Brain tumor segmentation is an important and time-consuming part of the usual clinical diagnosis process. Multi-class segmentation of different tumor types is a challenging task, due to the differences in shape, size, location and scanner parameters. Many 2D and 3D convolution neural network architectures have been proposed to address this problem achieving a significant success. It is well known that 2D approach is generally faster and more popular in the most of such problems. However, the usage of 3D models allows us to simultaneously improve the quality of segmentation. Accounting the context along the sagittal plane leads to the learning of 3-dimensional features that we used for computationally expensive 3D operations what in its turn increases the learning time as well as decreases the speed of operation.In this paper, we compare the 2D and 3D approaches on 2 datasets with MRI images: the one from the BraTS 2020 competition and a private Siberian Brain tumor dataset. In each dataset, any single scan is represented by 4 sequences T1, T1C, T2 and T2-Flair, annotated by two certified neuro-radiologist specialists. The datasets differ from each other in the dimension, grade set and tumor type. Numerical comparison was performed based on the Dice score index. We provide the case by case analysis for the samples that caused most difficulties for the models. The results obtained in our work demonstrate the significant over performing of 3D methods keeping robustness in a regard of data source and type that allow us to get a little closer to AI-assisted diagnosis.