Lung Cancer Detection Using Transfer Learning with DenseNet121 for Improved Diagnostic Accuracy

Abstract

The improved survival rates of the patients with lung cancer highly depend on the appropriate and timely early detection of this disease and the proper treatment planning. In this paper, we are proposing a transfer learning based convolutional neural network (CNN) model using the DenseNet121 architecture to classify lung cancer images into Malignant and Normal classes with high accuracy. A respiratory lung cancer dataset was previously made publicly available, and was pre-processed into training, validation, and testing sets. The DenseNet121 model was fine-tuned by freezing the convolutional layers and adding fully connected layers together with the dropout regularization layers to avoid overfitting. Final test result on the full dataset gave me a test accuracy of 97.97% with Malignant precision = 1.00 and Normal precision = 0.95. The classification report, confusion matrix, ROC curve, and training performance curves together indicate the robustness and generalization ability of proposed approach. The achieved results indicate that the model based on DenseNet121 outperforms proposed CNN and transfer learning models as well as some traditional models, which makes the presented model efficient and reliable solution for an automated automatic lung cancer detection.