Classification of Lung Tumors by using Deep Learning

Sammanfattning: The aim of this thesis is to apply deep learning on medical images in order to build an image recognition algorithm. The medical images used for this purpose are CT scans of lung tissue, which is a three dimensional image of a patients lungs. The purpose is to design an image recognition algorithm that is able to differentiate between tumors and normal tissue in lungs. The algorithm is based on artificial neural networks and therefore the ability of a convolutional neural network (CNN) to predict a tumor is studied. Two different architectures are designed in this thesis, which are a three and six layer CNN. In addition, different hyper-parameters and optimizers are compared in order to find suitable settings. This thesis concludes that no significant difference exists between the results of the two architectures. The architecture with three layers is faster to train and therefore 100 trainings with same settings are completed with this architecture in order to get statistics of the trainings. The mean accuracy for the test set is 91:1% and the standard-deviation for the test set is 2:39%. The mean sensitivity is 89:7% and the mean specificity is 92:4%.