Efficient Multi-Class Classification of NSCLC Subtypes with Transfer Learning-Enhanced CNNs on Augmented CT Data

The second most cancer related deaths are caused by lung cancer, since this cancer type has highest mortality rate. The early diagnosis of Lung cancer is crucial as there are possibilities to increases the survival rate. Conventional diagnostic methods often rely on human interpretation of medical images which is time consuming and are prone to error. In this research we proposed a Novel hybrid model LungNetB5 for multi-class lung cancer classification and prediction. CNN and Efficient Net are integrated into single architecture which blends traditional CNN feature extraction layers with the powerful feature learning capabilities of EfficientNet. The experiments are carried on “MCLCI” dataset for NSCLC lung cancer patients. To overcome the biases, class imbalance problem was addressed using data augmentation techniques. The developed LungNetB5 model attained 92% accuracy. The predicted results show that EfficientNet class based LungNetB5 outperforms other CNN models in terms of efficiency and accuracy. In addition, it is faster and need very less parameters to train when compared to other CNN models, making it a viable experiment for extensive clinical settings and a promising tool for automated detection of lung cancer from CT images.