融合多尺度注意力和累积学习的白血病分类识别

Leukemia classification and recognition based on multi⁃scale attention and cumulative learning

急性淋巴细胞白血病(ALL)图像数据集中有着类间形态学相似、数据不平衡的问题。文中设计了一种包含多尺度空间注意力和通道注意力的卷积模块,可以更好地提取不同类别图像的细颗粒特征信息,用于分类器的预测分类。使用加权交叉熵损失函数惩罚样本数量多的类,让模型学习不会偏向多数类。在此基础上引入累积学习策略,随着训练进程动态地调整正常损失函数和加权损失函数的比重,避免了加权损失函数对表征学习的损害,保持了对分类器的促进效果。最终在开源白血病细胞图像数据集C-NMC验证该设计方法的可行性,实验结果表明,测试集F1分数达到96.2%,对白血病细胞图像有着良好的识别效果。

The image data set of acute lymphoblastic leukemia(ALL)has the deficiencies of morphological similarity between classes and data imbalance.Therefore,a convolution module that includes multi⁃scale spatial attention and channel attention is designed,which can better extract the fine⁃grained feature information of different types of images for the prediction and classification of classifiers.The weighted cross⁃entropy loss function is used to punish the classes with large number of samples,so that the model learning will not be biased towards the majority of classes.On this basis,the cumulative learning strategy is introduced,and the proportion of normal loss function and weighted loss function is adjusted dynamically along with the training process,so as to avoid the damage of weighted loss function to representation learning and maintain the promotion effect on classifier.Finally,the feasibility of the designed method was verified in the open source leukemia cell image data set C⁃NMC,and the F1 score of the test set reached 96.2%.Therefore,the proposed method has good recognition effect on leukemia cell image.