【论文笔记】CVPR2020 Multi-scale Domain-adversarial Multiple-instance CNN for Cancer Subtype Classificatio

又一篇CVPR2020的histo image的文章，cancer subtype classification

Contribution

• cancer subtype classification 面临三个问题：

1. tumour and non-tumour regions are mixed, patch label is unavailable
2. staining conditions vary greatly
3. when the magnification of image changes, different features would be observed

• 作者结合以下framework，对于解决上述三个问题，目的是为了mimic real pratice:

1. multi-instance learning
2. domain adversarial normalisation (学习了domain adversarial training)
3. multi-scale learning

Methods

• multi-instance learning
  人n的WSI分成B_n个Bag，bag b分成I_b个instance，instance i 叫x_i。
  
  使用WSI-level的label作为supervision，将每个bag的1/0的probability平均起来成为一个人的预测结果：
  

• domain adversarial normalisation
  看了看paper: Domain-Adversarial Training of Neural Networks，真神奇。

  通过gradient reversal layer to reverse the gradient by multiplying it by a negative scalar
  during the backpropagation. In this way, the loss of the domain classifier is maximised.

  通过削弱model判断domain的能力，让model以为所有sample都来自一个domain。这样操作，model在不同domain之间判断不出区别，所以更好的利用domain-invariant features （对于判断domain无用的特征）来实现任务，更好的实现domain adaptation。
  
  在这篇文章中，作者将每个人作为一个domain，the staining condition of each person’s slide can be ignored.
  

• multi-scale learning
  训练分为2 stages：先train single-scale DA-MIL，然后再把每个scale训练好的feature extractor放到一起做multiscale。
  

Results