Lung cancer remains a leading cause of cancer mortality worldwide, underscoring the need for improved early detection methods. While deep learning models, particularly Convolutional
Neural Networks (CNNs), have shown promise in analyzing low-dose computed tomography (LDCT) scans for pulmonary nodule classification, a comparative analysis of modern CNN architectures for this specific task is limited. This study aims to conduct a rigorous head-to-head evaluation of ResNet, DenseNet, and EfficientNet for benign-malignant classification of lung nodules. Using the publicly available LIDC-IDRI dataset of thoracic CT scans, we developed a deep learning pipeline for binary nodule classification. The dataset was preprocessed, and ۸,۴۲۰ annotated nodule patches were extracted and augmented. Three CNN architectures—ResNet۵۰, DenseNet۱۲۱, and EfficientNetB۰—were initialized with ImageNet pre-trained weights and fine-tuned. Model performance was evaluated on a held-out test set using accuracy, precision, recall, F۱-score, and the area under the receiver operating characteristic curve (AUC). Interpretability was assessed using Grad-CAM visualizations. DenseNet۱۲۱ achieved the highest performance with an accuracy of ۹۵.۶% and an AUC of ۰.۹۸۲, demonstrating a superior balance of sensitivity and specificity and producing the fewest false negatives. EfficientNetB۰ followed closely (AUC: ۰.۹۷۶) with the shortest training time, while ResNet۵۰ was a strong but slightly less sensitive baseline (AUC: ۰.۹۶۴). Grad-CAM analysis confirmed that all models focused on relevant nodule regions, with DenseNet and EfficientNet exhibiting more localized activations. This study demonstrates that architectural choice significantly impacts performance in lung nodule malignancy classification. DenseNet emerged as the most effective backbone for this task, offering high discriminatory power and critical sensitivity to malignant cases.
