AI lung cancer screening shows promise but falls short of clinical autonomy

Deep learning systems that read low-dose CT scans can detect lung cancer with pooled sensitivity of 85.4 percent and pooled specificity of 83.5 percent, according to a systematic review and meta-analysis published in Diagnostic and Prognostic Research. The figures sound encouraging, but the authors of the review are careful about what they mean. Only five studies out of eleven met the strict requirements needed for quantitative pooling, and none of the included studies was free of risk of bias across all assessed domains. The verdict of the review is that artificial intelligence assistance for lung cancer screening is promising yet preliminary, best suited to a supporting role alongside radiologists rather than autonomous diagnosis.

The review, led by Dagem Gebremichael and colleagues, set out to answer a deceptively simple question: how accurately do deep learning systems detect lung cancer at the level of individual patients in screening settings? The distinction matters. Lung cancer was the most frequently diagnosed cancer and the leading cause of cancer death worldwide in 2022, with an estimated 2.5 million cases and 1.8 million deaths. Low-dose CT screening has been shown to reduce lung cancer mortality in high-risk populations in the landmark NLST and NELSON trials, and current guidelines endorse screening for eligible adults. But screening is not a simple image-classification problem, and that is where the technical subtleties begin.

Most nodules found on screening CT scans are benign. In the NLST trial, 96.4 percent of positive screens were false positives, meaning the vast majority of flagged scans led to follow-up imaging, surveillance, and sometimes invasive procedures for lesions that were never cancer. This is why specificity, the ability to correctly identify people without cancer, is so consequential in screening. A modest drop in specificity can cascade into more recalls, more radiation exposure from repeat imaging, and more biopsies of harmless nodules. The authors argue that clinical performance must therefore be judged within threshold-based management pathways such as Lung-RADS, which account for nodule characteristics, prior scans, and growth over time.

Deep learning models, in principle, offer a way to improve malignancy risk estimation by integrating three-dimensional imaging data, nodule morphology, whole-lung context, and longitudinal change between scans. Earlier work, including a well-known study by Ardila and colleagues, reported strong discrimination using both current and prior low-dose CT data, and subsequent models reinforced the value of longitudinal imaging. Yet the clinical value of any model depends critically on its intended role: whether it operates as a stand-alone classifier, a prescreening filter that rules out clearly negative scans, a second reader checking a radiologist’s work, or a concurrent assistant reading alongside the radiologist. Each deployment choice shifts the balance of sensitivity, specificity, radiologist workload, and downstream patient management.

Previous reviews on this topic have often mixed screening populations with symptomatic patients, incidental findings, nodule-enriched cohorts, and lesion-level analyses, which can inflate apparent performance. The new review deliberately restricted inclusion to deep learning-assisted low-dose CT in screening or screening-relevant populations, prioritised participant-level accuracy over nodule-level metrics, and separated eligibility for the systematic review from eligibility for meta-analysis. The researchers searched MEDLINE, Embase, and Web of Science from January 2010 to December 2025. Two reviewers independently screened the 2,207 records identified, assessed 56 full texts, and ultimately included eleven studies published between 2019 and 2025, of which eight were retrospective and three prospective, with seven being multicentre efforts.

The quantitative synthesis was deliberately conservative. To enter the meta-analysis, a study had to provide a complete or uniquely reconstructible patient-level two-by-two table at a defined operating threshold, meaning true positives, false positives, false negatives, and true negatives had to be recoverable for a specific decision threshold. The authors permitted reconstruction only when the disease-positive and disease-negative denominators and the reported sensitivity and specificity yielded a single internally consistent table, and they explicitly refused to infer counts from AUC values or rounded accuracy figures. Five studies, comprising 2,220 participants, 232 lung cancer cases, and 1,988 non-cases, cleared this bar. The remaining six studies were synthesised narratively because their threshold-specific patient-level data were unavailable or not reconstructible, even after the authors were contacted directly.

Across those five meta-analysed studies, which ranged from 100 to 1,139 participants, sensitivity spanned 75.0 to 95.5 percent and specificity spanned 64.0 to 92.7 percent, revealing substantial heterogeneity. The pooled estimates came from bivariate random-effects and hierarchical summary receiver operating characteristic models, which account for the correlation between sensitivity and specificity and for between-study variation. The positive likelihood ratio was 5.17, the negative likelihood ratio 0.18, and the diagnostic odds ratio 29.53. Notably, between-study variability was greater for specificity than for sensitivity, a pattern with direct clinical consequences given how sensitive screening outcomes are to false positive rates. The thresholds themselves differed across studies, from a Lung-RADS category of 3 or higher in the Ardila study to probability cutoffs, Youden-index-derived points, and a score threshold of 0.55, so the pooled values represent summary accuracy across different operating points rather than the performance of a single deployable algorithm.

The risk-of-bias picture was sobering. No study achieved low risk across all four QUADAS-2 domains. High risk of bias was most frequent for patient selection, affecting eight of eleven studies, and for the index test, affecting seven. Every single study, eleven of eleven, was judged at high risk in the flow and timing domain, and the reference standard was unclear in eight studies. An additional assessment of artificial intelligence reporting completeness, adapted from the CLAIM 2024 and STARD-AI guidelines, found that all studies described their data sources and ten of eleven reported data partitioning, but only three reported explainability methods, four addressed model availability, and not one clearly stated regulatory status. These reporting gaps limit reproducibility and complicate clinical translation.

The review also examined threshold-independent discrimination, reported as AUC values in ten of the eleven studies, with figures ranging from roughly 0.76 to 0.99. The authors caution that these values were not directly comparable because they came from different models, cohorts, prediction horizons, and units of analysis, and that AUC alone cannot define clinical action. A high AUC describes discrimination across all possible thresholds without specifying what recall rates, surveillance imaging, or biopsy burdens would follow at any chosen operating point. Planned subgroup analyses, meta-regression, and sensitivity analyses could not be performed because only five studies were quantitatively eligible, and although Deeks’ funnel plot asymmetry test found no statistical evidence of small-study effects, with a p-value of 0.368, the test was underpowered with so few studies and cannot exclude publication bias.

What, then, should happen next? The authors conclude that the evidence does not support autonomous AI diagnosis or AI-triggered invasive management, and that artificial intelligence should not replace radiologist interpretation or serve as a stand-alone rule-in test. The relatively favourable negative likelihood ratio supports rule-out applications more than rule-in ones, making prescreening and triage the most plausible near-term roles: flagging clearly negative examinations to reduce workload, or prioritising suspicious scans for urgent review. Workflow evidence, though limited, is supportive, with concurrent AI improving specificity without reducing sensitivity in multinational reader studies and prescreening showing favourable net benefit. The authors recommend that future studies be prospective, multicentre, and externally validated, that they follow reporting standards including CLAIM, STARD-AI, TRIPOD + AI, and DECIDE-AI, and that they compare human-only, AI-only, and human-AI strategies while reporting recalls, biopsies, interval cancers, workload, and cost-effectiveness. Until such evidence arrives, deep learning in lung cancer screening remains a decision-support tool within radiologist-led pathways, awaiting the prospective validation that will determine whether its laboratory promise survives contact with real clinical workflows.

Subject of Research: Diagnostic accuracy of deep learning-assisted low-dose CT screening for lung cancer detection

Article Title: Diagnostic accuracy of deep learning-assisted low-dose CT screening for lung cancer detection: a systematic review and meta-analysis

Article References: Gebremichael, D., Zimme, Z. A., Mengesha, L. Y., Jibat, N., Berihun, A., Ayele, A. A., Ashenafi, S. A., Tegegn, H. G., & Gete, K. Y. (2026). Diagnostic accuracy of deep learning-assisted low-dose CT screening for lung cancer detection: a systematic review and meta-analysis. Diagnostic and Prognostic Research, 10(1), Article 21. https://doi.org/10.1186/s41512-026-00235-w

Image Credits: AI Generated

DOI: 10.1186/s41512-026-00235-w

Keywords: deep learning, low-dose CT, lung cancer, screening, diagnostic accuracy, meta-analysis, QUADAS-2, sensitivity, specificity, radiology, artificial intelligence, Lung-RADS