Accurate detection of brain tumors is critical for early diagnosis and treatment planning, yet it remains challenging due to the irregular shapes, blurred boundaries, and low contrast of lesions, particularly in Gliomas. To address the limitations of existing methods in balancing accuracy with computational efficiency, this paper proposes a lightweight brain tumor detection framework based on an improved YOLOv12n. We introduce three core innovations to enhance feature representation: (1) the A2C2f-CGLU-DYT module, which integrates dynamic activation functions to strengthen local nonlinear modeling and adaptability to input variations; (2) the C2TSSA module, which utilizes token statistics self-attention to efficiently capture global long-range dependencies without the heavy computational cost of traditional transformers; and (3) the CGAFusion module, which employs content-guided attention to effectively fuse shallow geometric details with deep semantic features. Experimental results on the Brain Tumor dataset demonstrate that the proposed method achieves a precision of 92.7%, a recall of 87.4%, and an mAP@0.5 of 93.8%, outperforming the baseline YOLOv12n by 0.7%, 0.9%, and 1.2%, respectively. Notably, in the challenging Glioma category, the model attains an accuracy of 84.6%, significantly surpassing YOLOv10n by a margin of 8.6%. Furthermore, validation on the Blood Cell Count dataset yields an mAP@0.5 of 94.1%, confirming the model’s robust cross-dataset generalization ability.
Liu et al. (Sun,) studied this question.