Deep learning driven early detection of lung cancer from CT images using transfer learning approaches

Lung cancer remains one of the foremost causes of cancer-related mortality worldwide, primarily due to its frequent late-stage diagnosis and the limited efficacy of available treatments at advanced stages. This study seeks to advance early detection through the application of deep learning techniques for the automated classification of lung CT images into benign, malignant, and normal categories. A comprehensive methodological framework was adopted, encompassing data acquisition, preprocessing, model development, training, validation, and performance evaluation, with patient-level data splitting employed to prevent slice-level data leakage. Among six tested architectures (VGG16, Custom CNN, MobileNetV2, ResNet50, InceptionV3, and EfficientNetB0), the fully fine-tuned VGG16 model optimized via transfer learning and trained on the IQ-OTH/NCCD lung cancer dataset exhibited the strongest overall performance, achieving a mean test accuracy of 89.39% ± 2.10% (test loss 0.329) across three independent runs, and 89.70% ± 2.75% under 5-fold patient-grouped cross-validation. One-way ANOVA (F = 17.55, p < 0.0001) followed by Tukey's HSD post-hoc test confirmed that VGG16 significantly outperformed ResNet50 and EfficientNetB0. To enhance interpretability, Grad-CAM visualizations were generated for all three classes, indicating that the model's attention broadly corresponded to anatomically relevant lung regions. Trained and evaluated on a Google Colab GPU environment, the proposed system demonstrates potential as an assistive tool for automated lung cancer screening, warranting further validation on larger and multi-institutional datasets before clinical application.