چكيده به لاتين
Diagnosis of brain tumors is one of the challenging tasks in the field of medical imaging. Due to its nature, the possibility of the emergence of a tumor in the brain with different size and shape makes diagnosis more complicated. Doing it manually may result in human errors. Furthermore it is time-consuming and expensive. These problems indicate that we need an automatic method for this task. In this thesis, deep learning methods based on convolutional neural networks are used to solve these problems. The proposed architecture is an improved version of U-Net. In this architecture, we have tried to modify the original U-Net, to improve the overall performance of the network. We have also used a learning procedure called Multi-View for better learning which, according to the evaluation results, has improved the performance. This kind of training is intended to reduce the problems of two-dimensional training. Also, due to the imbalance of the brain tumor segmentation classes, a combination of the Cross-Entropy Function and the Generalized Dice Function is used as a cost function. The proposed network has shown a great performance in Brats 2018 challenge. The mean of the Dice Score obtained from this method is 89.46, 82.31, and 81.30 for whole tumor, core tumor and enhancing tumor respectively, using the 2018 validation data set.
Keywords: Deep Learning, Brain Tumor, MRI, Segmentation, Deep Convolutional Neural Networks
