Real-time detection method of brain tumor based on deep learning

Objective In the field of brain tumor detection, fast and high-precision object detection models have important clinical value. However, as the complexity of medical imaging data continues to increase, so does the requirement for model detection accuracy and efficiency. Therefore, this paper aims to improve the existing RCS-YOLO model to further improve its detection performance and meet the requirements of medical image analysis for subtle feature detection.Methods In this paper, an improved RCS-YOLO model (MARCS-YOLO) is proposed, which integrates multiple attention mechanisms (including SE Attention, CBAM and GAM) in the RCS-OSA module and optimizes feature fusion strategies to realize the complementarity of attention mechanisms in different dimensions and significantly improve the detection performance of the model. In addition, this paper introduces a learnable channel shuffle mechanism to flexibly control the information flow between channels, avoid the limitations of fixed shuffle, and improve the robustness of feature expression.Results Experimental results on the Br35H dataset show that the improved model achieves 0.95, 0.955, 0.953 and 0.742 in terms of accuracy, recall, mAP50 and mAP50:95, respectively, which are improved compared with the original RCS-YOLO model. The improvement of these indicators shows that the improved model is able to identify brain tumor boundaries more accurately, especially in detecting small tumors.Conclusions The improved MARCS-YOLO model significantly improves the performance of brain tumor detection by integrating multiple attention mechanisms and optimizing feature fusion strategies, and meets the high requirements of medical image analysis for subtle feature detection. Experimental results show that the proposed model is better than the original model in key indicators such as accuracy, recall and mAP, showing its potential value in clinical application.

研究目标：在脑肿瘤检测领域，快速高精度的目标检测模型具备重要的临床应用价值。然而，随着医学影像数据复杂度持续提升，对模型检测精度与效率的要求也同步提高。为此，本文旨在对现有RCS-YOLO模型进行改进，进一步提升其检测性能，以满足医学影像分析对细微特征检测的需求。
研究方法：本文提出一种改进的RCS-YOLO模型（MARCS-YOLO），该模型在RCS-OSA模块（RCS-OSA）中集成了多种注意力机制，包括SE注意力（SE Attention）、CBAM注意力（CBAM）与GAM注意力（GAM），并优化了特征融合策略，实现了不同维度下注意力机制的互补性，显著提升了模型的检测性能。此外，本文引入了可学习通道洗牌机制，可灵活控制通道间的信息流，规避了固定洗牌的局限性，进而提升了特征表达的鲁棒性。
实验结果：在Br35H数据集（Br35H）上的实验结果表明，改进后的模型在准确率、召回率、mAP50与mAP50:95指标上分别达到0.95、0.955、0.953与0.742，各项指标均优于原始RCS-YOLO模型。上述指标的提升表明，改进后的模型能够更精准地识别脑肿瘤边界，尤其在小肿瘤检测任务中表现优异。
研究结论：本文提出的改进型MARCS-YOLO模型通过集成多种注意力机制并优化特征融合策略，显著提升了脑肿瘤检测性能，满足了医学影像分析对细微特征检测的高要求。实验结果表明，所提模型在准确率、召回率与mAP等核心指标上均优于原始模型，展现出其在临床应用中的潜在价值。