Abstract: Accurate brain tumor detection is crucial for clinical diagnosis and treatment planning. However, existing deep learning models for brain tumor MRI scans often faced challenges such as excessive reliance on anchor boxes, limited receptive fields, and insufficient feature fusion capabilities. To address these issues, an efficient brain tumor detection algorithm based on an improved DEIM was proposed. First, a Multi-Scale Hybrid Gated Block (MS-HG Block) was designed within the backbone network. By integrating a gating mechanism and a SimAM 3D attention module, the adaptive capture of key spatial details of tumors was enhanced. Second, a Multi-Scale Contextual C2f Block (MS-C2f Block) was constructed in the feature fusion neck network. Multi-scale parallel dilated convolutions were utilized to expand the receptive field and improve contextual perception for tumors of different sizes. Finally, an Efficient Channel-Shuffle Upsampling Block (ESCUB) was proposed. Special channel shift and shuffle operations were designed to reduce information loss during upsampling, thereby generating higher-quality feature maps. Experimental results on the Brain Tumor MRI dataset demonstrate that, compared to the baseline model DEIM-D-FINE-N, the proposed algorithm achieves improvements of 2.6 and 4.6 percentage points in AP@[0.5:0.95] and AP@0.75, respectively, while the parameter count and computational cost (GFLOPS) increase by only 0.1M and 0.3 GFLOPs. Further comparisons with mainstream detectors such as RT-DETR-R50, YOLOv11n, and YOLOv12n show that the proposed algorithm significantly outperforms these methods in both AP@[0.5:0.95] (68.6% vs. 67.5%, 67.0%, 66.4%) and AP@0.75 (76.4% vs. 74.1%, 75.6%, 75.4%). This algorithm achieves a balance between accuracy and efficiency in brain tumor detection tasks, providing a reliable and efficient technical solution for auxiliary medical diagnosis.

摘要： 脑肿瘤精准检测对于临床诊断与治疗规划至关重要，然而现有深度学习模型在处理脑肿瘤磁共振成像(MRI)影像时常面临过度依赖锚框、感受野有限以及特征融合能力不足等挑战。为解决上述问题，文中提出一种基于改进DEIM的脑肿瘤高效检测算法。首先，在主干网络中设计多尺度混合门控模块(MS-HG Block)，通过融合门控机制和SimAM三维注意力模块，增强对肿瘤关键空间细节的自适应捕获能力；其次，在特征融合颈部网络中构建多尺度上下文C2f模块(MS-C2f Block)，利用多尺度并行空洞卷积扩大感受野，提升对不同尺度肿瘤的上下文感知能力；最后，提出高效通道移位上采样模块(ESCUB)，通过设计特殊的通道移位与混洗操作减少上采样过程中的信息损失，生成更高质量的特征图。实验结果表明，在Brain Tumor MRI数据集上，相较于基准模型DEIM-D-FINE-N，所提算法在参数量和浮点运算量仅增加0.1M和0.3GFLOPs的情况下，AP@[0.5:0.95]和AP@0.75指标分别提升了2.6和4.6个百分点。进一步对比RT-DETR-R50、YOLOv11n和YOLOv12n等主流检测算法，所提算法在AP@[0.5:0.95](68.6% vs. 67.5%、67.0%、66.4%)和AP@0.75(76.4% vs. 74.1%、75.6%、75.4%)指标上均显著优于对比方法。该算法在脑肿瘤检测任务中实现了精度与效率的平衡，为辅助医疗诊断提供了可靠高效的技术方案。