Abstract
Background Lung cancer is the leading cause of cancer-related mortality worldwide, with late-stage diagnosis contributing to poor survival rates. Early detection remains a critical challenge, hindered by diagnostic delays, radiologist shortages, and the limitations of current imaging workflows. Recent advances in artificial intelligence (AI), particularly deep learning, offer new avenues to enhance diagnostic accuracy and efficiency in radiology. Objective To develop and evaluate a deep learning model integrating Residual Network 50 Version 2 (ResNet50V2) with Squeeze-and-Excitation (SE) blocks for automated classification of lung cancer subtypes from computed tomography (CT) images. Methods A total of 1,000 anonymized lung CT images were obtained from a publicly available Kaggle dataset, categorized into four classes: adenocarcinoma, large cell carcinoma, squamous cell carcinoma, and normal tissue. The dataset was split into training (70%), validation (10%), and test (20%) sets. A fine-tuned ResNet50V2 architecture with SE blocks was used to enhance channel-wise feature recalibration. The model was trained using categorical cross-entropy loss with label smoothing and optimized via Adam. Performance was evaluated using accuracy, area under the receiver operating characteristic curve (AUC), precision, recall, and F1-score. Results The model achieved a test accuracy of 90.16% and an overall AUC of 0.9815. Class-wise AUCs were high across all categories: 0.9523 for adenocarcinoma, 0.9879 for large cell carcinoma, 0.9977 for normal tissue, and 0.9880 for squamous cell carcinoma. Precision ranged from 0.81 (large cell carcinoma) to 1.00 (normal tissue), while recall ranged from 0.85 (adenocarcinoma) to 0.98 (large cell carcinoma). F1-scores were consistently strong, ranging from 0.88 to 0.96. Conclusion The integration of SE blocks with ResNet50V2 yielded a high-performing model capable of accurately classifying lung cancer subtypes from CT images. The approach shows promise for assisting radiologists in diagnostic decision-making, particularly in settings with limited expert availability. Future work should focus on external validation, model interpretability, and exploration of emerging architectures such as Vision Transformers for enhanced performance and clinical adoption.