Accurate classification of brain tumor subtypes is critical for effective clinical decision-making and patient recovery. In this study, for the classification and detection of brain tumors using MRI scans we present a high-precision deep learning framework: RT-DETR+DL model (Real Time Detection Transformer with Dice Loss). The dataset comprises publicly available images from the “MRI for Brain Tumor” dataset on Kaggle and the “Glioma_of_test” dataset from Roboflow. Comprehensive preprocessing—including normalization, data augmentation, and dataset partitioning—was applied to improve model robustness. We evaluated six state-of-the-art deep learning models: ResNet18, ResNet101, ResNet152, Faster R-CNN, YOLO v12, and RT-DETR. To mitigate overfitting and optimize computational performance, dropout layers were introduced and batch sizes were tuned. Among all models, RT-DETR demonstrated superior performance, achieving an average classification precision of 96.6% and a mean Average Precision at IoU threshold 0.5 (mAP50) of 96.3%. These results outperform conventional models, underscoring the efficacy of the RT-DETR architecture in leveraging cross-attention mechanisms to bypass the redundancy and hyperparameter sensitivity associated with traditional NMS-based methods such as Faster R-CNN and YOLO. Further enhancements were realized by integrating the Dice Loss function, yielding a 0.6% improvement in average recall and a 4.8% increase in frames per second (FPS) over the baseline RT-DETR. These findings highlight the clinical viability of the presented model in delivering accurate and efficient brain tumor classification. Subsequent ablation studies identified the optimal Dice Loss weight as λ = 0.8. With this configuration, 5-fold cross-validation confirmed that the RT-DETR+DL model demonstrated superior generalization and stability compared to YOLOv12.