基于改进SE-Net网络与多注意力的脑肿瘤分类方法

Classification Method of Brain Tumor Based on Improved SE-Net Network and Multi-attention

手工筛选肿瘤图像来预测脑肿瘤类别的方法非常耗时,而将深度学习与医学图像相结合的方式,可以在一定程度上帮助医生解决这一问题,因此提出改进的SE-Net网络。首先,将Swish激活函数代替批归一化和特征融合后的ReLU激活函数,使模型更好地学习有效特征;其次,在第一层和第二层卷积层后分别添加ECA和改进的BAM注意力模块,在空间和通道2个方向并发进行特征提取,使目标特征充分被利用;最后,在SE注意力模块中添加全局最大池化,利用双通道池化层提取有效特征,抑制无效特征,提高模型准确率。在Kaggle公开的数据集中进行训练与测试,最终结果表明,该方法在脑肿瘤分类测试集中的准确率、召回率、精确率和F1值分别达到99.47%、99.42%、99.45%和99.43%,充分验证了改进模型的有效性。

An improved SE-Net network is proposed for the reason that the combination of deep learning and medical images can help doctors to a certain extent in solving the time-consuming problem of predicting brain tumor category through manual screening of tumor images.Firstly,the Relu activation function is replaced by the Swith activation function after batch normalization and feature fusion,so that the model can learn effective feacures better.Secondly,ECA and improved BAM attention modules are respectively added to the convolution layer of the first layer and the second layer for extracting features concurrently in space and channel directions so as to fully utilize the target feature.Finally,the global maximum pooling is added to the SE attention module and the effective features are extracted by the double channel pooling layer to suppress the invalid features and improve the accuracy of the model.Training and testing are carried out in the Kaggle public dataset.The final results show that the accuracy rate,recall rate,precision rate and F1 values of the method in the brain tumor classification test set have respectively reached 99.47%,99.42%,99.45%and 99.43%,fully verifying the effectiveness of the improved model.