<p>Brain tumor segmentation entails detecting and outlining abnormal tissue regions in brain images, a task essential for diagnosis, treatment planning, and monitoring. While machine learning and deep learning techniques have advanced automated MRI segmentation, enabling classification by tumor type and grade, and tracking pathological changes. This task remains challenging due to the brain’s structural complexity, inter-patient variability, and imaging artifacts. Addressing these issues requires sophisticated image processing methods and deep learning models specifically designed for brain image analysis. This paper proposes an explainable deep learning framework for automatic brain tumor detection, localization, and sub-region segmentation. The proposed model adopts a hybrid architecture that integrates U-Net with Dense blocks as its backbone, aiming to improve feature reuse, gradient flow, and overall segmentation accuracy. This approach is motivated by the limitations of standard segmentation pipelines, which often rely on encoders originally designed for classification tasks and generic preprocessing steps. The innovation of our method lies in its task-specific design that combines dense connectivity, adaptive loss weighting, and spatially focused preprocessing, resulting in a more robust and precise segmentation pipeline tailored to the complexity of brain tumor morphology. The evaluation of pretrained backbones and a loss function ablation study confirmed that the proposed Dense-UNet architecture offers superior feature extraction, convergence stability, and segmentation performance across both binary and multiclass brain tumor tasks, using two datasets and multiple evaluation metrics. The models performances have been evaluated on binary and multiclass brain tumor segmentation tasks, using two different datasets and several evaluation metrics. In multiclass segmentation, the model achieved a Dice score of 0.94 for enhancing tumor (ET), 0.84 for tumor core (TC), and 0.91 for whole tumor (WT), with corresponding mean IOU score of 0.81. For binary segmentation, the proposed model reached a Dice score of 0.95 and an IOU of 0.85 for the whole tumor region. The best recorded performance also includes a Precision of 98% and 99% for Accuracy and Precision scores, with a minimum validation loss value of 0.10, confirming the effectiveness and stability of the proposed pipeline. The model’s visual explainability, achieved through Grad-CAM and backpropagation, enhances transparency and supports expert interpretation, demonstrating the effectiveness of the proposed XAI approach in building trust and understanding model behavior.</p>