An improved YOLOv8-based model named YOLO-RMR is proposed for brain tumor detection in MRI images. The model integrates an enhanced re-parameterizable vision transformer (RepViT) for feature extraction, a multi-Scale fusion Block (MS-Block) for multi-scale fusion, and a receptive field block (RFB) for contextual modeling. On the Roboflow dataset, YOLO-RMR improves mAP50 and mAP50:95 by 7.9% and 5.4%, respectively, while accuracy and recall increase by 2.5% and 7.7%. On the Br35H dataset, mAP50 and mAP50:95 increase by 7.5% and 5.5%, and accuracy and recall improve by 2.0% and 7.5%, which demonstrates that the YOLO-RMR model has strong ro-bustness and generalization.