【Bilinear Pooling】《A Novel DR Classfication Scheme based on Compact Bilinear Pooling CNN and GBDT》

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/bryant_meng/article/details/82350404

Journal of Information Hiding and Multimedia Signal Processing（JIH-MSP） - 2018

Diabetic Retinopathy（DR） 糖网病——“糖尿病视网膜病变”的简称，糖尿病会引起视网膜血管失调。

---

目录

---

1 Background and Motivation

Diabetic retinopathy (DR) is one of the leading causes of blindness, but the classi cation of DR requires experienced ophthalmologist to distinguish the presence of various small features, which is time-consuming and difficult.

Therefore, automated DR classi cation is essential for medical treatment.

DR progresses through five stages:

• normal 正常
• mild non-proliferative DR (NPDR) 轻度非增殖 DR
• moderate NPDR 中度非增殖 DR
• severe NPDR 重度非增殖 DR
• proliferative DR (PDR) 增殖期 DR

automated DR classi cation schemes

• conventional image-analysis based schemes（low-level，在小数据集上，易overfeat，poor in actual scenes）
• deep learning based schemes（ignore the local pairwise feature，用的时softmax classifier）

为了处理 deep learning based schemes 的 two problems，作者提出

• a compact bilinear pooling CNN
• GBDT 替代 softmax

2 Innovation

Compact Bilinear Pooling CNN and Gradient Boosted Decision Tree

• 针对在糖网病分级中具有鉴别力的特征位于眼底图像的某个局部，特征粒度细的特点，将糖网病自动分级看做细粒分类问题，引入双线性池化算法获取眼底图像中重要局部位置信息。

• 考虑到双线性流池化方法需要计算外积导致计算代价大，算法引入紧凑双线性池化方法，通过采用 Count Sketch projection 方法对双线性池化进行降维处理，减小计算代价

• 考虑到CNN模型所采用的分类器为softmax，难以很好的表达复杂的局部位置信息特征，算法采用GBDT分类器替换 softmax分类器以提升自动分级算法的性能

3 Advantages

achieves a superior performance in terms of the Kappa score, F1-score and AUC

4 Methods

4.1 Compact Bilinear Pooling CNN

4.1.1 Compact Bilinear Pooling

stream A 和 stream B 计算出来的特征图大小都为 C×H×H，经过out product（$X*X^T$） 计算后，dimension 为 C*C，本算法中 C = 512，$C^2$ 达到了惊人的 26万之多，这是不能接受的。为了减少计算量，作者采用了 Count Sketch projection 方法，在不损失精度的同时将维度从$c^2$降到了d，其中 $d<<c^2$。

4.1.2 Network Architecture

VGG 把 fc6 和fc7 去掉，换成 compact bilinear pooling 层+ Signed Square Root+L2 normalization

4.2 Gradient Boosted Decision Tree

用的 XGboost

5 Dataset and Data Preparation

5.1 Dataset

EyePACS

80% for training
10% for validation
10% for test

5.2 Data Preparation

5.2.1 Image preprocessing

• resize
  images ranges from 2592×1944 to 4752×3168，作者为了减少计算量，都 resize 到 448×448

• color enhancement
  

  

G(x,y,$\rho$) represents the Gaussian lter with a standard deviation of $\rho$
* denotes the convolution operator
$\alpha, \beta, \rho, \gamma$ designed empirically as 4, -4, 10, 128 respectively.

5.2.2 Data Augmentation

• re-sampling
• spatial translation rotation
• crop

6 Experiments and Results

GTX TITAN X GPU
mini-batch of 64

6.1 Kappa score

proposed scheme 相比 Compact bilinear CNN，区别在于把 softmax 改成了 GBDT

6.2 F1-score

注意到 Mild NPDR效果普遍很差，因为该类细胞很小，resieze 后基本是 a small round dot，很多会被误认为时 Normal

6.3 ROC and AUC