• 题目：Local and Global Consistency Regularized Mean Teacher for Semi-supervised Nuclei Classification
• 时间：2019
• 会议：MICCAI
• 研究机构：佛罗里达大学
• 参考博客：https://blog.csdn.net/JYZhang_CVML/article/details/102724783

1 缩写 & 引用

• SE: self-ensembling
• SSDL: semi-supervised deep learning半监督
• TE: temporal ensembling
• EMA: exponential moving average

2 abstract & introduction

• 问题定义：半监督分类问题
• 输入：一部分标记，一部分没标记的数据
• Label Propagation标签传播算法介绍：https://blog.csdn.net/zouxy09/article/details/49105265

半监督学习算法，核心思想：相似的数据应该具有相同的label
两大步骤：1）构造相似矩阵；2）传播

• Siamese损失函数介绍：https://blog.csdn.net/u013841196/article/details/89877340

相似性度量方法：最小化把相同类的数据之间距离，最大化不同类之间的距离

• Mean Teacher介绍：https://blog.csdn.net/u011961856/article/details/84063663

两个网络同时训练
student网络参数根据损失函数梯度下降法更新得到
teacher网络参数通过student网络的参数迭代得到

2.1 motivation

1. 半监督学习很好，但是开销大

This requirement makes Temporal Ensembling model heavy when learning on large datasets…To alleviate this problem, Mean Teacher utilizes two models…

2. 忽视了local和global的一致性

• Local consistency refers to that samples from the same class are likely to lie in the same vicinity in the feature space
  局部一致性是指同一类样本可能位于特征空间的同一邻域
• Global consistency means that samples from the same global structure are likely to share the same label
  全局一致性意味着来自相同全局结构的样本可能共享相同的标签

To enforce the local and global consistency, in this paper, we propose a novel loss function that is computed over a graph constructed via Label Propagation

Siamese loss is employed to pull the data from same class closer and push those from different classes further away. Therefore,
the two consistencies are enforced

3 方法论

1. 标签传播，得到预测值

A label propagation step is integrated into the SE method, and a graph is constructed using the label propagation predictions that encode the local and global data structure

2. Siamese损失函数，关注local and global一致性

A Siamese loss is used to learn the local and global consistency from the graph.

3. 通过mean Teacher方法完成self-ensembling

Our implementation is based on the state-of-the-art self-ensembling method Mean Teacher

4 结果

• 数据集：Ki-67、MoNuseg
• MoNuseg有四种类型的细胞核：上皮细胞核、炎性细胞核、成纤维细胞细胞核、纤维母细胞细胞核
• Ki-67也是四种类型的细胞核：免疫阳性(非)肿瘤细胞核和免疫阴性(非)肿瘤细胞核