Abstract: Brain tumors are one of the most lethal cancers in the world, and their accurate classification is crucial for developing effective diagnostic and treatment plans. However, the heterogeneity and complexity of brain tumors pose challenges to traditional diagnostic methods. A deep learning-based brain tumor classification algorithm using MRI images is proposed. The algorithm achieves precise and efficient classification through the fusion of multi-scale features and a dual-channel attention mechanism. Specifically, a multi-scale feature fusion module is employed to replace traditional convolution operations to extract rich feature information at different scales. Then, a dual-channel attention mechanism is developed to dynamically filter key features and enhance the focus on the target region. Subsequently, transfer learning is utilized to initialize the model that can accelerate the learning process and enhance its generalization ability. Finally, a joint loss function is constructed by combining sparse cross entropy loss and center distance loss to oversee the model training process. Additionally, three public data resources are integrated to build a large-scale and well-distributed mixed dataset, and data augmentation techniques are applied to enrich sample diversity, providing sufficient data support for model training. Experimental results show that the proposed algorithm achieves a classification accuracy of 99.33% on the test dataset, which is a 4.08% improvement over the baseline algorithm. Ablation experiments and comparative experiments further validate the effectiveness and efficiency of the algorithm.

Key words: Brain tumor classification, Multi-scale feature fusion, Dual channel attention, Transfer learning

摘要： 脑肿瘤是全球致死率极高的癌症之一，其准确分类对于制定有效的诊断和治疗方案至关重要。然而，脑肿瘤的异质性和复杂性给传统人工诊断带来了严峻挑战。为此，研究了一种基于深度学习的MRI影像脑肿瘤分类算法。该算法通过融合多尺度特征和双通道注意力机制，实现了精确且高效的分类。具体而言，首先采用多尺度特征融合模块替代传统卷积操作，以提取不同尺度的丰富特征信息；然后设计双重通道注意力机制，动态筛选关键特征，增强对目标区域的关注；继而利用迁移学习技术初始化模型，加速学习过程并提升泛化能力；最后结合稀疏交叉熵损失和中心距离损失构建联合损失函数，以监督模型训练过程。此外，研究整合了三个公开数据资源，构建了一个规模较大且分布均衡的混合数据集，并结合数据增强技术丰富样本多样性，为模型训练提供了充分的数据支持。实验结果表明，所提出的算法在测试数据集上实现了99.33%的分类准确率，较基准算法提升了4.08%,消融实验和对比实验也进一步验证了算法的有效性与高效性。

关键词: 脑肿瘤分类, 多尺度特征融合, 双重通道注意力, 迁移学习