ABSTRACT
The paper explores the ensemble approach for the segmentation of brain tumors from Magnetic Resonance Images. For an automated brain tumor segmentation, well-known U-net architecture has been modified for the number of layers and input image type. In this modified architecture, the total number of layers is reduced to three, which have a reduced number of parameters required for tumor segmentation. Moreover, an ensemble of three architectures is used to segment the tumor. From three different views i.e., axial, coronal and sagittal, segmentation result is combined considering the mean of the output of these networks. The highest dice similarity coefficient achieved by this approach is 0.78.
