Comparison of artificial intelligence methods in diagnosis of lung cancer

Lung cancer is one of the deadliest types of cancer. Approximately 2 million new cases were detected in 2020. About 20% of cancer patients have lung cancer. There are two subspecies: small cell and non-small cell. Non-small cell carcinomas are divided into 3 subgroups: adenocarcinoma, squamous cell carcinoma, and large cell carcinoma. Tumor structures of subspecies differ from each other. Therefore, it can be differentiated in computed tomography images. In this study, a diagnosis of lung cancer is made through computed tomography images. At the same time, a subtype of lung cancer is detected. After the classification process was completed, the best classification algorithm was determined. In this study, Naive Bayes, k nearest neighbor, support vector machine, gradient descent with momentum, Levenberg-Marquardt, scaled conjugate gradient, convolutional neural networks, VGG16 and ResNet101 models are compared. The images obtained from the Kaggle Chest CT-Scan dataset were standardized by reducing them to 80x40 size. Images are converted to gray level so that the images can undergo wavelet transform. Feature extraction was performed by applying the 3rd level wavelet transform to the obtained images. In order to ensure the randomness of the data, the data were separated by the 5-fold cross-validation method. Since machine learning and artificial neural networks receive one-dimensional input, the images are reduced to one dimension and given as input to the models. Since deep learning models can automatically extract features in image processing, the data are given directly to the models. The most unsuccessful model is Naive Bayes with an F1 score of %51,28 and a kappa score of 0,335. The most unsuccessful model after Naive Bayes is the artificial neural network optimized by the gradient descent with momentum algorithm with an F1 score of 74,22% and a kappa score of 0,646. The two most successful models are the ResNet101 and convolutional neural network, which have an F1 score of 98,78% and a kappa score of 0,983.