Deep Feature Extraction Based Fine-Tuning

Abstract

In this study, an optimal deep feature set is determined for the classification of brain tumor tissue types Glioma, Meningioma and Pituitary. It is a good alternative to use a pre-trained network that is trained with millions of data in areas such as medical imaging where the amount of available data is not sufficient to train a CNN. Another method is to train the last layers of the network known as fine-tuning by configuring the problem at hand. In this study, AlexNet is used. Then, an SVM model is trained using features extracted from pre-trained and fine-tuned networks. Experimental results show that the 3D feature vectors extracted from the fine-tuned network yield superior results in a different test set, compared to the feature vectors extracted directly from the pre-trained one and increase the generalization ability of the model.