<p>Segmentation of brain tumors in cluttered regions of magnetic resonance imaging (MRI) is essential for precise diagnosis and therapy planning. Despite several automated methods, challenges persist due to variations in tumor location, size, shape, intensity, and the overlap between normal and malignant tissues in cluttered regions. Moreover, the loss of spatial and contextual information in cluttered areas exacerbates segmentation challenges. This study proposes OBL-RA Fusion Net, which integrates oblique learning with residual connected attention and multi-scale context fusion. The oblique learning enhances feature discrimination in cluttered and uncluttered tumor regions by computing regional statistical cues and forming an oblique projection that separates overlapping tumor features. A hybrid module integrates residual connections and attention mechanisms, where the residual part facilitates the effective training of deeper networks, while the attention mechanism adapts to complex tumor structures and captures heterogeneity within tumor regions. Additionally, a multi-scale context fusion module enhances the recovery of spatial and contextual information across different scales. The proposed model was evaluated on FLAIR LGG and BraTS2020 datasets and achieved a dice similarity coefficient (DSC) of 0.9151 and an intersection over union (IOU) of 0.8435 on FLAIR LGG. On the BraTS2020, the model attained DSC of 0.9727 and IOU of 0.9354. The integration of an explainable AI (XAI) method facilitates visual interpretations of tumor locations identified through oblique learning, demonstrating its transparency and reliability. A comparative analysis with state-of-the-art methods demonstrates improved segmentation in cluttered regions and generalization for complex tumor structures. XAI-based segmentation improves diagnostic and therapeutic decision-making.</p>