Accuracy and Impact of AI Software for Nodule and Cancer Detection at Lung Cancer Screening CT

Take-Away Points

• ■ Major Focus: To systematically evaluate the accuracy and impact of Conformité Européenne–marked (CE-marked) artificial intelligence (AI) software for lung nodule and cancer detection at lung cancer screening CT examination.

• ■ Key Results: In a systematic review of 11 studies, AI assistance at CT lung cancer screening resulted in faster reading times and improved sensitivity for detecting actionable and malignant nodules at the expense of decreased specificity.

• ■ Impact: AI-assisted lung cancer screening software shows promise for reducing workload and increasing lung cancer detection rates. Specificity of these systems must be improved to prevent unnecessary surveillance and interventions.

Randomized controlled trials show decreased lung cancer–specific mortality in individuals who undergo low-dose CT screening. Various AI algorithms have been developed to assist in lung cancer detection during screening, demonstrating potential to reduce radiologist reading times and workload. However, possible harm for screening participants or additional workload should be considered.

Geppert et al systematically searched multiple databases for studies published from 2012 to March 2023 that reported on accuracy and impact of CE-marked AI software for lung nodule detection at CT lung cancer screening. Secondary outcomes included reading time and Lung CT Screening Reporting and Data System (Lung-RADS) categorization. Readings by experienced radiologists served as the reference standard for presence or absence of a lung nodule. The analysis included 11 studies comprising 19 770 screened participants from six different countries and six different CE-marked AI software packages for CT lung cancer screening. In comparison to unaided reading, AI-assisted reading improved sensitivity for detecting and categorizing actionable nodules (+5% to +20%) and malignant nodules (+3% to +15%). Specificity decreased for detecting and categorizing individuals without actionable nodules (-7% to -3%) and malignant nodules (-8% to -6%). Across three multireader, multicase studies, AI assistance resulted in significantly faster reading times and increased proportions of actionable nodules (classified as Lung-RADS 3–4).

This study found that AI software used in lung cancer screening reduces radiologist reading times and increases lung cancer detection rates but at the cost of more false-positive findings and potential unnecessary surveillance. Further innovations may enhance the specificity of AI software for CT-based lung cancer screening.

Highlighted Article

• Geppert J, Asgharzadeh A, Brown A, et al. Software using artificial intelligence for nodule and cancer detection in CT lung cancer screening: systematic review of test accuracy studies. Thorax 2024;79(11):1040–1049. doi: https://doi.org/10.1136/thorax-2024-221662