Glioma is a type of brain tumor that is the most typical and most aggressive tumor. Magnetic resonance imaging (MRI) has a widespread utilization as an imaging method for assessing the tumor; however, a lot of information obtained from MRI would prevent manual segmentation in an acceptable timeframe. Therefore, we need reliable and automatic segmentation techniques. Nonetheless, broad structural and spatial variability amongst the brain tumors made automatic segmentation a difficult issue. Capsule Network (CapsNet) is an improved convolutional neural network, which involves responding to challenges. Consequently, CapsNet has been considered to be a perfect candidate to perform brain tumor segmentation. In this paper, a new model containing CapsNet, that operates in the Del vector domain and works with inhomogeneous voxels, is presented. Flair and T1 MR images are first transformed from the time domain to the vector domain using the Del operator. MVGC (Mean Value Guided Contour) algorithm is then applied on the Flair image to segment ROI (Region of Interest). Inhomogeneous voxels are then extracted from ROI of Flair and T1 MR images. In the final phase, a new two-path CapsNet architecture is applied to classify voxels. The introduced technique is verified according to BRATS 2015 and BRATS 2013 datasets. The outputs show that this technique achieves a competitive result with the average Dice similarity equal to 0.90 for the whole tumor in Brats2013 and 0.88 for the entire tumor in Brats2015. In addition, this technique took just ~86.2 s for the segmentation of a patient case.
