Zhang Li, Zhang Meng-Qian, Lv Xuerui
School of Computer Science and Technology, Soochow University, Suzhou, 215006, Jiangsu, China.
Provincial Key Laboratory for Computer Information Processing Technology, Soochow University, Suzhou, 215006, Jiangsu, China.
Med Biol Eng Comput. 2022 Nov;60(11):3113-3124. doi: 10.1007/s11517-022-02646-5. Epub 2022 Sep 13.
In medicine, identifying the indirect immunofluorescence of human epithelial type 2 (HEp-2) cells plays a decisive role in the diagnosis of autoimmune diseases. The manual interpretation of Hep-2 cell images may lead to some limitations, such as subjectivity, inconsistency and low efficiency. Therefore, it is very important to automatically identify HEp-2 images. Inspired by the outstanding performance of neural networks in image classification tasks, we propose a multi-class and multiple-binary classifier (MCMBC) for the classification of HEp-2 cells. MCMBC is an ensemble learner that contains two kinds of sub-classifiers: multi-class (MC) and multiple-binary (MB). The MC sub-classifier adopts a multi-scale convolutional neural network (MSCNN) that increases the efficiency of information transmission between layers. On the basis of classification results of the MC sub-classifier on validation sets, we can find easy-to-confuse class pairs. An easy-to-confuse class pair is two classes that are not easy to be identified from each other. The MB sub-classifiers adopt multiple-binary pre-trained VGG16 networks that are used to deal with these class pairs. The final prediction for a sample possibly belonging to an easy-to-confuse class is decided by the assembled features extracted from the last fully connected layer of MC and the output of MB sub-classifiers. To evaluate the proposed model, experiments were conducted on the ICPR 2014 Task-2 dataset. Experimental results show that MCMBC performs better than the state-of-the-art method (84.68% vs. 83.35% on the criterion of average classification accuracy (ACA) and 82.89% vs. 82.67% on the criterion of mean classification accuracy (MCA)).
在医学领域，识别人类上皮2型（HEp-2）细胞的间接免疫荧光在自身免疫性疾病的诊断中起着决定性作用。对HEp-2细胞图像进行人工解读可能会导致一些局限性，如主观性、不一致性和低效率。因此，自动识别HEp-2图像非常重要。受神经网络在图像分类任务中出色表现的启发，我们提出了一种用于HEp-2细胞分类的多类和多二元分类器（MCMBC）。MCMBC是一个集成学习器，包含两种子分类器：多类（MC）和多二元（MB）。MC子分类器采用多尺度卷积神经网络（MSCNN），提高了层间信息传输的效率。基于MC子分类器在验证集上的分类结果，我们可以找到容易混淆的类别对。容易混淆的类别对是指两个不容易相互区分的类别。MB子分类器采用多个预训练的二元VGG16网络来处理这些类别对。对一个可能属于容易混淆类别的样本的最终预测由从MC的最后一个全连接层提取的组合特征和MB子分类器的输出决定。为了评估所提出的模型，在ICPR 2014任务2数据集上进行了实验。实验结果表明，MCMBC的性能优于现有方法（平均分类准确率（ACA）标准下为84.68%对83.35%，平均分类准确率（MCA）标准下为82.89%对82.67%）。
