Medical images in clinical practice are highly heterogeneous and often differ in quality from those used in academic research. Preprocessing methods frequently fail in extreme cases where anatomical variations, artifacts, or imaging protocols deviate from standard conditions. Thus, robust segmentation methods are crucial. We propose a deep learning approach for fast and accurate segmentation of the human brain into 132 regions. The model is based on an efficient U-Net-like architecture that leverages intersection points across orthogonal 2D planes and hierarchical relationships for label fusion during end-to-end training. Weakly supervised learning enables whole-brain segmentation and intracranial volume (ICV) estimation using partially labeled data. Additionally, data augmentation generates realistic MRI variations, enhancing model robustness while preserving data privacy. Unlike conventional methods, the proposed approach does not require preprocessing steps such as skull stripping or artifact removal and remains robust across diverse imaging conditions. Extensive experiments with different atlases demonstrate superior segmentation accuracy and generalization compared to state-of-the-art methods across intra- and inter-domain datasets.