Explainable hybrid ensemble framework for lung cancer classification using deep CNNs and random forest

Lung cancer is still a significant cause of cancer death in the world. Thoracic computed tomography (CT) classification can be helpful in earlier diagnosis and better patient management. However, many deep-learning techniques require large annotated datasets, large amounts of computational resources and the techniques are not easily interpretable. This manuscript introduces a comprehensible hybrid approach to multiclass lung cancer classification based on two different transfer-learning backbones, DenseNet169 and MobileNetV2. They used several ensemble techniques to combine their class-probability outputs. For qualitative visual explanations of the image regions contributing to each prediction, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the image after inference. The framework was evaluated with two publicly available thoracic CT datasets that include both benign and malignant, normal and non-small-cell lung cancer categories. The accuracy of the Random Forest stacker for the Chest CT dataset was 99.50% and 99.54% for the IQ-OTH/NCCD dataset. The results show that DenseNet169 and MobileNetV2 complementary probability outputs could be used as an effective decision-fusion method in lung cancer classification from different datasets with different class structures and diagnostic categories. The Grad-CAM technique also made the model predictions more transparent by identifying image areas that influenced the prediction. In general, the proposed framework presents a practical and comprehensible method for automated lung cancer classification, and serves as a valuable guideline in future studies regarding the development of robust, explainable, and clinically relevant computer-aided diagnosis in medical imaging.