Ma, Hongwei

8021 Background: Lung cancer is the leading cause of cancer-related mortality worldwide, primarily due to delayed diagnosis and diagnostic variability in histopathological assessment. Advances in deep learning applied to medical imaging offer the potential to enhance early detection accuracy, reduce interobserver variability, and improve diagnostic efficiency in lung cancer. Methods: In this retrospective study, we evaluated DAVINCI using a dataset of 8,962 patients with confirmed lung malignancies. The model utilizes a specialized architecture combining convolutional neural networks (CNNs) and residual blocks to optimize feature extraction from high-resolution pathological slides. Performance was measured via detection accuracy, recall, specificity, and Area Under the Curve (AUC). Results: A total of 8,962 patients were included in the analysis. The DAVINCI model correctly identified lung cancer in 93% of cases, corresponding to 8,334 of 8,962 patients, with a 95% confidence interval (CI) of 86%–100% (7,707–8,962 patients). Model specificity was also 93% (95% CI, 86%–100%), indicating a high true-negative rate. Sensitivity (recall) was 87%, with correct detection in 7,797 patients and a 95% CI of 80%–94% (7,170–8,425 patients). The model demonstrated strong discriminatory performance, achieving an area under the curve (AUC) of 0.91 (95% CI, 0.84–0.98). In comparison, active pathologists achieved a mean diagnostic accuracy of 83%, corresponding to 7,438 of 8,962 cases, with a 95% CI of 76%–90% (6,811–8,066 cases). In terms of efficiency, DAVINCI processed whole-slide images in 16–19 seconds, substantially faster than conventional manual review, which typically requires several minutes per slide. Conclusions: DAVINCI demonstrated superior accuracy, robust discriminatory performance, and markedly faster processing compared with human review, supporting its potential role in early lung cancer detection at scale.