A Hybrid LECNN Architecture: A Computer-Assisted Early Diagnosis System for Lung Cancer Using CT Images

Lung cancer is one of the most common causes of cancer-related death. Therefore, early diagnosis of this cancer is crucial for planning patient treatment. This paper proposes a hybrid Lung Ensemble Convolutional Neural Network (LECNN) architecture for the computer-aided early diagnosis of lung cancer via CT images. The proposed hybrid approach integrates a transfer learning (TL) mechanism with ensemble learning (EL) on the basis of majority voting. Initially, CNN architectures (GoogLeNet, EfficientNet, DarkNet19, and ResNet18) are trained via TL, and the resulting CNN models are used as inputs in EL. The outputs from all the CNN architectures are evaluated via majority voting to identify the top-performing triple CNN combination, which is then utilized in the hybrid approach. The performance of the proposed method was assessed via the widely used IQ-OTH/NCCD dataset. Additionally, the impact of the elastic transformation method, a data augmentation technique, on performance improvement was investigated in the proposed method. The triple combination of the GoogLeNet, EfficientNet, and DarkNet19 CNN architectures, as part of the EL method in the hybrid approach, achieved superior performance on both the raw and augmented datasets according to the obtained performance results. The performance evaluations revealed that the proposed approach achieved more than a 5% improvement with the augmented dataset compared with the raw IQ-OTH/NCCD dataset, resulting in the highest performance. The proposed hybrid approach achieved 99% accuracy, 98.82% sensitivity, 99.48% specificity, 99.06% precision, and 98.94% F1 score on the augmented IQ-OTH/NCCD dataset. When compared with findings from previous studies using the same dataset, the proposed hybrid approach outperformed state-of-the-art methods. In conclusion, it demonstrates significant potential as a robust tool for computer-aided early lung cancer diagnosis systems and may also contribute to the development of future hybrid approaches in this field.