BrainYOLO-MCA: An Improved YOLOv11 with Multi-Scale Channel Attention for Brain Tumor Detection.

INTRODUCTION

Brain tumor detection in magnetic resonance imaging is crucial for early diagnosis and clinical treatment planning; however, it remains a challenging task due to the difficulty in detecting tumors with blurred boundaries, high visual similarity to normal brain tissue, and tiny early-stage lesions. To address these limitations, this study proposes a novel brain tumor detection framework named BrainYOLO-MCA.

METHODS

BrainYOLO-MCA is an improved YOLOv11 architecture with a core design of a novel Multi-scale Channel Attention (MCA) mechanism to enhance multi-scale feature representation and fuse critical texture and background information, thereby improving the recognition of indistinct tumor boundaries. On this basis, three targeted optimizations were integrated: a dynamic upsampling strategy (DySample) to preserve the spatial details of tiny lesions, an additional small-object detection head to enhance sensitivity to early-stage small tumors, and a sparse self-attention mechanism to strengthen global contextual feature representation. The proposed model is comprehensively validated on three public brain tumor datasets (Br35H, MBrT, Figshare).

RESULTS

Experimental results show that the study's method achieves 98.6% precision and 97.5% mAP50 on Br35H, 95.3% precision and 97.5% mAP50 on MBrT, and 96.1% precision and 96.3% mAP50 on Figshare, outperforming mainstream YOLO-based models including the baseline YOLOv11, while maintaining a lightweight architecture with only 3.1M parameters.

DISCUSSION: These results demonstrate that the proposed model improves the detection of tumors with blurred boundaries and tiny lesions, while achieving consistent performance across multiple datasets. The Multi-scale Channel Attention and feature fusion strategies enhance feature representation and improve robustness and detection accuracy.

CONCLUSION

The proposed BrainYOLO-MCA enables more accurate and reliable brain tumor detection, which can support clinical decision-making and facilitate early diagnosis of brain tumors.