In this paper, we extend the previous work on the robust pre-processing technique which allows to consider all available information from MRI scans by composition of T1, T1C and FLAIR sequences in the unique input. Such approach enriches the input data for the automatic segmentation process and helps to improve the accuracy of the segmentation performance. Proposed method also demonstrate significant improvement on the multi-class segmentation problem with respect to Dice metrics compare to similar training / evaluation procedure based on any single sequence regardless of the chosen neural network architecture. Obtained results demonstrate significant evaluation improvement while combining three MRI sequences in the 3-channel RGB like image for considered problem of multi-class brain tumor segmentation. We also provide results of comparison of various gradient descent optimization methods and of different backbone architectures. We found that different algorithms worked best for different tumors, but no single algorithm ranked in the top for all types of tumors simultaneously. Final improvements on the test part of our dataset are in the range of 6 - 9% on the trained model according to the Dice metric with the best value of 0.949.