Examiner: Irene Gu, Dept of Electrical Engineering
Brain cancer, in particular glioma, is a disease with a very low survival rate. The five-year relative survival rate for patients diagnosed with high grade Glioma is 3.6%. An important procedure for the medical clinician when diagnosing and planning treatment for patients is the segmentation of brain tumors into different classes. Magnetic resonance imaging is the imaging modality mainly used during clinical work with brain tumors. Deep learning methods have shown great potential for brain image segmentation and other biomedical applications.
This thesis studies brain tumor segmentation from MRIs in different modalities by deep learning methods. We study different deep learning architectures in combination with feature fusion. Two different architectures are studied and compared for semantic segmentation of brain tumors using fully convolutional networks. The architectures take inspiration from the well renowned U-Net structure, first used for semantic segmentation of cells. Finally, the size of 3D tumor volume is estimated where segmented 3D tumors can be visualized graphically.
The first architecture takes a multimodal MRI-image as input in a single-stream approach, while the second architecture processes each modality separately in a multi-stream fashion followed by feature fusion. A number of experiments were conducted for testing and optimizing different regularization parameters and hyper-parameters. In the multi-stream approach training is implemented in two ways: one is in an end-to-end manner and another in a sequential manner. The results of the experiments in this thesis suggests that the multi-stream approach trained in an end-to-end manner outperforms the other approaches.
Student project presentation
Landahlsrummet (room 7430), Hörsalsvägen 9, 7th floor
28 January, 2020, 14:30
28 January, 2020, 15:30