Prediction of tumor data from lung CT images with different deep learning models

Lung tumors are common today and are a dangerous disease that commonly causes death in people. Accurate and early diagnosis of lung tumors is very important in the prevention and treatment of lung cancer. In addition, with early and accurate diagnosis, the chances of survival of patients with cancer increase. This type of cancer is detected by medical imaging methods such as magnetic resonance (MR), positron emission tomography (PET) and computed tomography (CT), and doctors manually follow up and treat patients with this disease. However, manual examinations can often lead to incorrect diagnosis, thus putting people's health at risk. Instead, computer-assisted, automatic, accurate and detailed early cancer diagnosis is needed. Such an application can also make significant contributions to doctors working in this field of health. Instead of classical methods, computer-aided applications based on deep learning are of great benefit to experts working in this field of lung tumors. For this reason, in this study, studies on lung diseases were examined in detail. In the first stage of the study, 1190 lung tomography images were prepared and preprocessed for the proposed deep learning models. In the second stage, the detection of different lung tumors and normal lung images was performed using deep learning models such as Convolutional Neural Network (CNN), Deep Neural Network (DNN), ResNet, Xception, AlexNet, VGG16 and MobileNet. The deep learning methods used were modeled with different network structures and different tumors were classified quickly and accurately with this network structure. The layers, parameter values, classification accuracy and loss graphs used for the proposed DNN, CNN, ResNet, Xception, AlexNet VGG16 and MobileNet models are shown separately. The accuracy of each model used was calculated with different evaluation metrics such as sensitivity, precision and F1-Score, and the results were compared. In addition, performance analyzes were performed for each model, and complexity matrices and ROC analyzes were presented for training, testing and valid images. The test successes of the proposed deep learning models DNN, CNN, ResNet, Xception, AlexNet, VGG16 and MobileNet were found to be 98.18%, 98.18%, 95.15%, 89.09%, 82.42%, 92.12% and 96.36%, respectively.