Comparison of deep learning models in detection of lung cancer

Lung cancer is one of the cancer types with the highest mortality rates worldwide, and delays in diagnosis significantly reduce patient survival. In this study, modern deep learning architectures (VGG16, ResNet-50, EfficientNet-B0, MobileNetV2, and InceptionV3) were comparatively evaluated under a standardized and data leakage–free experimental protocol for the automatic detection of lung cancer in computed tomography (CT) images. The study employed the international and publicly available IQ OTH/NCCD dataset collected from two major oncology centers in Iraq. The dataset contains three classes comprising a total of 1097 CT slices, including 561 Malignant, 120 Benign, and 416 Normal cases. To obtain a realistic clinical scenario and prevent data leakage, all CT images were stratified into training, validation, and locked test subsets at ratios of 70 percent, 15 percent, and 15 percent, respectively, before any preprocessing or augmentation was applied. Class imbalance was addressed only in the training subset using a fixed-target oversampling strategy, increasing each class to 1000 images, while no augmentation or balancing was applied to the validation and test subsets. During the preprocessing stage, CT images were subjected to windowing, followed by Otsu thresholding, morphological opening and closing operations, hole filling, and single-pixel erosion to automatically segment the lung regions. All deep learning models were initialized with ImageNet weights and retrained using transfer learning in the MATLAB R2024b environment. The models were evaluated on the locked test dataset using statistical performance metrics including accuracy, sensitivity, precision, specificity, F1 score, and ROC-AUC. According to the results, the EfficientNet-B0 model achieved the highest performance with an accuracy of 98.17 percent. The remaining models achieved accuracies of 96.95 percent for InceptionV3, 96.34 percent for MobileNetV2, 93.90 percent for VGG16, and 89.63 percent for ResNet-50. Sensitivity for the clinically critical Malignant class reached its highest value of 98.81 percent with the EfficientNet-B0 model. Moreover, all models achieved 100 percent precision for the Malignant class, indicating that none of the samples classified as cancerous were false positives. Grad CAM-based explainable artificial intelligence analyses demonstrated that the networks based their decisions on clinically meaningful regions such as nodules, opacities, and lesions. In conclusion, this study provides a reliable and reproducible evaluation framework by comparatively analyzing five deep learning architectures on multi-center real CT images under a standardized training protocol that eliminates data leakage and preserves class balance, thereby contributing robust evidence to the literature.