Deep Learning-Based Lung Cancer Classification and Grad-Cam With Lime-Supported Explainability Analysis

This study aims to classify lung cancer using deep learning-based methods and to interpret the obtained model results using explainable artificial intelligence (XAI) approaches. The dataset used in this study consisted of 1933 chest computed tomography images, which were classified as normal, benign, and malignant. In the classification process, EfficientNet-B4, MobileNetV3-Large, and ResNet50 deep learning architectures were trained using a transfer learning approach, and their performance was evaluated using a 10-fold cross-validation method. The performance of the models was analyzed using accuracy, precision, recall, F1-score, and ROC-AUC metrics. According to the results obtained in the study, the MobileNetV3-Large model showed the highest overall performance with 97.05% accuracy and 99.70% ROC-AUC value. The EfficientNet-B4 and ResNet50 models also provided high and balanced performance values, achieving effective results in lung cancer classification. Grad-CAM++ and LIME methods were used to make model decisions more clinically reliable and interpretable. Grad-CAM++ analyses reveal that the models primarily focus on anatomically significant regions within the lung parenchyma during classification. LIME analyses, on the other hand, have enabled the superpixel-level explanation of local regions that contribute most to classroom decisions. Explainability maps obtained for normal, benign, and malignant classes showed that each class exhibited distinct spatial attention and contribution patterns. In conclusion, this study demonstrates that deep learning-based lung cancer classification offers a more reliable and transparent framework for clinical decision support systems, not only when evaluated with high-performance metrics but also when considered in conjunction with Grad-CAM++ and LIME-supported explainability analyses. The findings demonstrate that explainable artificial intelligence approaches play a significant role in improving model reliability in medical image analysis.