Lung Cancer Diagnosis Based on Convolutional Neural Networks Ensemble Model

Abstract

Lung cancer is a lethal disease that can be treated efficiently if diagnosed in an early stage. Screening is a technology involving using CT scan to diagnose whether the lung is attacked by malignant tumors. This study proposes a CNN-based framework to help classify if the CT scan detects a cancer or not. In the analysis, several individual CNN models, including AlexNet, VGG, DCNN and DenseNet, are applied to make predictions and their performances are compared. Subsequently, selected individual models are ensembled by voting and stacking strategy that synthesize their predicting results. According to the results, the best individual model is DenseNet with average pooling layers, which gains a 97.48% accuracy and a 0.99019 AUC score. In comparison, the best ensemble model turns out to be assembling predicting results of best three individual models by stacked generalization, which reaches a 99.37% accuracy and a 0.99984 AUC score. These results show that it is useful to apply ensemble algorithm to improving the performance above individual models in this lung cancer diagnosis framework. Moreover, the final ensemble structure is efficient and reliable on figuring out lung scan images with malignant tumors.