Magnetic Resonance Imaging (MRI) plays a pivotal role in the diagnosis and treatment of brain tumors, making its accurate assessment critically important. However, the inherent three-dimensional nature of MRI presents several challenges, leading to the common practice of conducting analyses on two-dimensional projections. While this simplification reduces complexity, it also introduces potential biases. Conversely, the more time-intensive three-dimensional evaluations, such as segmentation, can yield precise estimates of various spatial characteristics, enhancing our understanding of disease progression. Recent research focusing on segmentation tasks has demonstrated that Deep Learning techniques outperform traditional computer vision algorithms, although the problem remains complex. In this paper, we introduce a deep cascaded approach for the automatic segmentation of brain tumors. Our method, akin to contemporary object detection techniques, leverages neural networks and includes modifications to the 3D UNet architecture and augmentation strategies to effectively process multimodal MRI data. Additionally, we present a method to improve segmentation quality by incorporating contextual information from models of the same architecture operating on downscaled datasets. We assess our proposed approach using the BraTS 2018 dataset and provide a discussion of the results.