文献阅读（84）PFA-ScanNet

• 题目：PFA-ScanNet: Pyramidal Feature Aggregation with Synergistic Learning for Breast Cancer Metastasis Analysis
• 时间：2019
• 会议：MICCAI
• 研究机构：港中文/中科院

1 缩写 & 引用

• TNM: tumor, node, and distant metastasis (TNM)
• PFA: Pyramidal Feature Aggregation
• ROI: Region of Interest

2 abstract & introduction

本篇论文的主要贡献：

1. 金字塔特征聚合Pyramidal Feature Aggregation
2. 具有语义指导的协同学习

synergistic learning for training the main detector and extra decoder with semantic guidance

3. 池化层提高速度

high-efficiency inference mechanism is designed with dense pooling layers

本篇论文的目的：乳腺癌转移分析检测，检测转移的存在，并测量到四个转移类别的程度

主要难点

• 图像像素多
• 正常和癌变区域之间存在hard mimic
• 不同转移类型之间的显著大小差异

2.1 相关工作

• 教师学习网络、迁移学习
  Invasive cancer detection utilizing compressed convolutional neural network and transfer learning 2018 MICCAI
• 全卷积网络
  Fast scannet: fast and dense analysis of multi-gigapixel whole-slide images for cancer metastasis detection 2019 IEEE Trans. Med. Imaging 1
• 条件随机场考虑空间相关性
  Cancer metastasis detection with neural conditional random field 2018 arXiv

3 方法

3.1 Pyramidal Feature Aggregation金字塔特征聚合

• 通过大的卷积核获得大的感受域
• 15x15的卷积核拆成1x15+15x1和15x1+1x15两个卷积

To further reduce the computation burden and number of parameters

3.2 FCN与池化层

The pooling strides {128,64,32} are associated with feature level {3,4,5} in the training phase and will be converted to {128/α,64/α,32/α} in the inference phase

It allows dense and fast predictions when α increases

• L_p: 训练patch的size
• L_R: 推理input size
• S_p: scanning stide
• S_R: scanning stide for refetching ROIs
• L_m: size of predicted probability tile

Our model falls into the category of FCN architecture, which is equivalent to a patch-based CNN with input size Lp and scanning stride Sp , but the inference speed becomes much faster by removing redundant computations of overlaps

3.3 语义指导与协同学习

• BM用残余结构对边界对准进行建模

BM models the boundary alignment in a residual structure to take advantage of the local contextual information and localization cue

• 同时对detector和decoder训练
• 截断形式的交叉熵同时减少segmentation loss和分类loss

it is hard to minimize the classification loss and segmentation loss simultaneously in one iteration

The segmentation loss will clip outliers at the truncated point γ ∈ [0,0.5] when p(t|x;W) < γ

p(t|x;W) is the predicted probability for the ground truth label t given the input pixel x

When γ=0 it will degrade into binary cross-entropy

4 实验结果

• 数据集：Camelyon16和Camelyon17
• 随机森林分4类：正常、ITC、Micro、Macro
• 精度高，速度快