CE-Net: Context Encoder Network for 2D Medical Image Segmentation 论文阅读

date
Oct 26, 2022
Last edited time
Mar 27, 2023 08:45 AM
status
Published
slug
CE-Net论文阅读
tags
DL
CV
summary
转
type
Post
origin
https://www.notion.so/lazurite/CE-Net-Context-Encoder-Network-for-2D-Medical-Image-Segmentation-75b29b995ff44c8ebe18e1b1d878064b
Field
Plat
CE-Net: Context Encoder Network for 2D Medical Image Segmentation
Medical image segmentation is an important step in medical image analysis. With the rapid development of convolutional neural network in image processing, deep learning has been used for medical image segmentation, such as optic disc segmentation, blood vessel detection, lung segmentation, cell segmentation, etc. Previously, U-net based approaches have been proposed.
CE-Net: Context Encoder Network for 2D Medical Image Segmentation
https://arxiv.org/abs/1903.02740
CE-Net: Context Encoder Network for 2D Medical Image Segmentation
1903.02740.pdf
2833.2KB
github.com
https://github.com/Guzaiwang/CE-Net
[论文笔记]CE-Net与PsP-Net For Semantic Seg
（今天我们来介绍两篇论文,以CE-Net为主,因为CE-Net用到了PsP-Net中的block,所以我们顺带一起讲一下。semantic seg是逐像素点的分类,所以某种意义上讲semantic seg也可以称为 dense seg） 《CE-Net: Context Encoder Network for 2D Medical Image Segmentation》是一篇将 空洞卷积和 金字塔池化 结合,专门用在2D医学图像分割的paper,文章发表在IEEE transcations on medical imaging(TMI2019).
[论文笔记]CE-Net与PsP-Net For Semantic Seg
https://zhuanlan.zhihu.com/p/172485462
[论文笔记]CE-Net与PsP-Net For Semantic Seg

Abstract

  1. 提出了DAC block(dense atrous convolution )提取更多的去全局语义信息。利用空洞卷积,通过增大感受野获取更多high-level的info,作者说high-level的信息有助于分割精确度的提升。
  1. 提出了residual multi-kernel pooling (RMP) ，它是受空间金字塔池化的启发，对DAC的输出进行多尺度的信息编码。利用不同kernel-size的池化操作,保留不同scale的空间信息。
notion image

DAC block(to capture high-level semantic feature map)

受到Inception-ResNet-V2和以上空洞卷积的启发,本文中作者提出了以下DAC结构：
notion image
简单解释一下Inception,它加宽(widen)了卷积操作,是一个并行结构,用不同size的卷积核分别对同一个feature map进行卷积操作,得到的结果进行concat操作,也可以叫做Multi-scale的操作。
提到了Inception加宽了卷积操作,同样的也可以加深(deepen)卷积操作,这就是何恺明大神著名的ResNet思想啦,可以使网络结构很深同时保证梯度不爆炸不消失。
那现在作者将Inception和ResNet思想结合起来,理论上讲可以同时拥有两种结构的优点。

RMP block(to detect objects at different sizes)

Pyramid Scene Parsing Network
Scene parsing is challenging for unrestricted open vocabulary and diverse scenes. In this paper, we exploit the capability of global context information by different-region-based context aggregation through our pyramid pooling module together with the proposed pyramid scene parsing network (PSPNet).
Pyramid Scene Parsing Network
https://arxiv.org/abs/1612.01105
Pyramid Scene Parsing Network
1612.01105.pdf
4394.1KB
这个block结构上跟PsP-net是一样的。
notion image
结构上都是对同一个feature map并行不同尺寸的pooling操作,最后把结果concat起来。这样可以得到不同size的feature map,为了减少训练参数每次pooling后进行一个1*1卷积。

实验结果

作者认为传统的交叉熵损失不适合小目标(即前景/背景很小)的医学图像,所以损失函数选用Diceloss+reg(正则化项用来防止网络过拟合)
作者分别对视神经、视网膜血管、肺部、细胞进行了分割实验。
先来看指标性结果：
notion image

© Lazurite 2021 - 2024