Segmentation of MRI images is an indispensable step in the diagnosis and clinical decisions in therapy of brain tumors. The goal of segmentation is to distinguish the tumor from healthy tissue and to identify different parts of the tumor. It is a tedious task that is in clinical practice still performed manually by experienced neuroradiologists. With the advancement of machine learning, automatic segmentation is also within reach.
Current algorithms require complete, three-dimensional MRI images in high resolution. In clinical practice, however, not all sequences are usually recorded in full 3D resolution, making the use of the algorithms more difficult.
This project investigates how existing algorithms can be modified for practical use. A deep neural network, which has already been proven in practice, serves as the basis. It can be shown that different MRI sequences (T1, T1CE, T2 and FLAIR) include notable redundancies and that algorithms, despite incomplete data, still allow segmentation with relatively high accuracy.