Deep CNNs for Diabetic Retinopathy Detection笔记

Deep CNNs for Diabetic Retinopathy Detection笔记

1.主要工作

使用卷积神经网络实现DR的两种分类(都是2分类模型);

2.数据集

Kaggle

an epoch was set to 2000 training examples

number of postive(眼底正常类) and negative(1-4级非正常类) examples are equal.

3.预处理

图片大小变为256*256;

each image was rescaled to have the same radius (the eyeball) and each pixel had its color subtracted by the local average;

The edges of the images were also clipped since there is a great variation on the boundaries or edges of the images;

4.训练过程

首先是simpler task(2分类:是否with DR)

采用了两个model:都是使用了googleNet Inception V3模型(已经预训练)最后再加了两层全连接层。

模型的区别是:model1是冻结了v3部分,只训练加的两层全连接层;而model2不仅训练全连接层,而且训练V3模型的top two blocks(layer 172 to 217)

并且只从0级和4级数据集中选取总共1665张图片 80%用做training set;剩下20%用作验证和测试集。

最后训练结果表明: model2的训练集准确率和验证集准确率存在large gap(约10%),表明overfitting了。

所以增加了data augmentation(random vertical and horizontal reflections ,Gaussian nosie, random crops, random shear)。

采用data augmentation后,准确率明显提高(整体比第一次好)。

Detecting referable DR

和上面一样的两个model

referable DR: moderate or worse DR (which excludes mild DR), and corresponds to a grading of 2-4 in our dataset

两类:(0和1) &(2-4)

4.模型结果对比:

Deep CNNs for Diabetic Retinopathy Detection笔记_第1张图片

你可能感兴趣的:(眼底图片识别)