融合注意力的主动迭代优化白细胞图像分类模型

Active Iterative Optimization of Leukocyte Image Classification Model with Fused Attention

白细胞的数量和结构特征蕴含了人类健康状况的重要信息。通过对不同种类的白细胞进行计数,可为多种疾病的早期治疗提供重要依据。但目前采集与标注白细胞数据集成本较高、现有的白细胞数据集数量较少,给计算机辅助白细胞自动分类带来了挑战。为此,本研究提出了一种融合注意力的主动迭代优化白细胞图像分类模型,通过在ResNet18主干网络上附加用于主动学习的LossNet网络,从大量未标注样本中挑选最具代表性的样本进行标注,减少了需要人工标注的样本量。同时,为了应对白细胞数据集类间不平衡对主动学习的影响,加入主动迭代扩增模块,挑选训练过程中的困难样本进行含有随机因子的数据扩增,自下而上形成了双向的信息交互,以增强模型对不平衡数据集的适应力。最终,在比较了3种注意力模块后,本研究选择加入CBAM注意力模块,以增强模型对白细胞特征区域的关注、提高模型的性能。采用包含14514张白细胞显微镜图像的Raabin-WBC数据集进行方法验证。实验结果表明,所提出的模型在使用训练集28%、37%、52%的样本时,分类准确率分别达到92.35%、93.64%、94.86%,相比原ResNet18分别提升了5.14%、9.24%、2.37%,模型有效降低了白细胞数据集的标注成本,在缺乏标注的医学数据集上具有较为广泛的应用前景。

The number and structural characteristics of leukocytes contain important information about human health,and the counting of different types of leukocytes can provide important basis for the early treatment of many diseases.However,the high cost of collecting and labeling leukocyte data integration and the small number of available leukocyte datasets currently pose challenges for automatic computer-aided leukocyte classification.To address these challenges,an active iterative optimization leukocyte image classification model incorporating attention was proposed in this paper.By attaching a LossNet network for active learning to the ResNet18 backbone network,the most representative samples were selected from a large number of unlabeled samples for labeling,reducing the amount of samples that need to be manually labeled.Meanwhile,to cope with the impact of inter-class imbalance in the leukocyte dataset on active learning,this paper added an active iterative augmentation module to select difficult samples in the training process for data augmentation containing random factors,which formed a two-way information interaction from the bottom up and enhanced the adaptability of the model to imbalanced datasets.Finally,after comparing three attention modules,this paper chose to incorporate the CBAM attention module to enhance the model′s focus on the leukocyte feature regions and improve the model′s performance.In this study,the Raabin-WBC dataset containing 14514 leukocyte microscopy images was used for method validation,and the experimental results showed that the classification accuracy of the model proposed in this paper reached 92.35%,93.64%,and 94.86%when using 28%,37%,and 52%samples of the training set,respectively,which was 5.14%,9.24%,and 2.37%higher than the original ResNet18,respectively,and the model greatly reduced the labeling cost of leukocyte dataset,showing wide application prospectives in medical datasets that was lack of labelling.