Lung cancer classification with convolutional neural networks

Lung cancer, one of the leading causes of cancer-related deaths, remains an important public health problem because it is diagnosed at a late stage and early symptoms are not prominent. Early and accurate diagnosis is of great importance to reduce preventable deaths. In this study, we investigate advanced techniques for lung cancer diagnosis using Convolutional Neural Networks (CNNs) and transfer learning methods. At the beginning of the study, binary classification was performed by combining the images of different lung cancer patients in the dataset with the computed tomography images of normal patients. At this stage, experiments were conducted using different architectures and hyperparameter optimizations to determine the best model. As a result of the experiments, the best performing model among the architectures achieved 93% accuracy. In the next stage, the data set was initialised and multiclassed and the performances of various optimisation algorithms were compared. As optimization algorithms; Adam, SGD, RMSprop, Adadelta, Adagrad, Adamax and Nadam algorithms were used and an accuracy rate of 87% was obtained with the Adam algorithm. The same study was repeated using more lung images by increasing the data in the dataset and an accuracy of 96% was achieved with the RMSprop optimization algorithm. Furthermore, state-of-the-art models were trained using transfer learning on the existing dataset and test accuracies were obtained. The proposed method achieved higher accuracy values than both transfer learning based approaches and similar studies in the literature. The developed model will contribute to the decision-making process of doctors by providing successful results.