深層学習による細胞の機能過程の同定—第2報—

深層学習による細胞の機能過程の同定—第2報—
复制标题

使用深度学习识别细胞功能过程——第 2 部分——

DOI:
10.11517/jsaikbs.127.0_08
复制
发表时间:
2022
期刊:
JSAI Technical Report, SIG-KBS
影响因子:
--
通讯作者:
坂井 恵子
坂井 恵子
中科院分区:
--
文献类型:
--
作者:
福井 凜;山本 泰生;狩野 旬;坂井 恵子

文献摘要

相似文献

在之前的工作中,我们解决了使用卷积神经网络识别电场刺激下成肌细胞分化过程的问题。然后,我们观察到基于 VGG19 的模型通过使用突起周围的空间特征实现了细胞分类的高精度。在本研究中,我们首先使用新获得的数据集评估基于 VGG19 的模型的适用性。接下来,我们应用现有的持续学习方法(称为 EWC),使用新数据集增量更新模型。由于要分析的细胞图像的数量不断增加,重新学习迄今为止生成的所有数据集是不可行的。在这里,我们的目的是阐明现有持续学习方法的可行性。通过Grad-CAM可视化的实验结果,我们发现EWC模型集中了不同的感受野,导致其预测精度与再学习模型相比显着下降。
In the previous work, we have addressed the problem of identifying the differentiation process of myoblasts under electric field stimulation using convolutional neural networks. Then, we observed that the VGG19-based model achieved high accuracy for cell classification by using spatial features around their protrusions. In this study, we first evaluate the applicability of the VGG19-based model using the newly obtained dataset. Next, we apply an existing continual learning method, called EWC, to incrementally update the model with the new dataset. It will be infeasible to re-learn all the dataset which has been generated so far, as the amount of cell images to be analyzed is continuously increasing. Here, we aim at clarifying the feasibility of the existing continual learning method. Through the experimental result with visualization by Grad-CAM, we found that the EWC model concentrated different receptive fields which caused a significant decrease of its prediction accuracy, compared with the re-learning model.