Background and Aims: Lung cancer is one of the most lethal cancers worldwide, and early detection plays a vital role in improving treatment outcomes. Due to the challenges associated with manual analysis of CT scan images,
Convolutional Neural Networks (CNNs) have been considered a precise and automated solution for detecting pulmonary lesions. This study aims to systematically evaluate the performance of
CNN models in detecting lung cancer from CT scans and to identify methodological trends. Methods: This scoping review was conducted in accordance with PRISMA guidelines. Articles were retrieved from PubMed and Web of Science databases using the keywords 'CNN', 'Lung Cancer', 'Detection', and 'CT Scan' up to the year ۲۰۲۴. Original studies that focused on using
CNN for lung cancer detection and were published in English were included. Review papers, editor letters, and protocols were excluded. Key features extracted from the studies included evaluation metrics (accuracy, sensitivity, AUC), datasets used, types of algorithms, and the most effective
CNN models for lung cancer detection. Results: Out of ۳۰ identified studies, ۱۱ met the inclusion criteria and were analyzed in this scoping review. The most commonly used
CNN models were
InceptionResNetV۲ and ResNet۱۵۲V۲. On average, these models achieved an accuracy of ۹۹.۶۴%, sensitivity of ۹۷.۸۴%, specificity of ۹۷.۲%, and AUC ranging from ۰.۹۳۶ to ۰.۹۶۱, indicating high performance in detecting lung cancer. The most frequently used datasets were
LUNA۱۶ (۵۴.۵۴%) and
LIDC-IDRI (۴۵.۴۵%). Several studies showed that combining CNNs with optimization techniques such as
Particle Swarm Optimization (PSO) or hybrid models like
UNet + Transformer significantly enhanced diagnostic performance. Additionally, using
SMOTE to address data imbalance proved effective in some studies. Conclusion: The reviewed studies demonstrate that CNNs, particularly advanced models like
InceptionResNetV۲ and ResNet۱۵۲V۲, show strong potential for detecting lung cancer from CT scan images. This can lead to earlier diagnosis and improved treatment outcomes for patients. However, there remains a need to develop lightweight models suitable for clinical settings and to integrate hybrid approaches for improving accuracy in early-stage detection. The findings of this study may serve as a guide for designing AI-based systems for lung cancer screening.
