<p>Diagnosing brain tumors is a complex and time-consuming task that heavily relies on radiologists’ expertise. Deep learning models have revolutionized this field by enabling more accurate and efficient assessments, with attention-based architectures emerging as promising tools for highlighting critical features in complex medical images. In this study, we propose AT-YOLOv10, a modified version of YOLOv10, specifically designed to enhance both efficiency and robustness. Key improvements include the replacement of the Partial Self-Attention (PSA) mechanism with the Normalization-based Attention Module (NAM) to strengthen focus on relevant regions, and the adoption of MobileNetV3 as the backbone to reduce model complexity while preserving feature extraction capabilities. A Gaussian filter was employed to smooth MRI scans, and data augmentation techniques were applied to promote generalization. The proposed model was trained and evaluated using brain MRI images from the Gazi Brains 2020, BRATS 2019, and Br35H datasets. These datasets were combined to enable classification into twelve clinically significant categories: glioblastoma, meningioma, astrocytoma, metastases, central nervous system lymphoma, pituitary adenoma, ependymomas, medulloblastomas, oligodendrogliomas, hemangioblastomas, gliomas, and healthy brain tissues. This strategy ensured broad coverage of the morphological heterogeneity of brain tumors encountered in clinical practice. Across the experiments, AT-YOLOv10 achieved consistent superior performance. Specifically, on the BRATS 2019 dataset, it reached an accuracy of 98.73%, precision of 98.27%, recall of 98.33%, and an F1-score of 98.71%. On the Gazi Brains 2020 dataset, it obtained an accuracy of 98.51%, precision of 97.98%, recall of 98.11%, and an F1-score of 98.05%. On the Br35H dataset, AT-YOLOv10 achieved an outstanding accuracy of 99.93%, precision of 99.88%, recall of 99.45%, F1-score of 99.69%, MCR of 0.09%, specificity of 99.81%, PPV of 99.85%, NPV of 99.87%, and IoU of 98.56%. These results demonstrate a remarkable balance between sensitivity, specificity, and spatial localization accuracy. Furthermore, the integration of Grad-CAM enhanced the interpretability of the model by visualizing the regions of attention during classification decisions. Despite the promising results, some limitations remain, including dependency on high-quality input images and challenges in early-stage tumor detection and domain generalization. Future directions will focus on exploring domain adaptation techniques, semi-supervised learning strategies, model compression for resource-constrained environments, and the incorporation of advanced interpretability frameworks to further strengthen the clinical applicability of AT-YOLOv10.</p>