Explainable Transfer Learning for Lung Cancer Classification from CT Images Using Grad-CAM

Lung CT imaging plays an important role in identifying lung abnormalities, but deep learning models for medical image classification often face two practical challenges: class imbalance and limited interpretability. This study presents an explainable transfer learning framework for classifying lung CT images into normal, benign, and malignant categories. The IQ-OTH/NCCD lung cancer CT dataset was used, and the processed experimental set included 1097 CT images: 561 malignant, 416 normal, and 120 benign images. The dataset was divided into stratified training, validation, and test sets to preserve the class distribution. Three ImageNet-pretrained transfer learning models, ResNet50, DenseNet121, and EfficientNetB0, were trained and compared under the same preprocessing conditions. A fine-tuned ResNet50 model was also evaluated. Class weights and moderate data augmentation were applied to reduce the effect of class imbalance. Model performance was assessed using accuracy, precision, recall, F1-score, Macro F1, weighted F1, and Macro ROC-AUC. Grad-CAM was used to visualize model attention for correct predictions and high-confidence wrong predictions. ResNet50 achieved the best Macro F1 score of 0.7150, with an accuracy of 0.8364 and Macro ROC-AUC of 0.9064. Although fine-tuned ResNet50 achieved slightly higher accuracy, its Macro F1 decreased due to poor benign-class recall. Error analysis showed that benign cases were most frequently misclassified as normal. These results show that accuracy alone can be misleading for imbalanced lung CT classification. The study highlights the importance of combining class-sensitive evaluation metrics with explainability methods to better understand model behavior in AI-assisted tomography applications.