Impact of Artificial Intelligence Assistance on Radiologist Performance in CT-Based Pulmonary Nodule Detection: A Systematic Review and Meta-Analysis of Detection Accuracy Studies.
Authors
Affiliations (2)
Affiliations (2)
- Faculty of Medicine, Medical University of Varna, 9002 Varna, Bulgaria.
- Department of Orthopaedic Surgery, East and North Hertfordshire NHS Trust, Stevenage SG1 4AB, UK.
Abstract
Lung cancer remains a major cause of cancer-related death; prompt detection and monitoring of pulmonary nodules through computed tomography (CT) is critical to prevent mortality. CT interpretation is time-intensive and can be prone to fatigue-related errors. The application of artificial intelligence (AI) may aid in the enhanced detection of pulmonary nodules when used in conjunction with radiologists. Thus, possibly decreasing interpretation time and enhancing diagnostic reliability. However, studies have reported uncertainty regarding the extent to which true positives are always diagnosed, as studies have indicated an increase in false positive rates with AI-assisted pulmonary nodule detection. This systematic review and meta-analysis focuses on the impact of AI assistance on sensitivity and false positive rates in comparison to unaided reading. Adhering to the PRISMA guidelines, an intensive literature search was carried out on Cochrane, PubMed, Medline and Embase. Research composed of studies that compared AI-assisted reading to unaided radiologists in the context of nodule detection, which also provided additional outcomes that were utilised in the secondary outcome analysis. ROBINS-1 tool was used to assess risk of bias and data quality. Risk ratios and risk differences with 95% CI were reported for dichotomous outcomes, whereas mean differences with 95% CI were utilised for continuous outcomes. Sensitivity analyses were performed for the imputation of missing standard deviations. Seven studies reported mean sensitivity in either second reader (SR) mode or concurrent reader (CR) mode. Five of the included studies implemented second reader designs; AI-assisted sensitivity ranged from 38% to 85.66%, whereas unassisted reading sensitivity ranged from 20.9% to 73.46%. Additionally, three studies reported concurrent reading, with AI-assisted reading ranging from 80% to 84%, while unassisted reading varied from 64% to 72%. According to our research, AI-assisted reading had a significantly higher sensitivity (RR of 1.39; CI 95% 1.08-1.79; <i>p</i> = 0.01; I<sup>2</sup> = 95%). Five studies reported false positives per scan. Four reported a numerical increase with AI assistance, but incomplete variance reporting and study heterogeneity precluded quantitative pooling. Reading time was increased under AI assistance as a second reader (MD 77.54; CI 95% -57.50 to 212.6, <i>p</i> = 0.132; I<sup>2</sup> = 64.9%) but did not yield significance. Additionally, AI-assisted reading under concurrent reading did not yield statistical significance for reading time (MD -11.39; 95% CI -498.7 to 476.0, <i>p</i> = 0.816; I<sup>2</sup> = 41.3%). AI-assisted reading significantly improves sensitivity for pulmonary nodule detection compared to unaided reading, but the trade-offs with false positives and workflow disruption remain uncertain. Substantial heterogeneity and the limited number of studies preclude definitive clinical recommendations. These findings should be considered exploratory, and the implementation of AI in lung cancer screening requires prospective validation.