Yolo-c2ecasa: a lightweight attention-enhanced model for three-class brain tumor localization and classification in...

Brain tumor localization and type classification in MRI are important tasks in computer-aided medical image analysis. However, achieving competitive localization and classification performance with lightweight models remains challenging because reducing model complexity may weaken feature representation and localization ability. In this paper, we propose YOLO-C2ECASA, a lightweight attention-enhanced object detection framework based on YOLOv11 for three-class tumor localization and classification in tumor-positive MRI images. The present study focuses on glioma, meningioma, and pituitary tumors and does not address tumor-versus-no-tumor screening. The proposed C2ECASA module replaces the original self-attention design with a parameter-efficient spatial-channel attention mechanism. It integrates improved channel attention, lightweight spatial attention, C2-based feature preservation, and adaptive fusion to enhance lesion-relevant feature representation with limited computational cost. The proposed model was compared with representative object detection models. In the single-run comparison with representative object detectors, YOLO-C2ECASA maintained competitive detection performance while requiring only 2.4M parameters, 6.2G FLOPs, a model size of 5.01 MB, and a GPU inference time of 1.9 ms per image. More importantly, across ten independent runs under scratch initialization, YOLO-C2ECASA achieved mean mAP@0.5 and mAP@0.5:0.95 values of [Formula: see text] and [Formula: see text]. Under pretrained initialization, the corresponding repeated-run results were [Formula: see text] and [Formula: see text], respectively. Under CPU-only deployment, YOLO-C2ECASA achieves an inference time of 27.4 ms per image, further demonstrating its feasibility for resource-constrained environments. These results indicate that the proposed model achieves competitive detection performance while maintaining a lower computational cost than the compared detectors. Ablation and repeated-run experiments further demonstrate that the complete C2ECASA configuration achieves the highest average performance among the evaluated configurations while maintaining low computational complexity. Overall, the proposed method provides a favorable accuracy-efficiency trade-off for lightweight brain tumor localization and type classification in tumor-positive MRI images.