Answers are grounded in this paper and cite their sources.
Background: Pulmonary nodules are common on chest computed tomography (CT), and small or subtle lesions may be missed because of reader variability and workload. Deep learning–based computer-aided detection (CAD) may assist radiologists, but its effect on detection, false-positive findings, reading time and reader agreement needs evaluation in routine tertiary care practice. The aim is to evaluate artificial intelligence (AI)–assisted radiologist interpretation of chest CT for pulmonary nodule detection and to compare it with conventional interpretation. Materials and Methods: This retrospective, single-centre, paired multi-reader diagnostic accuracy study, reported according to STARD, included 300 adult chest CT examinations performed between January 2025 and June 2026 in the Department of Radiodiagnosis of a tertiary care teaching hospital in India. Two radiologists read every examination twice, once conventionally and once with a deep learning CAD system, separated by a washout of at least four weeks. The reference standard was consensus review by two experienced thoracic radiologists. The primary outcome was nodule-level sensitivity. Secondary outcomes were patient-level sensitivity, specificity, positive and negative predictive values, accuracy, false-positive findings per examination, interpretation time and interobserver agreement (Cohen’s kappa). Paired comparisons used McNemar’s test and the paired t-test or Wilcoxon signed-rank test; p < 0.05 was considered significant. Results: (Illustrative placeholder values, to be replaced with verified study data before submission.) The 300 examinations included 180 nodule-positive examinations (60.0%) containing 280 reference nodules. AI assistance increased nodule-level sensitivity from 77.14% to 88.93% and patient-level sensitivity from 83.33% to 93.33%; the gain was largest for nodules smaller than 6 mm (66.4% to 83.6%). Patient-level specificity changed from 88.33% to 85.00%, accuracy from 85.33% to 90.00%, and false-positive findings per examination from 0.093 to 0.130. Mean interpretation time fell from 12.90 to 10.49 minutes (18.7% reduction), and Cohen’s kappa rose from 0.771 to 0.866. Conclusion: In this dataset, AI assistance improved nodule detection, particularly for small nodules, and was associated with shorter reading time and better interobserver agreement, at the cost of more false-positive findings. AI should be used as a supplementary tool under radiologist supervision, and prospective multicentre validation is needed
No takes yet. Share an insight, caveat, or question.
Kuttikkat et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: