<p>Gliomas, a highly aggressive and malignant category of brain tumors, continue to pose a significant global health challenge. Originating from the abnormal and uncontrolled growth of glial cells within the brain, these tumors often result in severe neurological impairments and high mortality rates, underscoring the urgency for improved diagnostic and therapeutic strategies. Early diagnosis can significantly improve outcomes and survival. Consequently, precise segmentation of tumor tissue in medical images is critical for accurate brain tumor diagnosis and effective treatment. However, achieving high-precision segmentation remains challenging due to the complex and variable nature of tumor structures in medical images. To address this problem, we developed R2A-UNET, a U-shaped architecture that leverages the power of residual blocks and attention mechanisms. To enhance the model’s ability to capture critical and relevant information, we incorporated two advanced attention mechanisms. These mechanisms are designed to prioritize important features while suppressing irrelevant or redundant details, thereby significantly improving the efficiency and accuracy of feature extraction across varying datasets. Normalized Channel Attention (NCA) was integrated in each encoder stage, generating a squeezed vector with relevant features at the end of the contracting path. Normalized Spatial Attention (NSA) was included in the skip connection in the middle of the encoder and decoder, generating more concentrated feature maps before concatenation on the decoder side. These mechanisms enable the model to focus on specific pixel values that more accurately localize abnormalities. In our study, we evaluated the performance of our model using two MRI image datasets: the LGG (Lower-Grade Glioma) segmentation database and the BraTS 2018 dataset. Our method achieved a DSC of 92% and an IoU of 86% on the LGG dataset. On the BraTS dataset, it achieved a DSC of 94.79% and an IoU of 90.12%, demonstrating consistent and accurate segmentation performance. To further evaluate our model’s generalizability, we conducted cross-dataset validation, and the results demonstrated good performance. When trained on the LGG dataset and tested on BraTS, the model achieved a DSC of 91.12%. Conversely, when trained on BraTS and tested on LGG, it achieved a DSC of 88.39%. These results highlight the model’s ability to adapt across different datasets with varying characteristics. We analyzed the predictions using Grad-CAM to visualize the decision-making process and applied the Wilcoxon rank-sum test for statistical comparison. These evaluations confirmed the model’s effectiveness in segmenting unseen medical images, supporting its applicability for clinical use.</p>