – effectively.3 Additionally, there are personal and economic implications associated with rolling out screening programs.
Another key consideration is the risk of cumulative radiation exposure, particularly for patients requiring ongoing surveillance of indeterminate nodules. Ultra-low-dose computed tomography (ULDCT) reduces radiation exposure of 0.22mSv vs. 1.5mSv with LDCT. Combined with image reconstruction algorithms, ULDCT significantly mitigates radiation exposure risk while maintaining high-quality imaging for solid nodules larger than 5mm. However, for smaller and non-solid nodules, further optimization may be required, limiting both the clinical and risk assessment benefits of this imaging advancement at the current time. Another advancement, photon counting detector CT (PCD-CT), offers the potential to further reduce radiation exposure with even more enhanced image resolution.7
From an insurance perspective, three key considerations stand out in LDCT lung cancer screening. The first is assessing the likelihood of current or future malignancy in a detected lung nodule. Most underwriting guidelines take a risk-based approach, evaluating both nodule characteristics and an applicant’s individual risk factors. In high-risk groups, LDCT frequently detects lung nodules, yet only a small percentage eventually prove to be malignant.3,4 False-positive rates can reach up to 13%. To address this, standardized algorithms in radiology reporting have been developed to reduce discrepancies. Prudent risk assessment must balance clinical protocols, disease prevalence, likelihood of malignancy, and risk appetite in the event of positive findings.
Beyond identifying high-risk nodules, LDCT screening frequently detects incidental findings that may require further clinical evaluation and risk assessment. The rate at which these findings are reported varies significantly, ranging from 23% in community centers to 67% in university settings. Despite these differences, initial lung cancer screening trials found that fewer than 10% of incidental findings were clinically significant, with coronary artery calcification and emphysema being the most prevalent and relevant.3 Both conditions are independent risk factors for overall mortality and lung cancer, necessitating follow-up and further assessment.
Interstitial lung disease (ILD) is another significant incidental finding due to its potential impact on mortality and morbidity, also requiring further follow-up. Other lung-related incidental findings include bronchiectasis, tuberculosis, and pleural plaques/thickening and effusions. Osteopenia and osteoporosis may also be detected, although bone density is not always reported. If identified – particularly with associated vertebral fracture – they become noteworthy from a risk assessment perspective. As reporting standards for incidental findings evolve, they will help clarify the clinical and risk assessment significance of these findings.
Finally, overdiagnosis remains a concern from both an incidence rate trend and critical illness perspective, owing to increased claims and possible need for definition review.3 Quantifying overdiagnosis is challenging and may be underestimated in randomized population-level trials. Still, data from lung cancer screening trials provide some insight. In the National Lung Screening Trial (NLST), the probability that lung cancer detected by LDCT was an overdiagnosis was estimated at 18.5% initially after a mean follow-up of 4.5 years, decreasing to 3% after 9 years. The NELSON trial reported similar findings, with estimates of 19.7% and 8.9% at 4.5 and 5.5 years after final screening.3
Minimizing overdiagnosis – and the resulting overutilization and overtreatment – is essential. However, when screening occurs outside of structured population screening protocols, the impact becomes more difficult to quantify.
Computer-aided detection (CAD) tools have long assisted radiologists as primary, secondary, or parallel image readers, relying on predefined rules and thresholds. More advanced AI systems, however, can autonomously detect, measure, segment, and classify nodules or other potentially malignant structures, reducing the burden on radiologists. Unlike CAD, these AI models develop their algorithms from vast imaging and clinical datasets, rather than relying on fixed rule sets.
Further advancements in AI, including deep-learning techniques such as convolutional neural networks and radiomics, support clinical problem-solving by optimizing patient management, improving risk stratification, personalizing screening, and improving reporting efficiency.
Radiomics, a rapidly evolving field, enables AI to extract additional morphological data from radiological images. Unlike other “omics” fields (genomics, proteomics, transcriptomics), radiomics analyzes imaging-based qualitative features rather than requiring invasive biopsies or molecular assays.8
While AI improves sensitivity in lung cancer screening – detecting more actionable and malignant nodules – this comes at the cost of specificity, leading to the detection of nodules that may never develop into cancer.
To quantify this potential impact, a recent systematic review of six AI-based software systems found that at an assumed prevalence of 0.5%, AI would detect an additional 150 to 750 cancers per million people screened but also result in an additional 59,700 to 76,000 people being unnecessarily surveilled.9 A recently published systematic review and meta-analysis of 39 studies of AI-based imaging for lung cancer screening found a pooled sensitivity and specificity of 0.87 across various AI types, regions, and data sources. While authors viewed this as promising, despite observed heterogeneity in study parameters, they emphasized the need for standardized protocols to maximize AI’s potential.10
Most lung cancer screening programs focus on high-risk populations, which usually include older (current or past) smokers. However, across Asia, LDCT screening is often performed more broadly, including for younger individuals without known risk factors as part of routine “health check-ups.”
A retrospective study from a single center in Sichuan, China, examined AI-assisted LDCT scan readings in individuals undergoing routine physical examinations.11 The findings were particularly relevant because the study cohort – comprising over 23,000 participants, including smokers, non-smokers, and some with a family history of lung cancer – aligned more closely with an insured population. AI scanning outperformed manual readings in detecting and differentiating lung nodules. More importantly, of the 0.8% confirmed lung cancer cases, 74% were adenocarcinoma, with most identified at early stages.
The next exciting frontier in AI for lung cancer screening may already be here. A tool called Sybil, developed by researchers at MIT and Massachusetts General Hospital (MGH), predicts lung cancer risk up to six years in advance using a single LDCT image – without a radiologist or additional demographic or clinical data. Remarkably, Sybil can make predictions even from scans that lack visible imaging features, such as lung nodules.12
This breakthrough surpasses current technology and radiologist assessment alone, paving the way for more personalized screening that identifies risk where it is most beneficial.
Sybil’s deep-learning model was trained on datasets from the NLST, MGH, and Chang Gung Memorial Hospital (CGMH) in Taiwan. The inclusion of CGMH is significant, as Taiwan screens non-smokers, making the model more generalizable to this population.
Globally, lung cancer is being increasingly diagnosed in non-smokers, particularly in women. Over 50% of female lung cancer cases occur in non-smokers, compared to 15% to 20% in men. Despite this trend, lung cancer screening remains underutilized for various reasons. AI models like Sybil could help bridge the gap between broader screening initiatives and more targeted efforts for at-risk populations.
Virtual biopsies are an emerging concept that non-invasively assess intra-tumoral heterogeneity and clinical outcomes from images. Radiogenomics and radioproteomics explore the combination of imaging phenotypes (graphic characteristics) with genetic and protein expression patterns to evaluate biological features related to clinical outcomes, potentially replacing the need for surgical biopsies and histopathological analysis. This advancement could also enable more personalized care.13
For example, Lee et al. identified nine radiomic “fingerprints” across the lung cancer imaging spectrum and developed a machine-learning virtual biopsy tool capable of accurately detecting adenocarcinoma invasion.14
Liquid biomarkers, including autoantibodies, and blood genomic and protein profiling, are also promising medical advances. While these tests are still undergoing validation, their integration with imaging could further optimize risk prediction models, improve diagnostic accuracy of indeterminate pulmonary nodules, and ultimately also reduce false-positive results.
As healthcare professionals and technology continue to co-evolve, AI-assisted early lung cancer detection tools will play an increasingly significant role – especially as more are clinically validated in large-scale prospective studies. The challenge for the industry will be determining how to best apply these tools to risk assessment while keeping pace with rapid technological advancements and their impact.
RGA experts are eager to engage with clients to better understand and tackle the industry’s most pressing challenges together. Contact us to learn more about RGA's capabilities, resources, and solutions.
