使用普遍可及、跨平台、AI驱动的应用程序进行肺癌智能检测（LUCID）

Introduction: Lung cancer represents the leading cause of cancer-related mortality globally. Only 29% achieve 5-year survival, primarily due to late diagnoses. In addition, the economic burden exceeds $3.9 trillion internationally. Only one-third of all lung cancer patients are diagnosed early enough so that effective treatment can be implemented. Deep learning architectures have revolutionized diagnostic imaging, extracting complex imaging patterns beyond human capability. EfficientNetB0, a state-of-the-art convolutional neural network achieves remarkable accuracy and computational efficiency. Its application to computed tomography (CT) imaging offers unprecedented potential for early lung cancer detection and mortality reduction. Methods: Anonymized CT images of normal, benign, and malignant lung (n = 1190) were proportioned into training (60%), validation (20%), and testing (20%) sets. Preprocessing and augmentation improved data quality and model generalizability. EfficientNetB0 was trained and optimized, with performance evaluated by accuracy, precision-recall, F1-score, F2-score, and area under the receiver operating characteristic curve (AUROC) on validation and test sets. The trained model was deployed in a universally accessible, cross-platform application for independent validation by experts globally. Results: EfficientNetB0 demonstrated exceptional performance in distinguishing normal, benign, and malignant lung lesions. The model achieved perfect accuracy, precision, and recall for both classes, with F1-scores of 1.00 (training and validation). AUROC values of 1.00 demonstrate outstanding discriminative capacity. The confusion matrix confirmed zero misclassifications across 137 lesions (24 benign, 113 malignant). The model's exemplary performance suggests substantial potential to augment clinical decision-making and advance equitable cancer care. Conclusion: EfficientNetB0 represents a paradigm-shifting diagnostic tool for lung cancer stratification on CT imaging, demonstrating clinical-grade accuracy that augments expert radiologist interpretation. The model's exceptional discriminative capacity, achieving high sensitivity and specificity, positions it as a transformative adjunct to conventional diagnostic workflows. Deployment via an accessible cross-platform application facilitates dissemination across resource-constrained environments, democratizing precision diagnostics. This deep learning architecture substantially augments early detection capabilities, potentially reducing mortality through timely clinical intervention. These findings establish a robust foundation for prospective validation studies and clinical integration to optimize lung cancer outcomes globally.