With a good balance between speed and accuracy, YOLO (You Only Look Once) framework has become one of the most effective algorithms for object detection. In the field of medical brain tumor detection, this paper proposes an improved and innovative algorithm, YOLOv8 - BBS, based on the YOLOv8 algorithm. It constructs a two - layer feature fusion framework by combining two attention mechanisms, namely the Bi - Level Routing Attention (BRA) and the Bi - modal Attention mechanism (BiFormer), and skillfully integrates this framework into the backbone network of YOLOv8. This enables the algorithm to comprehensively capture image details and global features. To address the problems of class imbalance in the dataset and the identification of difficult samples, this algorithm introduces the SlideLoss function, which dynamically adjusts the class weights, enhancing the accuracy and robustness of data classification. Experimental results on the brain tumor MRI dataset show that the model proposed in this paper outperforms YOLOv5, YOLOv8, YOLOv9, and YOLOv10 in terms of both speed and accuracy. The F1 - score of YOLOv8 - BBS is 1% higher than that of YOLOv8, with an accuracy rate of 81.8% and a recall rate of 72.7%. These improvements confirm the effectiveness of the model in brain tumor detection and provide new ideas for subsequent object detection applications in medical imaging and clinical disease diagnosis.