Abstract
Brain tumors remain among the most fatal diseases of the central nervous system, and early diagnosis is vital for improving patient survival outcomes. However, conventional diagnostic approaches relying on manual interpretation of MRI scans by radiologists often suffer from subjectivity, inefficiency, and a high rate of missed diagnoses. To overcome these limitations, this study introduces a systematically optimized deep learning framework designed to achieve real-time performance while preserving high detection accuracy. Specifically, we propose the A2C2f-Mona module, which enhances local and medium-range feature perception through multi-scale depthwise convolution and residual connections; design the C2PSA-DyT module, which replaces conventional normalization layers with element-wise normalization to enhance training stability and feature distribution consistency; and introduce the CGAFusion module, which integrates high- and low-level features via a channel attention mechanism, effectively improving the detection of tumors with blurred boundaries and small volumes. Experimental evaluations demonstrate that the proposed approach surpasses YOLOv12n across all performance metrics, achieving a precision of 93.8%, recall of 88.0%, and mAP@0.5 of 94.0%. Notably, the recall for pituitary tumors shows the greatest improvement, while the recognition accuracy for gliomas is also significantly enhanced. Overall, the proposed method achieves simultaneous gains in detection accuracy and stability, underscoring its substantial potential for advancing intelligent diagnosis in medical imaging.