SEMLCC: A Stacked Ensemble Model with Transfer Learning for High-Accuracy Lung Cancer Classification from CT Images

Lung nodules manifest as small tissue formations in the lungs, presenting as circular, white shadows on a computed tomography (CT) scan. These nodules are categorized as either benign or malignant. Benign nodules, being non-cancerous, do not extend to other body parts. Conversely, malignant lung nodules are cancerous that can extend to other body parts. This paper proposes a SMOTE-based stacked ensemble model with transfer learning that classifies lung cancer from Computed Tomography images. Computed Tomography-CT can detect abnormal features precisely. The IQOTH/NCCD benchmark dataset is used, which contains a total of 1,098 images, each of which is a CT scan slice from 110 different cases. It contains 3 classes benign, malignant, and normal. In the present work, first data augmentation and SMOTE are applied for the pre-processing of the dataset and to handle the imbalanced data. Transfer learning is used with fine tuning to train four deep learning architectures Xception + CNN, VGG19 + CNN, EfficientNetB7 + CNN, and Inception v3 + CNN. A comprehensive evaluation of these architectures is done and then these four trained models are stacked to generate the Ensemble model for lung cancer classification (SEMLCC). From the experimental results, it was found that the SEMLCC model outperforms all the existing methods. The SEMLCC model achieved an impressive 96.90% accuracy, 97.04% precision, 98.50% specificity, 96.90% sensitivity, and 96.90% F1 Score. This result shows that the SEMLCC performance is excellent in classifying data with high accuracy, high sensitivity(recall), and low FPR-false positive rate.