YOLO-DCS: A Brain Tumor Detection Algorithm Based on Dynamic Self-Attention Mechanism and Adaptive Fusion∗

To address the issues of high missed detection rates for small objects and interference from complex backgrounds in brain tumor detection, this paper proposes an improved model named YOLO-DCS based on YOLOv8. First, a C2fDAttention module is designed to replace the original backbone network, enhancing feature extraction in tumor regions through a dynamic attention mechanism. Second, the CARAFE (ContentAware ReAssembly of FEatures) upsampling operator is adopted to improve contour reconstruction accuracy for small tumors. Finally, the SEAMHead (Squeeze-and-Excitation Augmented Mask Head) is introduced into the detection head to achieve adaptive fusion of multi-scale features. Experimental results on the dataset demonstrate that the improved model achieves an mAP50 of 97.3 % with only a 0.06 % increase in parameter count, while the recall rate is improved by 3.3 % compared to the original YOLOv8, and the floating-point operations per second are reduced by 8.6 %. This provides a high-precision and lightweight solution for medical image-assisted diagnosis.